Access to this full-text is provided by Taylor & Francis.
Content available from GIScience & Remote Sensing
This content is subject to copyright. Terms and conditions apply.
RESEARCH ARTICLE
Multi-year mapping of ood autumn irrigation extent and timing in harvested
croplands of arid irrigation district
Ximin Qian
a
, Hongwei Qi
a
, Songhao Shang
a
, Heyang Wan
a
and Ruiping Wang
b
a
State Key Laboratory of Hydroscience and Engineering, Department of Hydraulic Engineering, Tsinghua University, Beijing, China;
b
Department of Scientific Research, Bayannur Institute of Water Conservancy Research, Bayannur, China
ABSTRACT
Flood irrigation after crop harvest, e.g. autumn irrigation (AI), is a common irrigation practice in arid
and semi-arid regions like Hetao Irrigation District (HID) in Northwest China to increase soil
moisture and leach soil salt. Detailed information about the extent, timing, and amount of AI is
imperative for modeling agro-hydrological processes and irrigation management. However, little
attention is given to the identication of the above AI factors. There are basically three major
diculties in estimating the annual changes in AI, including a suitable index to identify AI,
temporal instability of thresholds, and an eective validation method for irrigation timing.
Therefore, this study proposes a simple and eective threshold-based method to extract the
extent and timing of AI in the HID using MODIS water indices at a daily timescale. The Multi-
Band Water Index (MBWI) time series is rst reconstructed using an adaptive weighted Savitzky-
Golay lter and then used to identify the AI extent and time. The proposed model has a stronger
generalization capability both in time and space due to robust thresholds selected from the
Z-score normalized feature variable. The model is validated both at pixels generated by the
segmentation of Sentinel-derived MBWI using a threshold-based model and at sampling points
from the eld survey. Results show that the model performed well with an overall accuracy of more
than 90.0% for the irrigation area. The overall accuracies of irrigation timing are 76.4% and 91.7%
based on the middle-to-late and whole irrigation periods, respectively. We found a decreasing
trend in the AI area and a gradual delay in the starting time of AI in the HID, mainly due to changes
in cropping patterns, climate, and irrigation fees. Overall, the model is promising in identifying
ood irrigation extent and timing in large irrigation districts and is helpful for irrigation scheduling.
ARTICLE HISTORY
Received 06 May 2022
Accepted 14 September 2022
KEYWORDS
Flood irrigation; transient
water bodies; irrigation
timing; MODIS; water index;
arid agricultural region
1. Introduction
Flood irrigation in harvested croplands is benecial in
increasing soil water storage, regulating soil tempera-
ture, reducing pests in cropland, and/or leaching salt
(Xiong et al. 2021). Flood irrigation has been adopted
in many regions of the world (Jing et al. 2020; Negri
et al. 2020), especially in the irrigation districts of arid
and semi-arid regions in Northwest China. It is also
known as autumn or winter irrigation (AI) because it is
generally applied in late autumn to early winter after
crop harvest. AI is quite dierent from irrigation dur-
ing the crop growing season. The surface water
inundated by AI often reaches several decimeters
and takes several days to fully inltrate. AI, therefore,
consumes a large amount of water, e.g. AI accounts
for about 1/3 of the annual irrigation water diversion
to the Hetao Irrigation District (HID) from Mid-
September to early November (Lu et al. 2019).
Irrigation during the crop growing season
requires rational management, and so does AI. On
the one hand, improper time and amount of AI may
lead to secondary soil salinization issues and harm
crop production (Mao et al. 2017). On the other
CONTACT Songhao Shang shangsh@tsinghua.edu.cn
GISCIENCE & REMOTE SENSING
2022, VOL. 59, NO. 1, 1598–1623
https://doi.org/10.1080/15481603.2022.2126342
© 2022 The Author(s). Published by Informa UK Limited, trading as Taylor & Francis Group.
This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use,
distribution, and reproduction in any medium, provided the original work is properly cited.
hand, large irrigation districts in the arid regions of
China are subjected to water scarcity problems due
to the limited available water for irrigation.
Therefore, agricultural water conservation and sali-
nity control have become major issues for the sus-
tainable development of irrigation districts in arid
regions (Xu et al. 2010). Under the current circum-
stances of limited water availability, the ecient use
of available water to address the soil salinity and its
leaching eect is a major concern in these regions
(Wen et al. 2020). In addition, the decrease in water
supply for irrigation has led to changes in the spa-
tiotemporal allocation of irrigation water and crop-
ping patterns, which then have signicant impacts
on the hydrological processes, agricultural produc-
tion, and the environment at a regional scale (Shi
et al. 2020). Detailed irrigation information is an
indispensable input to quantifying these changes in
the regional agro-hydrological simulation models
(Wen et al. 2022). Therefore, timely and accurate AI
information is necessary for irrigation and agriculture
administrations as well as for researchers.
Recently, the applications of remote sensing tech-
niques (satellite data) to investigate irrigation during
the crop growing season are signicantly increased
(Karthikeyan, Chawla, and Mishra 2020; Ozdogan et al.
2010). The investigation is performed at dierent spa-
tial scales by adopting optical data (Deines et al. 2019;
Salmon et al. 2015), microwave data (Dari et al. 2021;
Gao et al. 2018), or both (Bazzi et al. 2019; Elwan et al.
2022). Moreover, the investigation of irrigation during
the growing season is mainly focused on the dier-
ences in the greenness index (Xie and Lark 2021) and/
or soil moisture between rain-fed and irrigated crops.
However, very few studies have investigated AI to
date because the above-mentioned methods may
not be able to detect AI. Therefore, there is
a signicant research gap in identifying AI on a large
scale. AI identication is almost similar to the identi-
cation of water bodies because water accumulates
to the depth of several decimeters after the AI appli-
cation. Water index (WI) can be used for the identi-
cation of AI. However, the identication of AI is more
dicult than that of water bodies because the surface
water ponding as a result of AI lasts only for several
days. Therefore, the identication of AI requires high
temporal resolution data, where Moderate Resolution
Imaging Spectroradiometer (MODIS) with the daily
temporal resolution is the most appropriate remote
sensing data for studying short-term inundation (Dao,
Mong, and Chan 2019).
Meanwhile, dierentiation between water, ice, and
snow is also an important aspect in the identication
of AI because the inundated irrigation water after AI
can freeze in some meteorological conditions, where
ice and snow may have WI values close to or higher
than water (Klein and Barnett 2003). Although freez-
ing of water manifests the condition of being irri-
gated, the increase in WI due to freezing can lead to
signicant inconsistencies in feature values for lands
irrigated at dierent times, which hinders the precise
identication of AI extent and time. Therefore, snow
and ice should be handled carefully while dealing
with AI. Although many studies have compared the
performances of dierent WIs for water body detec-
tion (Ji, Zhang, and Wylie 2009), little is known about
their sensitivity to surface water change and freezing.
Therefore, a robust index is required for AI identica-
tion which should be sensitive to changes in surface
water but not to freezing.
Threshold-based models are one of the main
approaches for identifying irrigated land or water
bodies based on the selected index (Pervez and
Brown 2010). The threshold-based model requires the
selection of appropriate feature variables, which can be
divided into two categories. The rst category is based
on the feature value in a specied day or the maxi-
mum/minimum values in the time series (L. Wang et al.
2021). The second category is based on the dynamic
time series, where the combined index of dierent
periods in a specic time prole is selected, or the
time variation of the index is considered (Yeom et al.
2021). It has been proved that the time prole is more
suitable to identify irrigation than the static spectral
signatures at a one-time point (Y. Chen et al. 2018), but
it emphasizes the processing of feature time series.
Another drawback of the threshold-based method is
that the thresholds cannot be compared temporally
(Xie and Lark 2021). Therefore, pre-classication (e.g.
wet and dry years classication (Pervez, Budde, and
Rowland 2014) or data normalization techniques (Xie
and Lark 2021)) were used to obtain thresholds avail-
able for multiple years to track annual changes in
irrigation extent and timing.
The Z-score normalization method is one of the
most promising techniques for normalizing tem-
poral trajectories (Santana et al. 2018). It provides
a statistical measure of deviations from the mean
GISCIENCE & REMOTE SENSING 1599
after taking into account the natural variability
(Kreyszig 2010), and facilitates the user to dene
thresholds without considering the variations in
detected objects due to disturbing factors such as
satellite view angle (DeVries et al. 2020). Z-score has
been successfully applied so far for re detection (De
Carvalho Júnior et al. 2015), tree planting detection
(Chen, Jin, and Brown 2019), ood detection
(DeVries et al. 2020), etc. Therefore, the Z-score-
based threshold model was considered in this
study to identify AI. Similar to the identication of
irrigated paddy rice elds and irrigation timing
(Jeong et al. 2012), the AI threshold can be deter-
mined by statistical data for a particular study area,
whereas AI time refers to the rst time at which the
threshold is reached.
In addition, a sucient number of widely distributed
and representative reference data is critical to identify
and assess the accuracy of remote sensing-based AI
extent and time. Currently, irrigation products are only
available during the growing season and therefore
cannot be used to validate AI identication. Visual
interpretation of the high spatial resolution image or
Google Earth is another common method of obtaining
validation samples for the detection of irrigation areas
(Jin et al. 2016). However, it is dicult to provide
validation for irrigation timing due to the requirements
of short satellite revisit periods and suciently accurate
auxiliary information. Therefore, a simple and fast
method that can extract irrigation distribution and
irrigation time from high-resolution images on a large
scale could be useful for the validation of AI mapping.
To summarize, there are still several key challenges in
detecting the multi-year variations in AI area and timing,
such as appropriate feature extraction, temporal incom-
parability of thresholds, and limited data available for
accuracy assessment for the irrigation time. In order to
address these problems, the main objective of this study
is to develop a threshold-based model capable of
extracting multi-year AI area and timing using remote
sensing-based daily WI time series data, where
a hierarchical decision tree classier is used to determine
the irrigation time for all irrigated pixels. The model is
applied to the HID and validated from multiple perspec-
tives using agricultural statistics, eld survey data, and
information extracted from high-resolution Sentinel-2
images on Google Earth Engine (GEE).
2. Study area
Hetao Irrigation District (HID), located in Inner
Mongolia, is one of the three largest irrigation dis-
tricts in China and the largest in the arid regions of
Northwest China. It is a representative irrigation dis-
trict that relies on AI for storing water and leaching
salt. It consists of ve sub-irrigation districts, includ-
ing Wulanbuhe (WLBH), Jiefangzha (JFZ), Yongji (YJ),
Yichang (YC), and Wulate (WLT) (Figure 1). The
whole HID covers an area of about 1.189 million
ha, where 0.733 million ha (62% of the total) is
irrigated cropland that is mainly sown with corn,
sunower, and wheat. Natural land is another
major land-use type and covers approximately 19%
of the total area of HID, which consists of sparse
grassland with scattered shrubs and abandoned
cropland due to high soil salinity. The groundwater
table depth in HID is generally shallow (Wen et al.
2020), with large annual cyclic variations from 0.5
to 3.0 m.
Agricultural production in HID is highly dependent
on irrigation due to the arid and semi-arid climate
conditions (low precipitation and strong evaporation
potential). As reported in the Bayannur water resources
bulletin, irrigation water is primarily diverted from the
Yellow River, accounting for about 1/7 of the total
water diversion in the Yellow River basin. The total
amount of water diverted from the Yellow River is
approximately 4.0–4.8 billion m
3
per year from 2010
to 2020. Water available for irrigation has decreased in
recent years as a result of decreased runo and inte-
grated water resources management in the Yellow
River basin. A proper irrigation management strategy
is therefore essential for high-ecient use of limited
irrigation water, particularly for AI, which accounts for
more than one-third of the total diversion of irrigation
water each year (Lu et al. 2019).
3. Materials and methods
The owchart (Figure 2) shows that the methodology
for this study can be divided into four major steps: (a)
data collection and pre-processing, (b) feature extrac-
tion, (c) development of the algorithm for retrieving
irrigation information (extent and time), and (d) accu-
racy assessment.
1600 X. QIAN ET AL.
3.1. Data sources
3.1.1 Remote sensing data and preprocessing
Remote sensing data used in this study include the
daily land surface reectance of the MODIS
(MOD09GA) from 2010 to 2020 and Sentinel-2 from
2016 to 2020.
The MOD09GA has 500 m spatial and daily tem-
poral resolutions, which can be downloaded from
https://ladsweb.modaps.eosdis.nasa.gov/. Sentinel-2
has a spatial resolution of 10–60 m and a temporal
resolution of 2–5 days. We used Sentinel-2ʹs Top of
Atmosphere (TOA) reectance product processed
Figure 1. Sketch of the Hetao Irrigation District and field sampling points. WLBH, JFZ, YJ, YC, and WLT refer to Wulanbuhe, Jiefangzha,
Yongji, Yichang, and Wulate sub-irrigation districts, respectively; DK, HH, LH, WY, and WLTQ refer to Dengkou, Hangjinhouqi, Linhe,
Wuyuan, and Wulateqianqi counties, respectively; AI and NAI refer to Autumn irrigation and Non-Autumn irrigation, respectively.
Figure 2. Flowchart of the proposed methodology. MODIS, MBWI, and MMAVE represent Moderate Resolution Imaging
Spectroradiometer, Multi-Band Water Index, and maximum value of the three-day moving average, respectively.
GISCIENCE & REMOTE SENSING 1601
using GEE because the surface reectance data is
available for a very limited time on GEE. Several stu-
dies have reported a promising accuracy for TOA data
in detecting surface water inundation (Seaton, Dube,
and Mazvimavi 2020). The remote sensing data pre-
processing included re-projection, clipping to the
study area, and cloud and snow removal. For
MOD09GA, the daily State Quality Assurance (QA)
ags were used to mask pixels aected by cloud and
snow. On the other hand, the QA60 bitmask band was
used for Sentinel-2 to remove the cloud.
We also evaluated the potential of another MODIS
product, MCD43A4, as a data source for AI identica-
tion. The MCD43A4 product corrects the problem of
MOD09GA being aected by sensor viewing angle
and solar zenith angle (Thompson and Paull 2017).
However, it has a higher percentage of missing data in
areas with high precipitation, aerosol concentration,
or snow cover (Colditz et al. 2018).
The study period of September 15 to November 30
of each year was selected based on the historical
statistics of starting and ending dates of AI. The com-
parison of the proportions of valid observations of the
above-mentioned three products in the study period
(given in Appendix A1, Figure 13) shows that
MOD09GA is more suitable for AI identication,
while Sentinel-2 can be used as a supplement to the
validation dataset.
3.1.2 Field sampling data
For validation purposes, sampling points were col-
lected in two eld surveys from 19th to
24 October 2019 and 21st to 25 October 2020. The
surveyed elds were divided into irrigation elds and
non-irrigation elds. The information of each sam-
pling eld including the longitude and latitude of
points along the eld boundary and azimuth is
recorded in detail. The longitude and latitude are
recorded with a Juno SA Handheld Computer with
GNSS (Global Navigation Satellite System) (Trimble,
USA). The eld-scale sampling data were then aggre-
gated within a spatial window of 500 m to determine
whether the corresponding MODIS pixel has been
irrigated. If more than 50% of the pixel area is irri-
gated, it is taken as an AI sample. If the irrigated area
of the pixel is less than 50% for the whole irrigation
period, then it is considered an NAI sample (Hao et al.
2022; Tsela et al. 2014). The numbers of irrigated and
non-irrigated pixels extracted from the eld survey in
2019 and 2020 is given in Table 1.
3.1.3 Other auxiliary data
Land-use maps of the study area (with 30 m spatial
resolution) for the years 2010 (Figure 3a) and 2020
(Figure 3b) are obtained from GlobeLand30 data sets
(http://globallandcover.com/). Globeland30 is
a global land cover product developed by the
Ministry of Natural Resources of the People’s
Republic of China. The overall accuracies (Kappa coef-
cients) of the GlobeLand30 datasets in 2010 and
2020 are 83.50% (0.78) and 85.72% (0.82), respectively
(http://globallandcover.com/). In addition, the classi-
cation accuracy of cropland was 92% based on the
data collected from an agricultural sample survey in
HID in 2020, where the remaining 8% of the cropland
is mainly classied as grassland.
The soil texture data (Figure 3d) obtained from the
Harmonized World Soil Database within the territory
of China in 1:1,000,000, was provided by the Cold and
Arid Regions Sciences Data Center at Lanzhou, China
(http://westdc.westgis.ac.cn/).
Statistical data of the AI area of each sub-irrigation
district from 2010 to 2020 (where the data for WLBH
and WLT are temporarily unavailable from 2019 to
2020) were provided by the Bayannur Institute of
Water Conservancy Research. Furthermore, the statis-
tical data of the AI area in the WLT after 2014 was not
used for model calibration and validation, as the data
for this period included only the Yellow River irriga-
tion area and was incongruent with other areas.
3.2. Masking of potentially irrigated area
In order to be consistent with MODIS data, the land-
use maps (Figures 3(a and 3b) were aggregated to
500-m pixel size using the nearest-neighbor
approach. All the croplands and grasslands north of
Table 1. The number of validation pixels extracted from field
sampling and Sentinel-2 images from 2016 to 2020.
Source Class
Year
2016 2017 2018 2019 2020 total
Field sampling AIL 56 117 173
NAIL 16 49 65
Sentinel-2 AIL 342 490 883 497 1577 3789
NAIL 464 484 510 725 511 2694
Note: AIL and NAIL represent the Autumn irrigation land and Non-Autumn
irrigation land, respectively.
1602 X. QIAN ET AL.
the ood bank in 2010 and 2020 land-use maps were
used as a mask of potentially irrigated areas for
further irrigation events detection. This is because
croplands and grasslands are often confused in land
cover classication (Shazadeh-Moghadam et al.
2021) and they are often converted to each other in
HID. Furthermore, croplands between the Yellow
River and the ood bank (Figure 1) generally do not
irrigate in autumn. However, these croplands exhibit
similar spectral characteristics to irrigated croplands
as a result of shallow groundwater levels and high soil
moisture due to the Yellow River and General Main
Canal close to the ood bank. Therefore, these crop-
lands were not considered to avoid misjudgment of
the AI location.
3.3. Method to extract irrigation information
3.3.1. Reconstruction of WI time series and WI
selection
A daily continuous time series of each index was rst
generated by linear interpolation based on cloudless
observation data and then ltered using an adaptive
weighted Savitzky-Golay (S-G) lter, a noise reduction
process proposed by Hu et al. (2020). The lter origin-
ally proposed for the MODIS-based Normalized
Dierence Vegetation Index (NDVI) is applied to the
MOD09GA. This lter eliminates the eect of negative
bias and improves the quality of the index time series.
In the ltering process, weights of indices on dierent
days were determined from the quality control para-
meters of the MOD09GA products. The four-bit range
of each reectance band provided by the QC band
has ve situations in the study area, i.e. 0000, 1000,
1011, 1101, and 1110. According to MODIS Surface
Reectance User’s Guide Collection 6 (Vermote,
Roger, and Ray 2015), they represent dierent quality
control aspects. In short, 0000 and 1000 are of high
quality; 1101 are of average quality; 1011 and 1110
are of low quality. Combined with the cloud and snow
conditions provided by the State band, the weight
(ω
i
) corresponding to the i-th point, is specied in
Table 2.
Initially, the lter window size is set to 7. In order to
reduce the computation complexity and reserve the
original high-quality data, the index values for points
with weight ω
i
= 1 remain unchanged in the tting
process. The lter window will expand adaptively
depending on the number of valid values (values
Table 2. Weights for the adaptive weighted Savitzky-Golay filter
determined from quality flags of the MOD09GA product.
Quality
band
name Flag Data quality description Weights
QC 500 m Bands 1–7
data
quality
Quality of all used bands is high 1.0
Quality of all used bands is above
average but not all high
0.8
Quality of all used bands is poor 0.2
Value obtained by linear
interpolation
0.1
The others 0.4
Figure 3. The Hetao Irrigation District maps: (a) Land-use map in 2010; (b) Land-use map in 2020; (c) Land-use changes from 2010 to
2020; (d) soil texture map.
GISCIENCE & REMOTE SENSING 1603
with weights greater than 0.1). The maximum lter
window size should not exceed 15, because data
beyond seven days before and after the current date
has little inuence on the central value.
We selected seven indices that could potentially be
used for AI identication, including Normalize
Dierence Water Index (NDWI) using bands 4 and 6
(NDWI46) and bands 1 and 7 (NDWI17) (Boschetti
et al. 2014), Multi-Band Water Index (MBWI) (X.
Wang et al. 2018), the 2015 water index (WI2015)
(Fisher, Flood, and Danaher 2016), Automated Water
Extraction Index for images with shadows (AWEIsh)
and without shadows (AWEInsh) (Feyisa et al. 2014),
and Soil Moisture Monitoring of Remote Sensing
index (SMMRS) (Zhan et al. 2006). We compared the
sensitivity of these indices to changes in water surface
and ice by analyzing all the eld samples from 2019
and 2020. Overall, MBWI is the most suitable index for
AI identication (see Appendix A2 for more details).
3.3.2. Selection of classification features
The feature variable derived from the MBWI curve is
selected based on the following two hypotheses: 1)
the peak MBWI value of irrigated land is higher than
that of non-irrigated land in the same sub-irrigation
district; 2) the uctuation of MBWI caused by rainfall/
snowfall, ice, and other factors during the study per-
iod is unlikely to cause a larger MBWI increase than AI
under normal circumstances.
However, we did not directly nd the maximum
MBWI value in the ltered MBWI time series in some
cases, which is mainly due to two reasons. First, some
of the peaks are not eective even after ltering the
MBWI because some uctuations may be caused by
external disturbances, e.g. subpixel clouds and incor-
rect ags in MODIS quality information (Y. Chen et al.
2018). Second, since the weighted S-G lter was
mainly employed to improve the quality of the
index time series without considering the smoothing
of curves, the ltered series may still have signicant
uctuations, making it dicult to extract the temporal
information of AI. The water surface in HID after AI can
be generally kept for 3–14 days depending mainly on
the applied irrigation depth and soil texture.
Therefore, a pixel having surface water for more
than three consecutive days should be identied as
an AI pixel. We further calculated the three-day mov-
ing average of the ltered MBWI time series, which
can make the time series smoother with a more
obvious changing trend, and thus help to determine
the AI time more precisely. Finally, we selected the
maximum value of the three-day moving average of
the ltered MBWI time series (MMAVE-MBWI) as the
feature variable.
3.3.3. Algorithm for identifying autumn irrigation
land
The key point of the threshold-based method is to
determine a threshold that can be used for multiple
years. The variation of reectance values of MOD09GA
products with sensor view-angle and solar zenith angle
and the application of S-G lter on MBWI time series
may lead to inconsistent inter-class separability across
years (Shao et al. 2016). Therefore, processing the data
is necessary to stabilize the thresholds. Several studies
have shown that the Z-score normalization method
can be used to nd outliers within the year, where
the resulting maps developed from Z-score normaliza-
tion are better than those derived from simple empiri-
cal threshold methods (B. Chen, Jin, and Brown 2019;
DeVries et al. 2020). Therefore, we used the Z-score
normalization method to eliminate the abnormal inter-
annual uctuations in each sub-irrigation district. The
Z-score can be calculated using the following equation:
Z¼aμ
δ(1)
where a and Z are the initial and normalized values of
the selected feature; μ and δ are the average and
standard derivation of the feature variable for
a reference land-use type in the study region. As an
objective baseline, the reference land-use type should
be able to reect inter-annual variations in feature
values due to factors unrelated to the study objectives,
and it should be available for each sub-irrigation dis-
trict. We compared three possible reference land-use
types, i.e. water bodies, all the croplands and grass-
lands, and part of croplands and grasslands with higher
MMAVE-MBWI, to nd an appropriate reference land-
use type (see Appendix A3 for more details). After
Z-score normalization, the threshold value of the fea-
ture variable (MBWI
Te
) was calibrated using the statis-
tical data for AI areas of each sub-irrigation district.
3.3.4. Algorithm for determining autumn irrigation
time
As a result of fragmented croplands owned by dier-
ent farmers, a cropland pixel of 500 m by 500 m would
1604 X. QIAN ET AL.
usually contain several pieces of cropland that may or
may not be irrigated at the same time. More and more
cropland in one pixel gets inundated with the appli-
cation of irrigation, which allows the coarse-
resolution sensors to start detecting cropland with
water cover. The change of MBWI in an AI pixel is
comprised of the following stages (Figure 4).
First, before the application of AI, a cropland pixel
changes from crop cover to bare soil immediately after
crop harvest. MBWI curve shows a valley point shortly
before AI (Point A in Figure 4) because the MBWI values
for cropland are generally higher than those of bare soil
(X. Wang et al. 2018). The MBWI at point A increases
when the cropland in the pixel is irrigated. After reach-
ing a specied threshold (Point B in Figure 4), a certain
percentage of the cropland in the pixel is irrigated and
thus the entire pixel can be considered to reach the
level of AI. Therefore, the irrigation time is determined
as the date of point B. The specied threshold (MBWI
T
)
corresponding to B is obtained by inverse normaliza-
tion using the calibrated MBWI
Te
, which is
MBWIT¼MBWITeδþμ(2)
MBWI rises continuously with AI until the irrigation
stops. At this specic point, the MBWI reaches its peak
(Point C in Figure 4) and the water depth of the entire
pixel reaches its maximum level. After that, MBWI
begins to decrease and reaches the threshold for the
last time (Point D in Figure 4). At this specic thresh-
old, it can be considered that the whole pixel is no
longer inundated, i.e. no longer in the AI state. MBWI
continues to decline and reaches the rst valley point
(Point E in Figure 4).
According to the characteristic points A, B, C, D,
and E of the MBWI time series, characteristic periods
can be dened, including peak range and initial, mid-
dle, and late irrigation stages. Peak range (points A to
E in Figure 4) is dened as the time range to the peak
between two valleys. If the ending valley is greater
than the threshold, the time range should extend to
the next valley below the threshold. Initial irrigation
stage (points A to B in Figure 4) illustrates that the
cropland in a pixel starts to be irrigated until the pixel
is in the state of AI with a certain proportion of
irrigated cropland. Middle irrigation stage (points B
to C in Figure 4) is the period when the water content/
depth of the entire pixel keeps increasing with con-
tinued irrigation until it reaches the peak and the
irrigation stops. Late irrigation stage (points C to
D in Figure 4) is the period when the whole pixel is
still inundated although irrigation has stopped, i.e.
the whole pixel is still in the AI state. Overall, the
whole irrigation period (points A to D in Figure 4)
combines the initial, middle, and late irrigation stages.
In general, the MBWI time series of a pixel may
have one or more peaks exceeding the threshold.
The following steps (Figure 5) are used to determine
the date of AI for dierent cases.
(1) For pixels with only one peak exceeding the
threshold (Figure 4a), the rst day when the
MBWI reaches the threshold is taken as the AI
date.
(2) For pixels having multiple peaks that exceed
the threshold and the highest peak appears
rst (Figure 4b), the rst day when the MBWI
Figure 4. Illustration of MBWI curves for three representative AI pixels and irrigation stages. Point A corresponds to The
Valley of the smoothed MBWI curve before autumn irrigation, Points B and D correspond to the threshold value in the
increasing and decreasing segments of the MBWI curve after autumn irrigation, Point C corresponds to the peak value of
smoothed MBWI curve, and Point E corresponds to the first valley of the smoothed MBWI curve after autumn irrigation. DOY
represents Day of the Year.
GISCIENCE & REMOTE SENSING 1605
reaches the threshold is taken as the AI date
because the subsequent peak(s) may be caused
by icing or other factors.
(3) For pixels having multiple peaks that exceed
the threshold where the highest peak did not
appear rst (Figure 4c), we developed
a “proximity algorithm” based on the assump-
tion that the AI dates of adjacent pixels should
be close or roughly the same. The proximity
algorithm is based on a 3 by 3 window centered
on this pixel with the following steps:
(a) If more than half of the AI dates of other pixels
in a 3 by 3 window are within a certain peak time
range of the central pixel (points A to E in Figure 5),
the AI dates in this 3 by 3 window are considered to
be consistent. In this case, the AI date of the central
pixel is the date when the MBWI rst reaches the
threshold within this consistent peak range.
(b) If the conditions in step (a) are not met, the AI
date is taken as the date when MBWI rst reaches the
threshold within the peak time range where the max-
imum rising segment of the MBWI curve intersects.
(c) Step (a) is performed iteratively until the irriga-
tion dates cannot be determined for more pixels.
(d) The irrigation date of the nal remaining points
is determined as the date when MBWI rst reaches the
threshold within the peak range of the maximum
peak.
3.3.5. Sensitivity analysis of the threshold
Sensitivity analysis of the threshold is performed
by taking the calibrated threshold in each sub-
irrigation district as the baseline (T
b
) and the
range of variation is set from −0.3 to 0.3. Then
the AI areas in the sub-irrigation districts are esti-
mated with dierent thresholds, and the sensitiv-
ity of the threshold (S
T
) can be calculated from
Wen et al. (2022)
ST¼ΔA
Ab
=ΔT
Tb
(3)
where S
T
represents the ratio of the relative change in
model output to the relative change in the threshold,
A
b
is the estimated AI area corresponding to T
b
, ΔA is
the variation of the simulated AI area, and
Figure 5. The hierarchical structure of the decision tree classifier for irrigation timing. AI and FDRT represent autumn irrigation and the
first day when the MBWI reaches the threshold.
1606 X. QIAN ET AL.
ΔT represents the change of the threshold. Generally,
a greater |S
T
| indicate the threshold is more sensitive.
3.4. Generating validation points from Sentinel-2
images
AI distribution maps are also veried using validation
points extracted from Sentinel-2 . The Sentinel-2
MBWI distribution maps from 2016 to 2020 were
generated using the GEE platform to produce refer-
ence AI points. MBWI was originally proposed based
on the Landsat 8 surface reectance product, where
the zero value represents the default threshold for
separating water from non-water bodies (X. Wang
et al. 2018). Since the water depth and surface area
of AI cropland are shallower and smaller than those of
water bodies, it can be assumed that the MBWI
threshold of water bodies is greater than or equal to
that of AI cropland. Therefore, the threshold of water
bodies is selected as the threshold of AI cropland to
ensure the accuracy of the validation sample for AI
lands. Since the TOA reectance product of Sentinel-2
is used, the default threshold from Landsat 8 surface
reectance product has to be redened by tting the
linear relationship between Landsat-based MBWI
(LMBWI) and Sentinel-2-based MBWI (SMBWI) with
images on the day when data is available from both
datasets. AI date of the corresponding pixel is the
date when the MBWI rst exceeds the threshold to
become a water-like body. Further, to comply with
the MODIS-derived results, the percentage of irri-
gated Sentinel-2 pixels to the total number of
Sentinel-2 pixels within each MODIS pixel (500 m)
was recorded.
A pixel with a percentage of AI over 50% is
extracted as an irrigated pixel (Hao et al. 2022) and
its irrigation date is recorded as the time when more
than 50% of the pixel is irrigated. However, the
threshold method is no longer applicable for the
selection of non-irrigated sample points. The multiple
snapshots of landscape from Google Earth, Sentinel-2,
and Landsat8 imageries are visually interpreted to
supplement the eld sampling data.
3.5. Calibration and validation of the model
Available studies have shown that the choice of
calibration and validation period has a signicant
impact on the evaluation of model performance
and further simulations (e.g. Myers et al. 2021).
Instead of arbitrarily choosing calibration and vali-
dation periods, we divide the dataset into training
and testing sets. The training dataset includes infor-
mation about the AI area of ve sub-irrigation dis-
tricts from 2010 to 2017. The testing dataset
includes recorded AI area data from 2018 to 2020,
eld sampling points from 2019 to 2020, and vali-
dation points extracted from Sentinel-2 from 2016
to 2020. A k-fold cross-validation method is used to
determine the appropriate threshold and evaluate
the performance of the model by iteratively with-
holding samples from the training dataset (Y. Chen
et al. 2018). Moreover, the testing dataset is used to
evaluate the accuracy of the model in greater detail.
3.5.1. Cross-validation in each sub-irrigation district
The thresholds in each sub-irrigation district are cali-
brated by minimizing the mean absolute relative error
(MARE) between simulated and observed annual
autumn irrigated areas, which is
MARE ¼1
nX
n
i¼1
^
yiyi
yi
��������(4)
where ^
yi and yi are the simulated and observed areas
at the year i, and n is the total number of years.
In order to select the threshold, we used k-fold
cross-validation, which involves training the model
using k−1 folds of training data and validating the
trained model using the remaining data. The data of
8 years acquired from 2010 to 2017 were divided into
four folds (k = 4), where data of 6 years is used to train
the model, and the remaining 2 years’ records are
used for model validation. However, the data from
2010 to 2014 was used for calibration and validation
in WLT, where data for three years was selected ran-
domly for calibration and the other two years for
validation. Consequently, a total of 28 models are
obtained for WLBH, JFZ, YJ, and YC, while 10 models
are for WLT after the cross-validation. The model with
the lowest MARE for each sub-irrigation district was
selected to further evaluate the accuracy.
3.5.2. Validation with testing dataset in the whole
HID
In the testing dataset, the irrigation extent is vali-
dated using statistical AI area data. Further, both
the irrigation extent and time are validated using
GISCIENCE & REMOTE SENSING 1607
eld sampling points and points extracted from
Sentinel-2.
Relative error (RE) and overall accuracy (OA) are
used to assess the accuracy of AI area and specic
locations, respectively, which are expressed as
REi¼^
yiyi
yi
(5)
OA ¼ ðNm=NÞ 100%(6)
where N
m
is the number of simulation objects match-
ing the observation object, and N is the number of the
observation objects.
OA is also used to assess the accuracy of the
determined AI date, where two dierent OAs are
used based on dierent irrigation periods. For the
middle to late period of irrigation (MTLP-based
OA), N
m
in the OA is the number of samples
whose observed irrigation dates are within the
date intervals of the Mid-to-late irrigation period
(points B to D in Figure 4). However, N
m
in OA
based on the whole period of irrigation (WP-based
OA) is the number of samples whose observed
dates are within the date intervals of the whole
irrigation period (points A to D in Figure 4).
WP-based OA is used in the validation of eld
sampling points because the recorded date may be
in any stage of AI for a pixel. For validation points
extracted from Sentinel-2, the recorded date is the
date on which most of the areas (50% or more) in
the pixel have been irrigated. Although the thresh-
old is high, it may also be lower than the value of
MBWI when the cropland has just begun to be
irrigated. Due to the revisit cycle and cloud cover,
Sentinel-2 may not be able to capture dates with
maximum MBWI value for many pixels. Therefore,
MTLP-based OA is more reasonable for points
extracted from Sentinel-2.
4. Results
4.1. Cross-validation results based on statistical
autumn irrigation areas
Based on the results of the inter-annual uctuations of
average feature variable (MMAVE-MBWI) for three dif-
ferent reference land-use types, i.e. water bodies, all
the croplands and grasslands, and part of croplands
and grasslands with higher MMAVE-MBWI, there is
a good agreement among the three land-use types
from 2010 to 2020 (Figure 15, Appendix A3). Since
most of the study area is covered by cropland and
grassland , we selected the average MMAVE-MBWI of
all the cropland and grassland pixels which is accurate
enough to reect the inter-annual changes of
MMAVE-MBWI for the study area.
After MBWI normalization with reference to the
average MMAVE-MBWI of all the cropland-grassland
pixels, 4-fold cross-validation was used to calibrate
and validate the model. It is observed the model per-
forms well when data from dierent periods was used
for calibration. The average MAREs during the calibra-
tion period of all models developed in the cross-
validation processes in each sub-irrigation district
are smaller and range from 2.7% to 7.6% (Figure 6,
left panel). The comparison shows that dierent cali-
bration periods results in slight dierences in the
MAREs . The dierence between maximum and mini-
mum MAREs in JFZ is only 2%, and the dierences in
other sub-irrigation districts are about 5%.
However, the ranges of MAREs in the validation
period are signicantly greater than those in the cali-
bration period (Figure 6, middle panel). These dier-
ences in MAREs might be associated with fewer years
(only 2 years) for validation. When the calibration and
validation periods were considered as a whole, the
variations in MAREs in each sub-irrigation district
becomes smaller and almost concentrated near the
mean or median (Figure 6, right panel). The MARE of
JFZ is the smallest with a mean value of 2.8%, while
WLBH, YJ, YC, and WLT have mean MAREs close to
each other ranging from 6% to 8%. These results
suggest that the threshold-dependent classier was
quite robust and produced reasonably good results.
Among the models developed for each sub-
irrigation district, we selected the model with mini-
mum MARE during the whole calibration and valida-
tion period as the nal model for further validation of
AI location and calculation of AI time. The nal MBWIT
and MBWITe of ve sub-irrigation districts are shown
in Table 3. The minimum and maximum MBWI
T
are
observed in WLBH and WLT, respectively. This is
mainly because WLBH has more sand while WLT has
more clay (Figure 3(d), resulting in dierent surface
water retention capacities.
The results of sensitivity analysis illustrated that the
values of S
T
are −0.22, 0.04, 0.21, −0.13, and −0.37 for
WLBH, JFZ, YJ, YC, and WLT, respectively. Since the
1608 X. QIAN ET AL.
autumn irrigated area decreases with the threshold,
the signs of S
T
values are opposite to those of MBWI
Te
(Table 3). In other words, a 100% increase in the
absolute values of the threshold results in the
decrease of identied AI area by 22%, 4%, 21%, 13%,
and 37%. In general, the models for dierent sub-
irrigation districts have a low to moderate sensitivity
to the threshold. Among the irrigation districts, JFZ
has the lowest sensitivity, probably due to its higher
percentages of AI in all years.
4.2. Model accuracy assessment based on testing
datasets
4.2.1. Validation points generated from Sentinel-2
images
The tting results between SMBWI and LMBWI are
SMBWI = 0.88*LMBWI−0.015, R
2
= 0.83, p < 0.05.
Therefore, we set the threshold value of SMBWI as
−0.015. If the cropland pixel’s MBWI value is greater
than −0.015 on a given day, the pixel is then identied
as an AI pixel. A total of 6483 testing samples (3789
irrigated and 2694 non-irrigated) from 2016 to 2020
were selected from Sentinel-2 images to further verify
the accuracy of the model (Table 1). The distribution
of validation points is shown in Appendix A4
(Figure 16).
4.2.2. Accuracy assessment of autumn irrigation area
and its distribution
The model performed well in the testing years, as
illustrated by smaller relative errors between
observed and simulated irrigation areas at WLBH in
2018 and JFZ, YJ, and YC from 2018 to 2020 (Table 4).
The relative errors are within the acceptable range,
among which 52% are in the range of −5.0% to 5.0%
and 81% from −10.0% to 10.0%. The MAREs for the
testing year (2018–2020) are 2.2%, 7.3%, 6.0%, and
4.5% for WLBH, JFZ, YJ, and YC, respectively. The
MAREs of sub-irrigation districts and the whole HID
from 2010 to 2020 are between 4% and 7%.
Therefore, the threshold-based model developed
using data from 2010 to 2017 can be applied to
identify AI in other years with satisfactory precision.
Moreover, the overall accuracy for MMAVE-MBWI-
based classication model was 90.0% and 97.0%
according to the eld sampling data and the refer-
ence data extracted from Sentinel-2, respectively
(Table 5). In summary, the good performance of
Figure 6. Box plot for MAREs of each sub-irrigation district for calibration period (left), validation period (middle), and the whole
calibration and validation periods (right).
Table 3. Threshold of each sub-irrigation district from 2010 to
2020.
Sub-irrigation district Range of MBWI
T
MBWI
Te
WLBH −0.53 to −0.46 0.26
JFZ −0.41 to −0.33 −0.08
YJ −0.51 to −0.40 −0.44
YC −0.40 to −0.30 0.24
WLT −0.37 to −0.28 0.51
Note: MBWI
T
and MBWI
Te
represent the threshold and the normalized
threshold, respectively.
GISCIENCE & REMOTE SENSING 1609
model validations with independent datasets sup-
ports the robustness and potential of the method
for extracting AI attributes.
4.2.3. Accuracy assessment of autumn irrigation
timing
The accuracy of estimating irrigation time is given in
Table 6. For eld sampling points, WP-based OA is
higher than 90.0% in both 2019 and 2020, with an
average of 91.7%., However, the MTLP-based OA for
points extracted from Sentinel-2 varies from 66.5% to
82.9% with an average of 76.4% when 50% of irri-
gated area is taken as the lower limit of the irrigated
fraction to dene an irrigated MODIS pixel. WP-based
OA has higher values than those from MTLP-based OA
because of the longer time interval.
Furthermore, the MTLP-based OA from 2017 to 2020
can be signicantly higher than in 2016. This may be
related not only to the identication results but also to
the reference data used for assessment. The pixels in
study area have average Sentinel-2 images of 7.4, 12.2,
13.6, 13.5, and 12.7 days during the study period from
2016 to 2020. A signicantly smaller number of images
in 2016 may cause a loss of eective information, thus
failing to provide a more realistic assessment of the
model quality. Therefore, we attempted to increase the
eective information by increasing the lower limit
(80%) for the irrigation fraction of the validated pixels
to account for the dierence in MTLP-based OA in
dierent years. It can be found that the average MTLP-
based OA will rise to 88.9%, with a signicant increase
in 2016 in particular (Table 6).
Overall, the observed AI time matches well with the
irrigation dates simulated by our model, which indicates
the method used to calculate irrigation time is reliable.
4.3. Spatial-temporal changes in the distribution of
irrigated croplands across HID
Figure 7 depicts the spatial distributionof AI areas in
HID from 2010 to 2020, where the irrigated area
depends mainly on the availability of runo from
Table 4. Relative error (%) of each sub-irrigation district and the whole HID from 2010 to 2020.
Year WLBH JFZ YJ YC WLT HID
2010 −4.1 −6.2 −3.8 −17.0 −8.7 −9.4
2011 0.2 −2.0 −3.7 −7.8 −14.1 −5.5
2012 −6.6 0.1 −0.5 −2.3 0.4 −1.5
2013 0.2 2.7 0.2 8.5 −0.5 3.1
2014 13.2 6.7 16.4 13.6 6.5 12.2
2015 16.1 −2.7 7.0 −8.2
2016 −12.0 1.2 −3.5 0.2
2017 9.9 0.2 9.3 0.3
2018 −2.2 16.2 13.3 6.3
2019 5.5 2.7 −2.6
2020 −0.1 −1.9 −4.7
MARE (%) 7.2 4.0 5.7 6.5 6.0 6.3
Note: MARE refers to mean absolute relative error. Underlined numbers are for the testing dataset.
Table 5. The overall accuracy of irrigated land mapping using reference data.
Data source Identified class
Actual class
AIL NAIL Total Correct ratio
Field sampling data AIL 154 5 159 96.9%
NAIL 19 60 79 75.9%
Total 173 65 238 OA = 90.0%
Sentinel-2 reference data AIL 3625 30 3655 99.2%
NAIL 164 2664 2828 94.2%
Total 3789 2694 6483 OA = 97.0%
Note: AIL and NAIL represent Autumn irrigation land and Non-Autumn irrigation land, respectively.
Table 6. Overall accuracy (OA) of irrigation timing from 2016 to 2020.
Source OA
Year
2016 2017 2018 2019 2020 Average
Field sampling WP-based 92.1% 91.3% 91.7%
Sentinel-2(50%) MTLP-based 66.5% 79.1% 82.9% 81.1% 72.4% 76.4%
Sentinel-2(80%) MTLP-based 100.0% 75.8% 96.7% 100.0% 86.5% 88.9%
Note: 50% and 80% are lower limits of the irrigated fractions to define an irrigated MODIS pixel.
1610 X. QIAN ET AL.
the Yellow River that varied by as much as 20%
among dierent years. Over the 11 hydrological
years, the irrigated area varied between a maximum
of 4758 km
2
in 2011 and a minimum of 4121.5 km
2
in
2019. The mean autumn irrigated area from 2010 to
2020 was 4444.3 km
2
with a standard deviation of
215.25 km
2
, providing insights into the inter-annual
variability in the extent of the autumn irrigated area.
The proportion of the total AI area in each sub-
irrigation district has remained stable for these years,
with average values of 10.54% for WLBH, 26.70% for
JFZ, 21.99% for YJ, 29.27% for YC, and 11.48% for WLT.
Although there have been slight uctuations in some
years, the irrigated area has shown a downward trend
during the past 11 years (Figure 8), with an average
decreasing rate of 53.76 km
2
/year in the whole HID.
Besides, YC and YJ also have higher decreasing rates
of 20.50 and 17.44 km
2
/year, respectively. Based on
the statistical data of planting areas acquired from
Water Conservancy Development Center of HID and
the estimated autumn irrigated areas in Figure 8, JFZ
has the highest proportion of AI within the total
planting area (83.5%), while WLBH has the lowest
proportion (58.1%). Despite the largest cropland
area in YC, its proportion of AI (70.8%) is considerably
lower than that of JFZ. Overall, the proportion of AI
has reached 72.0% for the whole HID.
Superimposing multi-year images of AI area can
eectively detect the maximum extent of irrigation
area and nd the autumn irrigated times for each
pixel from 2010 to 2020 (Figure 9 a). Some lands
have never been irrigated in autumn, mainly grass-
land, wasteland, and sunowers in WLBH and WLT.
Highly concentrated AI lands are mainly distributed in
the JFZ and YJ, the southern WLBH, eastern YC, and
northern WLT. The distribution of AI in the WLBH is
mainly concentrated near the main canal.
By aggregating the 500 m irrigation map into
a uniform 9 km
2
pixel and performing linear regression
on the irrigation area of each aggregated pixel, we
further studied the spatial trends in the variations of
irrigated area over time on a ne scale (Figure 9(b).
From Figures 7 and 9b, we can nd that the autumn
irrigated areas in the southern YC and the northern YJ
have signicantly decreased since 2014. The western
part of the WLT has hardly been irrigated in autumn
since 2016. The northeast corner of the JFZ has also
seen a reduction in AI area since 2015.
4.4. Spatial-temporal changes in autumn irrigation
timing
Figure 10 depicts the spatial distribution of AI dates
across HID from 2010 to 2020. The gure shows
Figure 7. Spatial distributions of autumn irrigated areas in the Hetao Irrigation District from 2010 (a) to 2020 (k).
GISCIENCE & REMOTE SENSING 1611
obvious dierences in the spatial distribution of AI
time among dierent years. In general, earlier irriga-
tion was found around the boundary between YC and
WLT and some areas in the north of HID. Compared
with highly variable irrigation time in areas with small
and scattered cropland, more synchronized irrigation
was found in the southern part of HID (mainly in JFZ
and YJ). The identied irrigation time of WLBH varies
greatly from year to year.
As an example, Figure 11 shows evolution process
of the irrigation area in 2019. A small-scale AI area rst
appeared in YJ and the east of YC. Before
October 15th, fewer areas in the HID have been irri-
gated, while a wide range of AI has appeared in JFZ,
YJ, and YC since October 15th. By November 15th, the
whole HID has almost completed AI. The results
(cumulative area diagram shown in Appendix A5,
Figure 17) show that irrigation was earlier in 2010,
2012, and 2018, while delayed in 2015 and 2020.
Figure 12 shows that the start time of AI is delayed
gradually from 2010 to 2020. Meanwhile, the propor-
tion of cropland irrigated from mid-September to
mid-October showed a downward trend and reached
a new low of 6% in 2020. Since 2012, more than 70%
of the cropland is irrigated from mid-October to mid-
November, and this proportion is still increasing.
5. Discussion
Our study found that AI in HID showed dierent
spatial distribution patterns each year, with
a downward trend in the area, and a delayed starting
time in the study period. This is consistent with our
Figure 9. Spatial distribution of autumn irrigation times (a) and variation trends in the irrigated area for aggregated pixels of 9 km
2
(b)
in the 11 years from 2010 to 2020.
Figure 8. Inter-annual changes of the autumn irrigation areas (S) in each sub-irrigation district and the whole HID from 2010 to 2020.
t is the year.
1612 X. QIAN ET AL.
Figure 10. Spatial distributions of autumn irrigation time in the Hetao Irrigation District for the period of 2010 (a) to 2020 (k).
Figure 11. The time transition diagram of the spatial distribution of autumn irrigation in 2019, divided into five periods with half-
month intervals: September 15th to September 30th (a), September 15th to October 15th (b), September 15th to October 31st (c),
September 15th to November 15th (d), and September 15th to November 30th (e).
GISCIENCE & REMOTE SENSING 1613
eld investigation and previous literature results (L.
Wang et al. 2021; Yang, Liu, and Liu 2017).
The distribution pattern of AI is inuenced by crop-
ping patterns and soil texture. The cropping patterns
dier in dierent sub-irrigation districts, especially the
percentage of sunower planting. The sunower is
more salt-tolerant and planted later than other
major crops, which can be irrigated in the spring
before sowing to leach salt and increase soil moisture.
Therefore, sunower has less demand for AI. As
a result, JFZ with a lower percentage of sunower
(Wen, Shang, and Rahman 2019) has the highest pro-
portion of AI within the total planting area. On con-
trary, YC and WLT with higher percentages of
sunower have lower proportions of AI.
Furthermore, WLBH has the lowest proportion of AI
mainly due to sandy soil texture (Figure 3(d). The
sandy soil with poor moisture retention capacity
results in poor eect in increasing the root zone
moisture, which reduces the farmers’ enthusiasm for
AI and leads to great variation in the AI time
(Figure 10) dependent on the farmers’ decisions.
The decreasing trend in the AI area (Figure 8) is
mainly related to the change in cropping patterns and
irrigation water fees. Increased area and extent of
sunower planting in recent years (Yu and Shang
2017) is the main reason for the decreasing trend of
AI area. The decreasing rate was greater in YC and YJ
and smaller in WLBH, JFZ, and WLT. This is mainly
because the southern part of YC and WLT has always
been the growing area of sunowers, but recently
sunowers are expanding toward the middle of HID,
i.e. the rest of YJ and YC. Besides, the water fee of AI is
higher than that of spring irrigation in HID, which
results in the decreased enthusiasm of the farmers
for AI and also leads to a decreasing trend in the AI
area.
The postponement of AI time is related to not only
cropping patterns but also some other factors such as
climate change. In recent years, the proportion of late
autumn crops tends to increase, such as maize har-
vested at the end of September, which leads to
delayed AI time in these croplands. Meanwhile, cli-
mate change characterized by temperature rising and
shortened freezing periods leads to the poorer eect
of early irrigation (Yang, Liu, and Liu 2017).
The changes in AI area and time signicantly
impact the hydrological cycle, agriculture, ecosystem,
and environment in HID. On the one hand, a decrease
in the AI area leads to a drop in groundwater levels.
Consequently, road slurry and land collapse caused
by frost heave and thaw collapse in mid-spring will
become less frequent (Yang, Liu, and Liu 2017). In
addition, although the area irrigated by water diver-
sion from the Yellow River is decreasing year by year
in WLT, the total AI area is still stable (Figure 8). It
indicates that the proportion of AI using other irriga-
tion methods is increased annually in WLT, which has
a good eect on saving irrigation water and control-
ling soil salinity (Mao et al. 2017). On the other hand,
the changes in irrigation extent and timing will lead to
corresponding changes in salinity and moisture
Figure 12. Proportions of irrigated areas from September 15th to October 15th and from October 15th to November 15th from 2010 to
2020 and their corresponding trend lines.
1614 X. QIAN ET AL.
distribution (Li et al. 2012). The multi-year water-salt
balance in the HID will be disrupted, requiring arti-
cial adjustment of irrigation strategies.
Our results can be used as a reference for irrigation
and agricultural management in the HID. First, it can
provide AI information in near-real-time (by the end
of AI season) to guide spring irrigation and AI in the
coming year. Second, the results provide information
on croplands that are not irrigated in autumn, which
are useful for determining the main planting areas for
sunowers and guiding the rational adjustment of the
cropping patterns to prevent the single planted vari-
ety from aecting the sustainable development of
local agricultural production (J. Liu et al. 2015).
The present model has been applied successfully in
the HID. However, because it is an initial exploration
in identifying the extent and timing of AI, the model
still has some limitations and needs to be further
improved.
To better detect the AI time, MODIS data with the
daily temporal resolution is used. However, the mod-
erate spatial resolution of MODIS data may inuence
the identication precision because of the inevitable
problem of mixed pixels (Wardlow and Callahan
2014). The proportion of AI in HID is high, exceeding
70% of the crop planting area. Furthermore, the irri-
gated land is basically in large blocks, which provides
a better opportunity to use MODIS data to identify the
irrigated area. Although the model performs well in
the entire HID and four of ve sub-irrigation districts,
its performance in WLBH is not as good as in other
districts. This is mainly because the distribution of
cropland in WLBH is more fragmented and MODIS
cannot capture it well. The selection of remote sen-
sing data with appropriate temporal and spatial reso-
lutions depends mainly on the distribution of
cropland in the irrigation district and research pur-
poses (Velpuri et al. 2009). Generally, products whose
spatial resolution matches the cropland scale for irri-
gation application should be used (Dari et al. 2022;
Massari et al. 2021). However, we often need data
with high temporal resolution (e.g. daily scale) that
may not have the high spatial resolution to match the
irrigation scale (Durgun et al. 2020). In such cases, the
percentage of cropland in each pixel can be consid-
ered in the classication in future studies or an appro-
priate method for irrigation land fraction estimation
in pixel scale can be further developed (W. Liu et al.
2018; Zhang, Chen, and Lu 2015).
Another source of uncertainty may be associated
with the threshold, which is also an unavoidable pro-
blem of the threshold-based method. The key is how
to choose the threshold value. In some previous stu-
dies using MODIS data for identifying irrigation, the
thresholds came from typical sample plots (L. Wang
et al. 2021) that require the matching between sam-
ple purity and actual purity of the classication target;
otherwise, it may lead to misclassication (Y. Chen
et al. 2016). Therefore, we chose the conservative
calibration method with statistical data to obtain the
threshold. Second, the thresholds are normally not
stable over dierent years. Compared with a xed
threshold, using a exible Z-score normalized thresh-
old can make the model robust enough to be applied
in dierent years (B. Chen, Jin, and Brown 2019).
However, we still need to be cautious in applying
the thresholds to changed environmental conditions
in the future. Besides, MBWI thresholds need to be
adjusted for other irrigation districts with dierent
management practices, irrigation volumes, soil char-
acteristics, and cropping patterns.
Inaccuracies in land-use products can also lead to
uncertainties in the identication of irrigated land.
Identifying possible cultivated land is the rst step in
mapping irrigated land (Jin et al. 2016). Usually, the
extraction of potential cultivated land depends on the
existing land cover maps (Xie et al. 2019). However,
the classied maps were generally for country or even
global scales, which is seldom independently vali-
dated for a specied irrigation district (Gumma et al.
2020). Moreover, the temporal resolution of these
products is relatively low, and the updating speed of
these products generally cannot keep up with the
speed of actual land-use change (Pan et al. 2020).
According to the land-use maps for 2010 and 2020
(Figure 3), the distribution of lakes in WLBH has chan-
ged signicantly from 2010 to 2020. Since lakes are
possibly confused with the phenomenon of AI, land-
use maps in only two years cannot eliminate the
inuences of lakes in other years. Therefore, precision
land-use maps for the study region on an annual basis
are benecial for high precision in AI mapping.
For accuracy assessment, the mixed pixel problem
complicates the evaluation of the nal mapping accu-
racy. A threshold of 50% for AI or NAI area proportion
was chosen to extract MODIS validation pixels based
on sampling points and Sentinel-2 images with ner
scales, which is a common practice in binary
GISCIENCE & REMOTE SENSING 1615
classication with mixed pixels (Hao et al. 2022).
However, the most appropriate proportions for dier-
ent regions need to be further explored (Colditz et al.
2018). In future studies, AI classication results con-
sidering the percentage of cropland in each pixel are
useful to study mapping accuracies at dierent levels
of cropland abundance, which can also provide useful
insights to nd the most appropriate proportion to
identify a MODIS pixel as an AI pixel in the study
region.
Moreover, the data gap of Sentinel-2 may aect the
evaluation results on irrigation time. Studies have
shown that the number of images involved is crucial
when detecting objects with spatial-temporal hydro-
logical uctuations (Borro et al. 2014). The principles
we set for determining the irrigation time for samples
may not be applicable to all samples due to data gaps.
Therefore, the optimal number of images for valida-
tion sample selection should also be considered in the
future, which may aid in achieving a realistic and
more accurate assessment of the model.
6. Conclusions
Knowing exactly when and where irrigation is applied
is essential for eective irrigation water management.
Autumn irrigation has been adopted in many regions
of the world but is usually less investigated than
irrigation during the growing season. A new approach
was developed to map AI area and timing distribu-
tions using MODIS MBWI time series in the Hetao
Irrigation District (HID) of Northwest China. This
study addressed three key technical issues, i.e. selec-
tion of an appropriate WI for the freezing period,
temporal incomparability of thresholds, and diculty
in validating the identied irrigation results, especially
irrigation timing. The model is simple but eective
and suitable for rapidly updating AI lands in large
irrigation districts, especially for irrigated cropland
distributed in large blocks. The generated maps
from 2010 to 2020 for HID with a total land area of
1.189 million ha and an irrigated land area of
0.733 million ha delivered reasonable performance.
The overall accuracy of irrigated area is over 90%,
whereas the overall accuracies of irrigated dates
assessed with “MTLP” and “WP” are 76.4% and
91.7%, respectively. The proposed mapping metho-
dology and associated datasets provide the rst
detailed AI extent and timing distribution dataset for
HID to date and can help reveal new insights into
irrigated land use and water use change.
Based on the mapping results of 11 years of
data, although the AI area has increased or
decreased in some regions, there is no signicant
dierence in the annual spatial distributions of AI
extent compared to the signicant dierence in
the annual spatial distribution of AI time in HID.
Moreover, the area of AI tends to decline in the
whole HID, and the starting time of AI is gradually
delayed. These changes are related to many fac-
tors, including climate, irrigation fees, and changes
in cropping patterns. The results obtained in this
study can provide a reference for the rational for-
mulation of AI scheduling under water-saving irri-
gation conditions and can provide a basis for
reducing secondary salinization and promoting
agricultural production in the study area.
Acknowledgements
We are grateful to Dr. Khalil Ur Rahman from Tsinghua
University for proofreading the manuscript. We are also grate-
ful to the editors and anonymous reviewers, whose comments
and suggestions are helpful to improve the quality of the
manuscript.
Disclosure statement
No potential conict of interest was reported by the author(s).
Funding
This work was supported by the National Natural Science
Foundation of China [52279038, 51839006, and 51779119];
Research Program of the State Key Laboratory of Hydroscience
and Engineering, Tsinghua University [2020-KY-01].
Data availability
The data that support the ndings of this study are available on
request from the corresponding author.
References
Bazzi, H., N. Baghdadi, D. Ienco, M. El Hajj, M. Zribi,
H. Belhouchette, M. J. Escorihuela, and V. Demarez. 2019.
“Mapping Irrigated Areas Using Sentinel-1 Time Series in
Catalonia, Spain.” Remote Sensing 11 (15): 1836. doi:10.
3390/rs11151836.
1616 X. QIAN ET AL.
Borro, M., N. Morandeira, M. Salvia, P. Minotti, P. Perna, and
P. Kandus. 2014. “Mapping Shallow Lakes in A Large South
American Floodplain: A Frequency Approach on
Multitemporal Landsat TM/ETM Data.” Journal of Hydrology
512: 39–52. doi:10.1016/j.jhydrol.2014.02.057.
Boschetti, M., F. Nutini, G. Manfron, P. Brivio, and A. Nelson.
2014. “Comparative Analysis of Normalised Dierence
Spectral Indices Derived from MODIS for Detecting Surface
Water in Flooded Rice Cropping Systems.” PLoS One 9 (2):
e88741. doi:10.1371/journal.pone.0088741.
Chen, B., Y. Jin, and P. Brown. 2019. “Automatic Mapping of
Planting Year for Tree Crops with Landsat Satellite Time
Series Stacks.” ISPRS Journal of Photogrammetry and
Remote Sensing 151: 176–188. doi:10.1016/j.isprsjprs.2019.
03.012.
Chen, Y., D. Lu, L. Luo, Y. Pokhrel, K. Deb, J. Huang, and Y. Ran.
2018. “Detecting Irrigation Extent, Frequency, and Timing in
a Heterogeneous Arid Agricultural Region Using MODIS
Time Series, Landsat Imagery, and Ancillary Data.” Remote
Sensing of Environment 204 (October 2017): 197–211. doi:10.
1016/j.rse.2017.10.030.
Chen, Y., X. Song, S. Wang, J. Huang, and L. R. Mansaray. 2016.
“Impacts of Spatial Heterogeneity on Crop Area Mapping in
Canada Using MODIS Data.” ISPRS Journal of
Photogrammetry and Remote Sensing 119: 451–461. doi:10.
1016/j.isprsjprs.2016.07.007.
Colditz, R. R., C. Troche Souza, B. Vazquez, A. J. Wickel, and
R. Ressl. 2018. “Analysis of Optimal Thresholds for
Identication of Open Water Using MODIS-Derived Spectral
Indices for Two Coastal Wetland Systems in Mexico.”
International Journal of Applied Earth Observation and
Geoinformation 70: 13–24. doi:10.1016/j.jag.2018.03.008.
Dao, P. D., N. T. Mong, and H.-P. Chan. 2019. “Landsat-MODIS
Image Fusion and Object-Based Image Analysis for
Observing Flood Inundation in a Heterogeneous Vegetated
Scene.” GIScience & Remote Sensing 56 (8): 1148–1169.
doi:10.1080/15481603.2019.1627062.
Dari, J., L. Brocca, P. Quintana-Seguí, S. Casadei,
M. J. Escorihuela, V. Stefan, and R. Morbidelli. 2022. “Double-
Scale Analysis on the Detectability of Irrigation Signals from
Remote Sensing Soil Moisture over an Area with Complex
Topography in Central Italy.” Advances in Water Resources
161: 104130. doi:10.1016/j.advwatres.2022.104130.
Dari, J., P. Quintana-Seguí, M. J. Escorihuela, V. Stefan, L. Brocca,
and R. Morbidelli. 2021. “Detecting and Mapping Irrigated
Areas in a Mediterranean Environment by Using Remote
Sensing Soil Moisture and a Land Surface Model.” Journal
of Hydrology 596: 126129. doi:10.1016/j.jhydrol.2021.
126129.
De Carvalho Júnior, O. A., R. F. Guimarães, C. R. Silva, and
R. A. T. Gomes. 2015. “Standardized Time-Series and
Interannual Phenological Deviation: New Techniques for
Burned-Area Detection Using Long-Term MODIS-NBR
Dataset.” Remote Sensing 7 (6): 6950–6985. doi:10.3390/
rs70606950.
Deines, J. M., A. D. Kendall, M. A. Crowley, J. Rapp, J. A. Cardille,
and D. W. Hyndman. 2019. “Mapping Three Decades of
Annual Irrigation across the US High Plains Aquifer Using
Landsat and Google Earth Engine.” Remote Sensing of
Environment 233: 111400. doi:10.1016/j.rse.2019.111400.
DeVries, B., C. Huang, J. Armston, W. Huang, J. W. Jones, and
M. W. Lang. 2020. “Rapid and Robust Monitoring of Flood
Events Using Sentinel-1 and Landsat Data on the Google
Earth Engine.” Remote Sensing of Environment 240: 111664.
doi:10.1016/j.rse.2020.111664.
Durgun, Y. Ö., A. Gobin, G. Duveiller, and B. Tychon. 2020.
“A Study on Trade-Os between Spatial Resolution and
Temporal Sampling Density for Wheat Yield Estimation
Using Both Thermal and Calendar Time.” International
Journal of Applied Earth Observation and Geoinformation
86: 101988. doi:10.1016/j.jag.2019.101988.
Elwan, E., M. Le Page, L. Jarlan, N. Baghdadi, L. Brocca,
S. Modanesi, J. Dari, P. Quintana Seguí, and M. Zribi. 2022.
“Irrigation Mapping on Two Contrasted Climatic Contexts
Using Sentinel-1 and Sentinel-2 Data.” Water 14 (5): 804.
doi:10.3390/w14050804.
Feyisa, G. L., H. Meilby, R. Fensholt, and S. R. Proud. 2014.
“Automated Water Extraction Index: A New Technique for
Surface Water Mapping Using Landsat Imagery.” Remote
Sensing of Environment 140: 23–35. doi:10.1016/j.rse.2013.
08.029.
Fisher, A., N. Flood, and T. Danaher. 2016. “Comparing Landsat
Water Index Methods for Automated Water Classication in
Eastern Australia.” Remote Sensing of Environment 175:
167–182. doi:10.1016/j.rse.2015.12.055.
Gao, Q., M. Zribi, M. J. Escorihuela, N. Baghdadi, and P. Q. Segui.
2018. “Irrigation Mapping Using Sentinel-1 Time Series at
Field Scale.” Remote Sensing 10 (9): 1495. doi:10.3390/
rs10091495.
Gumma, M. K., P. S. Thenkabail, P. G. Teluguntla, A. Oliphant,
J. Xiong, C. Giri, V. Pyla, S. Dixit, and A. M. Whitbread. 2020.
“Agricultural Cropland Extent and Areas of South Asia
Derived Using Landsat Satellite 30-m Time-Series Big-Data
Using Random Forest Machine Learning Algorithms on the
Google Earth Engine Cloud.” GIScience & Remote Sensing
57 (3): 302–322. doi:10.1080/15481603.2019.1690780.
Hao, X., G. Huang, Z. Zheng, X. Sun, W. Ji, H. Zhao, J. Wang, H. Li,
and X. Wang. 2022. “Development and Validation of a New
MODIS Snow-Cover-Extent Product over China.” Hydrology
and Earth System Sciences 26 (8): 1937–1952. doi:10.5194/
hess-26-1937-2022.
Hu, S. S., C. X. Huang, B. Yang, and Z. F. Tan. 2020.
“Reconstruction of MODIS Vegetation Index Time Series by
Adaptive Weighted Savitzky-Golay Filter.” Science of Surveying
and Mapping 45 (4) 105–116. In Chinese with English abstract.
Jeong, S., S. Kang, K. Jang, H. Lee, S. Hong, and D. Ko. 2012.
“Development of Variable Threshold Models for Detection of
Irrigated Paddy Rice Fields and Irrigation Timing in
Heterogeneous Land Cover.” Agricultural Water
Management 115: 83–91. doi:10.1016/j.agwat.2012.08.012.
GISCIENCE & REMOTE SENSING 1617
Jing, M., H. M. Zhang, J. Yang, Y. M. Fan, F. G. Huang, and
L. Chen. 2020. “Study of Deep Water Saving and Control in
Ning-Meng Irrigation Area.” Yellow River 42 (9) 155–160. In
Chinese with English abstract.
Jin, N., B. Tao, W. Ren, M. Feng, S. Rui, L. He, W. Zhuang, and
Q. Yu. 2016. “Mapping Irrigated and Rainfed Wheat Areas
Using Multi-Temporal Satellite Data.” Remote Sensing 8 (3):
207. doi:10.3390/rs8030207.
Ji, L., L. Zhang, and B. Wylie. 2009. “Analysis of Dynamic
Thresholds for the Normalized Dierence Water Index.”
Photogrammetric Engineering & Remote Sensing 75 (11):
1307–1317. doi:10.14358/PERS.75.11.1307.
Karthikeyan, L., I. Chawla, and A. K. Mishra. 2020. “A Review of
Remote Sensing Applications in Agriculture for Food
Security: Crop Growth and Yield, Irrigation, and Crop
Losses.” Journal of Hydrology 586 (March): 124905. doi:10.
1016/j.jhydrol.2020.124905.
Klein, A. G., and A. C. Barnett. 2003. “Validation of Daily MODIS
Snow Cover Maps of the Upper Rio Grande River Basin for
the 2000–2001 Snow Year.” Remote Sensing of Environment
86 (2): 162–176. doi:10.1016/S0034-4257(03)00097-X.
Kreyszig, E. 2010. Advanced Engineering Mathematics. Hoboken,
New Jersey: John Wiley & Sons.
Li, R., H. Shi, G. N. Flerchinger, T. Akae, and C. Wang. 2012.
“Simulation of Freezing and Thawing Soils in Inner Mongolia
Hetao Irrigation District, China.” Geoderma 173–174: 28–33.
doi:10.1016/j.geoderma.2012.01.009.
Liu, W., J. Dong, K. Xiang, S. Wang, W. Han, and W. Yuan. 2018.
“A Sub-Pixel Method for Estimating Planting Fraction of
Paddy Rice in Northeast China.” Remote Sensing of
Environment 205 (September 2017): 305–314. doi:10.1016/j.
rse.2017.12.001.
Liu, J., S. Sun, P. Wu, Y. Wang, and X. Zhao. 2015. “Evaluation of
Crop Production, Trade, and Consumption from the
Perspective of Water Resources: A Case Study of the Hetao
Irrigation District, China, for 1960–2010.” Science of the Total
Environment 505: 1174–1181. doi:10.1016/j.scitotenv.2014.
10.088.
Lu, X., R. Li, H. Shi, J. Liang, Q. Miao, and L. Fan. 2019.
“Successive Simulations of Soil Water-Heat-Salt Transport
in One Whole Year of Agriculture after Dierent Mulching
Treatments and Autumn Irrigation.” Geoderma 344: 99–107.
doi:10.1016/j.geoderma.2019.03.006.
Macfarlane, C., A. H. Grigg, and M. I. Daws. 2017.
“A Standardised Landsat Time Series (1973–2016) of Forest
Leaf Area Index Using Pseudoinvariant Features and
Spectral Vegetation Index Isolines and A Catchment
Hydrology Application.” Remote Sensing Applications:
Society and Environment 6 (January): 1–14. doi:10.1016/j.
rsase.2017.01.006.
Mao, W., J. Yang, Y. Zhu, M. Ye, and J. Wu. 2017. “Loosely
Coupled SaltMod for Simulating Groundwater and Salt
Dynamics under Well-Canal Conjunctive Irrigation in
Semi-Arid Areas.” Agricultural Water Management 192:
209–220. doi:10.1016/j.agwat.2017.07.012.
Massari, C., S. Modanesi, J. Dari, A. Gruber, G. J. M. De Lannoy,
M. Girotto, P. Quintana-Seguí, et al. 2021. “A Review of
Irrigation Information Retrievals from Space and Their
Utility for Users.” Remote Sensing 13 (20): 4112. doi:10.
3390/rs13204112.
Myers, D. T., D. L. Ficklin, S. M. Robeson, R. P. Neupane, A. Botero-
Acosta, and P. M. Avellaneda. 2021. “Choosing an Arbitrary
Calibration Period for Hydrologic Models: How Much Does It
Inuence Water Balance Simulations?” Hydrological Processes
35 (2): e14045. doi:10.1002/hyp.14045.
Negri, C., E. Chiaradia, M. Rienzner, A. Mayer, C. Gandol,
M. Romani, and A. Facchi. 2020. “On the Eects of Winter
Flooding on the Hydrological Balance of Rice Areas in
Northern Italy.” Journal of Hydrology 590: 125401. doi:10.
1016/j.jhydrol.2020.125401.
Ozdogan, M., Y. Yang, G. Allez, and C. Cervantes. 2010. “Remote
Sensing of Irrigated Agriculture: Opportunities and
Challenges.” Remote Sensing 2 (9): 2274–2304. doi:10.3390/
rs2092274.
Pan, H., X. Tong, X. Xu, X. Luo, Y. Jin, H. Xie, and B. Li. 2020.
“Updating of Land Cover Maps and Change Analysis Using
GlobeLand30 Product: A Case Study in Shanghai
Metropolitan Area, China.” Remote Sensing 12 (19): 3147.
doi:10.3390/rs12193147.
Pervez, M. S., and J. F. Brown. 2010. “Mapping Irrigated Lands at
250-m Scale by Merging MODIS Data and National
Agricultural Statistics.” Remote Sensing 2 (10): 2388–2412.
doi:10.3390/rs2102388.
Pervez, M. S., M. Budde, and J. Rowland. 2014. “Mapping
Irrigated Areas in Afghanistan over the past Decade Using
MODIS NDVI.” Remote Sensing of Environment 149: 155–165.
doi:10.1016/j.rse.2014.04.008.
Sab, N. S. C. 2014. “Study on Stability of Reectance
Characteristics of Natural Features for Calibrating Remote
Sensing Data.” PhD diss., Universiti Teknologi Malaysia.
Salmon, J. M., M. Friedl, S. Frolking, D. Wisser, and E. Douglas.
2015. “Global Rain-Fed, Irrigated, and Paddy Croplands:
A New High Resolution Map Derived from Remote Sensing,
Crop Inventories and Climate Data.” International Journal of
Applied Earth Observation and Geoinformation 38: 321–334.
doi:10.1016/j.jag.2015.01.014.
Santana, N. C., O. A. De Carvalho Júnior, R. A. T. Gomes, and
R. F. Guimarães. 2018. “Burned-Area Detection in Amazonian
Environments Using Standardized Time Series per Pixel in
MODIS Data.” Remote Sensing 10 (12): 1904. doi:10.3390/
rs10121904.
Satterwhite, M. B., H. J. Mitchell, T. Hemmer, and J. D. Leckie. 2003.
“Field Spectral Signatures of Snow, Ice, and Water.” Algorithms
and Technologies for Multispectral, Hyperspectral, and
Ultraspectral Imagery IX 5093: 528. doi:10.1117/12.488363.
Seaton, D., T. Dube, and D. Mazvimavi. 2020. “Use of
Multi-Temporal Satellite Data for Monitoring Pool Surface
Areas Occurring in Non-Perennial Rivers in Semi-Arid
Environments of the Western Cape, South Africa.” ISPRS
Journal of Photogrammetry and Remote Sensing 167:
375–384. doi:10.1016/j.isprsjprs.2020.07.018.
Shazadeh-Moghadam, H., M. Khazaei, S. K. Alavipanah, and
Q. Weng. 2021. “Google Earth Engine for Large-Scale Land
Use and Land Cover Mapping: An Object-Based
1618 X. QIAN ET AL.
Classication Approach Using Spectral, Textural and
Topographical Factors.” GIScience & Remote Sensing 58 (6):
914–928. doi:10.1080/15481603.2021.1947623.
Shao, Y., R. S. Lunetta, B. Wheeler, J. S. Iiames, and
J. B. Campbell. 2016. “An Evaluation of Time-Series
Smoothing Algorithms for Land-Cover Classications Using
MODIS-NDVI Multi-Temporal Data.” Remote Sensing of
Environment 174: 258–265. doi:10.1016/j.rse.2015.12.023.
Shi, H. B., S. Q. Yang, R. P. Li, X. Y. Li, W. P. Li, J. W. Yan, Q. F. Miao,
and Z. Li. 2020. “Water-Saving Irrigation and Utilization
Eciency of Water and Fertilizer in Hetao Irrigation District
of Inner Mongolia: Prospect for Future Research.” Journal of
Irrigation and Drainage 39 (11): 1–12. In Chinese with English
abstract. doi:10.13522/j.cnki.ggps.2020155
Thompson, J. A., and D. J. Paull. 2017. “Assessing Spatial and
Temporal Patterns in Land Surface Phenology for the
Australian Alps (2000–2014).” Remote Sensing of
Environment 199: 1–13. doi:10.1016/j.rse.2017.06.032.
Tsela, P., K. Wessels, J. Botai, S. Archibald, D. Swanepoel,
K. Steenkamp, and P. Frost. 2014. “Validation of the Two
Standard MODIS Satellite Burned-Area Products and an
Empirically-Derived Merged Product in South Africa.”
Remote Sensing 6 (2): 1275–1293. doi:10.3390/rs6021275.
Velpuri, N. M., P. S. Thenkabail, M. K. Gumma, C. Biradar,
V. Dheeravath, P. Noojipady, and L. Yuanjie. 2009.
“Inuence of Resolution in Irrigated Area Mapping and
Area Estimation.” Photogrammetric Engineering and Remote
Sensing 75 (12): 1383–1395. doi:10.14358/PERS.75.12.1383.
Vermote, E. F., J.-C. Roger, and J. P. Ray. 2015. ““MODIS Surface
Reectance User’s Guide.” Collection 6 MayRetrieved
fromhttp://modis-sr.ltdri.org/guide/MOD09_UserGuide_v1.
4.pdf
Wang, L. X., M. Y. Cai, X. L. Bai, S. X. JI, B. Jia, X. W. Feng, and
X. L. Chang. 2021. “Estimation of Irrigation Area in Hetao
Irrigation District Based on MODIS Analysis.” Journal of Inner
Mongolia Normal University(Natural Science Edition) 50 (1):
44–52. In Chinese with English abstract. doi:10.3969/j.1001-
8735.2021.01.007
Wang, X., S. Xie, X. Zhang, C. Chen, H. Guo, J. Du, and Z. Duan.
2018. “A Robust Multi-Band Water Index (MBWI) for Automated
Extraction of Surface Water from Landsat 8 OLI Imagery.”
International Journal of Applied Earth Observation and
Geoinformation 68: 73–91. doi:10.1016/j.jag.2018.01.018.
Wang, R., C. Zeng, P. Li, and H. Shen. 2011. “Terra MODIS Band 5
Stripe Noise Detection and Correction Using MAP-Based
Algorithm.” In 2011 International Conference on Remote
Sensing, Environment and Transportation Engineering,
Nanjing, 8612–8615. doi:10.1109/RSETE.2011.5964181.
Wardlow, B. D., and K. Callahan. 2014. “A Multi-Scale Accuracy
Assessment of the MODIS Irrigated Agriculture Data-Set
(Mirad) for the State of Nebraska, USA.” GIScience & Remote
Sensing 51 (5): 575–592. doi:10.1080/15481603.2014.952546.
Wen, Y., S. Shang, and K. U. Rahman. 2019. “Pre-Constrained
Machine Learning Method for Multi-Year Mapping of Three
Major Crops in a Large Irrigation District.” Remote Sensing
11 (3): 242. doi:10.3390/rs11030242.
Wen, Y., S. Shang, K. U. Rahman, Y. Xia, and D. Ren. 2020. “A
Semi-Distributed Drainage Model for Monthly Drainage
Water and Salinity Simulation in A Large Irrigation District
in Arid Region.” Agricultural Water Management 230: 105962.
doi:10.1016/j.agwat.2019.105962.
Wen, Y., H. Wan, S. Shang, and K. U. Rahman. 2022. “A Monthly
Distributed Agro-Hydrological Model for Irrigation District in
Arid Region with Shallow Groundwater Table.” Journal of
Hydrology 609: 127746. doi:10.1016/j.jhydrol.2022.127746.
Xie, Y., and T. J. Lark. 2021. “Mapping Annual Irrigation from
Landsat Imagery and Environmental Variables across the
Conterminous United States.” Remote Sensing of Environment
260 (April): 112445. doi:10.1016/j.rse.2021.112445.
Xie, Y., T. J. Lark, J. F. Brown, and H. K. Gibbs. 2019. “Mapping
Irrigated Cropland Extent across the Conterminous United
States at 30 m Resolution Using a Semi-Automatic Training
Approach on Google Earth Engine.” ISPRS Journal of
Photogrammetry and Remote Sensing 155: 136–149. doi:10.
1016/j.isprsjprs.2019.07.005.
Xiong, L., X. Xu, B. Engel, Q. Huang, Z. Huo, Y. Xiong, W. Han, and
G. Huang. 2021. “Modeling Agro-Hydrological Processes and
Analyzing Water Use in a Super-Large Irrigation District (Hetao)
of Arid Upper Yellow River Basin.” Journal of Hydrology 603:
127014. doi:10.1016/j.jhydrol.2021.127014.
Xu, X., G. Huang, Z. Qu, and L. S. Pereira. 2010. “Assessing the
Groundwater Dynamics and Impacts of Water Saving in the
Hetao Irrigation District, Yellow River Basin.” Agricultural Water
Management 98 (2): 301–313. doi:10.1016/j.agwat.2010.08.025.
Yang, L. Q., Y. F. Liu, and K. Liu. 2017. “The Causes of the
Increase of Spring Irrigation Area in Yichang Irrigation
District.” Inner Mongolia Water Resour 4: 66–67. In Chinese.
Yeom, J.-M., S. Jeong, R. C. Deo, and J. Ko. 2021. “Mapping Rice
Area and Yield in Northeastern Asia by Incorporating a Crop
Model with Dense Vegetation Index Proles from
a Geostationary Satellite.” GIScience & Remote Sensing
58 (1): 1–27. doi:10.1080/15481603.2020.1853352.
Yu, B., and S. Shang. 2017. “Multi-Year Mapping of Maize and
Sunower in Hetao Irrigation District of China with High
Spatial and Temporal Resolution Vegetation Index Series.”
Remote Sensing 9 (8): 855. doi:10.3390/rs9080855.
Zeng, Q. Z., M. S. Cao, X. Z. Feng, F. X. Liang, X. Z. Chen, and
W. K. Sheng. 1984. “Study on Spectral Reection
Characteristics of Snow, Ice and Water of Northwest
China.” Scientia Sinica, Series B 27 (6): 647–656.
Zhang, C., Y. Chen, and D. Lu. 2015. “Detecting Fractional
Land-Cover Change in Arid and Semiarid Urban
Landscapes with Multitemporal Landsat Thematic Mapper
Imagery.” GIScience & Remote Sensing 52 (6): 700–722. doi:10.
1080/15481603.2015.1071965.
Zhan, Z. M., Q. M. Qin, A. Ghulan, and D. D. Wang. 2006. “NIR-
red Spectral Space Based New Method for Soil Moisture
Monitoring.” Scientia Sinica Terrae 36 (11) 46–52. In Chinese.
GISCIENCE & REMOTE SENSING 1619
Appendix A1. Proportions of valid observations from dierent remote sensing data sources
Proportions of valid observations from the three remote sensing data sources, i.e. Sentinel-2, MOD09GA, and MCD43A4, for
each day in the 2019 study period are shown in Figure 13. More than 60% of the dates contain more than 70% of MOD09GA data
that is usable. MCD43A4 has a lower percentage of available data since mid-October and has almost no data from November 9th,
which may result in a high degree of incomplete information in classication maps. Sentinel-2 data can be used for validation due
to its high spatial resolution (10m) and high percentage of data available on certain dates.
Appendix A2. Appropriate water index for mapping irrigated areas
The simplicity of the two-band indices makes them easy to use. Previous studies have shown that when using MODIS data for
calculation, normalized dierence water indices using bands 4 and 7 (NDWI47), 1 and 7 (NDWI17), 4 and 6 (NDWI46), 4 and 5
(NDWI45), or 1 and 6 (NDWI16) perform better (Boschetti et al. 2014; Ji, Zhang, and Wylie 2009). Among them, some studies have
found that NDWI46 has the best performance in water body recognition. It has a low magnitude of error and a stable threshold, and
performs well on mixed pixels (Colditz et al. 2018). Meanwhile, due to the repetitive stripes, we avoid using band 5 (R. Wang et al.
2011). Finally, NDWI46 and NDWI17 are selected that are more sensitive than other two-band WIs in low water percentage
conditions (Boschetti et al. 2014).
Multi-band indices considering more band information may be more suitable to distinguish water bodies when there are
many surfaces that can be easily confused with a water surface. Therefore, we also chose several multi-band indices
proposed in recent years, including MBWI (X. Wang et al. 2018), WI2015 (Fisher, Flood, and Danaher 2016), AWEIsh and
AWEInsh (Feyisa et al. 2014). In addition, a soil moisture index, SMMRS proposed by Zhan et al. (2006) is also compared with
these surface water indices. Six WIs and one soil moisture index used for comparison in this study are presented in Table A1,
which can be broadly divided into two types, i.e. two-band and multi-band indices.
Table A1. List of the seven indices compared in this study. The input is surface reectance (ρ) for bands b1 (ρb1) to b7 (ρb7)
derived from the MODIS multi-spectral bands.
In order to show the dierence of indices clearly in the graph, we normalize the indices which are not in the range of [−1, 1] to
[0, 1] by Equation (A1).
Figure 13. Time series of the proportion of pixels with no missing value in any of the bands used to calculate water index after
preliminary quality control during the study period in 2019 (DOY258–DOY334). DOY represents day of the Year.
Index Equation
NDWI46 ρb4ρb6ð Þ=ðρb4þρb6Þ
NDWI17 ρb1ρb7ð Þ=ðρb1þρb7Þ
MBWI 2 ρb4ρb1ρb2ρb6ρb7Þ
AWELnsh 4 ρb4ρb6ð Þ=ð0:25 ρb2þ2:75 ρb7Þ
AWELsh ρb3þ2:5ρb41:5 ðρb2þρb6Þ 0:25 ρb7
WI2015 1:7204 þ171 ρb4þ3ρb170 ρb245 ρb671 ρb7
SMMRS 1ρb1þMρb2ð Þ=ffiffiffiffiffiffiffiffiffiffiffiffiffiffi
1þM2
p
Note: M is the slope of the soil line, which is extracted using the near-infrared (NIR) and red bands reflectance of the study area (Zhan
et al. 2006).
1620 X. QIAN ET AL.
x¼xxmin
xmax xmin
(A1)
where x and x* are the original and normalized indices, and x
min
and x
max
are minimum and maximum values of the x series,
respectively.
By comparing the time proles of all sampling points, multi-band water indices are found more sensitive to changes in water
content. They usually have a more signicant increase after autumn irrigation, exhibiting an improvement than the two-band
normalized-dierence water indices (Figure 14). However, the inclusion of more bands also leads to more uctuations in the multi-
band water indices.
In addition, dierent water indices are increased with varying magnitudes in late November (around DOY 325 for the
representative irrigated pixel in Figure 14), possibly due to snow and/or ice. Among them, the increase of MBWI is small and
usually lower than the increase caused by irrigation, while it even does not increase when other indices have abrupt changes in
some cases. This is mainly because the number of absorption bands of ice and snow is similar to that of water, but their maximum
absorption is dierent from that of water (Zeng et al. 1984; Satterwhite et al. 2003). The weights of MBWI in the visible and near-
infrared bands reduce the inuence of ice and snow to some extent. In view of this, we nally chose MBWI to extract irrigation
information.
Appendix A3. Comparison of average feature values of dierent references
The reference should be selected to eliminate the eects of viewing angle, solar zenith angle, data pre-processing, and other
factors on the feature values. It should provide an objective baseline to measure feature values that can be related to autumn
irrigation in a given year.
Persistent water bodies are relatively stable in terms of reectance characteristics and experience little variations over the same
period of multiple years (Sab 2014). Therefore, they can be considered pseudo-invariant features to be used as the normalization
reference (Macfarlane, Grigg, and Daws 2017). However, not all sub-irrigation districts have enough water pixels that can be used as
references. Due to the fact that autumn-irrigated cropland pixels are covered by water during AI, they can be used as a proxy for
water bodies. By calculating the average of the feature values of all input croplands and grasslands pixels in the irrigation district, it
is found that the uctuation is also consistent with the uctuation of the water body pixels (Figure 15). It shows that in large-scale
irrigation districts when the irrigation ratio is high every year, the average value can reect the inherent uctuation of the index and
can be used as a reference to remove the inter-annual uctuations. Furthermore, since the proportion of AI areas in some other
irrigation districts may be small, pixels whose MMAVE-MBWI value ranks between 70% and 100% of all cropland and grassland
pixels were also selected. Figure 15 shows the inter-annual uctuations in the average MMAVE-MBWI for the above three
references. Finally, we selected a simple average of all the input data, because it is accurate enough to reect the inter-annual
changes of feature values in the study area due to factors unrelated to the study objectives. This method of computing baseline
statistics ensures that Z-scores are inuenced by autumn irrigation only.
Figure 14. Time profile of seven different indices for one representative irrigated pixel in the study area in 2019. DOY represents day of
the Year.
GISCIENCE & REMOTE SENSING 1621
Appendix A4. Validation points extracted from Sentinel-2 images from 2016 to 2020
As shown in Figure 16, a total of 3789 autumn irrigated points are selected based on the threshold method and 2694 non-autumn
irrigated points by visual interpretation from 2016 to 2020.
Figure 16. Validation points extracted from Sentinel-2 images from 2016 to 2020, including Autumn irrigation (AI) points (left) and
Non-Autumn irrigation (NAI) points (right).
Figure 15. The average values of the feature variable (MMAVE-MBWI) for water bodies, cropland-grassland, and part of cropland-
grassland with higher MMAVE-MBWI in the whole irrigation district from 2010 to 2020.
1622 X. QIAN ET AL.
Appendix A5. Cumulative areas of autumn irrigation in years from 2010 to 2020
Based on the evolution processes of the autumn irrigation area in dierent years, the cumulative areas of autumn irrigation from
2010 to 2020 was shown in Figure 17.
Figure 17. Cumulative areas of autumn irrigation in years from 2010 to 2020.
GISCIENCE & REMOTE SENSING 1623
Content uploaded by Songhao Shang
Author content
All content in this area was uploaded by Songhao Shang on Sep 27, 2022
Content may be subject to copyright.
