Content uploaded by Aqil Tariq
Author content
All content in this area was uploaded by Aqil Tariq on Jan 17, 2022
Content may be subject to copyright.
RESEARCH ARTICLE
Agricultural Field Extraction with Deep Learning Algorithm
and Satellite Imagery
Alireza Sharifi
1
•Hadi Mahdipour
2
•Elahe Moradi
1
•Aqil Tariq
3
Received: 19 May 2021 / Accepted: 2 December 2021
ÓIndian Society of Remote Sensing 2021
Abstract
Automatic detection of borders using remote sensing images will minimize the dependency on time-consuming manual
input. The lack of field border data sets indicates that current methods are ineffective. This article seeks to promote the
detection of field borders from satellite images with general process based on a multi-task segmentation model. ResUNet-a
is a convolutional neural network with a completely linked UNet backbone that supports sprawling and conditional
inference. The algorithm will significantly increase model efficiency and its generalization by re-constructing connected
outputs. Then individual field segmentation can be accomplished by post-processing model outputs. The model was
extremely exact in field mapping, field borders, and thus individual fields using the Sentinel-2 and Landsat-8 images as
inputs. The multitemporal images replacement with a single image similar to the composition time decreased slightly. The
proposed model is able to reliably identify field borders and remove irrelevant limits from the image to acquire complex
hierarchical contextual properties, thus outstriking classical edge filters. Our method is supposed to promote individual
crop field extraction on a scale, by minimizing overfitting.
Keywords Convolutional neural networks Edge detection Field borders Sentinel-2 Remote sensing
Introduction
Many of the automated agriculture promises support
farmers in monitoring agricultural fields during the season
of growth. With specific field border, a requirement for
field measurement has become essential and farmers are
often required to provide precise digital records of their
limits while they are signed up with the service provider
(Sharifi, 2018). This procedure is mainly manual and time
consuming and discourages people from doing so. Crop
yield forecasting and food safety control are also used for
predicting areas of crop yields through Earth Observation
programs (Paudel et al., 2021). The recurring extraction of
field borders through broad areas will benefit greatly in
such applications. Automation not only promotes the
inclusion of farmers and thus encourages more extensive
use of automated agricultural services, but also allows for
the provision by means of remote sensing (Tariq et al.,
2021). Several techniques were developed to extract field
borders from satellite images, which regularly and globally
cover high-resolution cropping areas. These methods are
divided into three models: edge-based models (Khan et al.,
2019), region-based models (Meher et al., 2019), and
hybrid models (Shi et al., 2020).
Edge-based methods are dependent on filters in an epoch
in which pixel values shift quickly to detect discontinuities.
Each filter describes a certain kernel that is converted into
emphasizing edges with an input image (Scharr, Sobel and
Canny operator are typical examples) (Kumar, Afzal, and
Afzal 2020). There are a variety of problems while dealing
with edge operators because their susceptibility in high
frequency sometimes produces misrepresentatives and their
parameterization is subjective and important to unique
&Alireza Sharifi
a_sharifi@sru.ac.ir
1
Department of Surveying Engineering, Faculty of Civil
Engineering, Shahid Rajaee Teacher Training University,
Tehran, Iran
2
Chief Innovation Office, Sinenta Corp., La Can
˜ada,
04120 Almeria, Spain
3
State Key Laboratory of Information Engineering in
Surveying Mapping and Remote Sensing (LIESMARS),
Wuhan University, Wuhan 430079, China
123
Journal of the Indian Society of Remote Sensing
https://doi.org/10.1007/s12524-021-01475-7(0123456789().,-volV)(0123456789().,-volV)
circumstances (Rabbi et al., 2020). Thresholds that are
post-processed and adapted locally can solve problems that
resulting in better established, closed borders. Region-
based methods are groups of adjacent entity pixels’
dependent on a certain criteria of homogeneity. The quest
for optimum segmentation for regional methods remains a
test and error procedure that is expected to produce inap-
propriate results. For example, objects with low parame-
terization might cease expanding, creating sliver polygons
and moving the extracted borders into them, before hitting
actual borders (Burdick et al., 2018). Methods dependent
on region often aim to oversegment areas with strong
internal variability and slight neighboring undersegment
areas (IdBenIdder & Laachfoubi, 2015; Sharifi, 2020). Any
of these adverse reactions can be mitigated by overseg-
menting images intentionally and determining whether
neighboring features are fused into machine learning (Park
et al., 2021). While edge-based and region-based methods
are available, the user group seems to have no use of these
methods and suggests a lack of fitness to do so. For
example, from crowdsourced, manually digitized polygons,
the only global field map is found (Ciobanu et al., 2019;S.
Wang et al., 2019).
Multitemporal image features appear redundant when
the borders of fields have been extracted from well-targeted
one-date images in certain situations. In addition, it is
impossible to produce a constant sequence of times in
places such as the tropics because of continuous cloud
coverage (Sudmanns et al., 2020). While multitemporal
data would likely improve their accuracy, particularly in
highly dynamic systems, we believe that large-scale field
border extraction and uptake are needed for models with
minimal preprocessing and parameter setting. Deep neural
networks have new ways of extracting field borders
because they need no designed features and since their
design is highly adaptable to new problems (Mohammadi
& Sharifi, 2021). Convolutional neural networks (CNNs)
are used more and more in image processing so they can
use hierarchical characteristics, from local to global images
(Krizhevsky et al., 2017). Their depth is based on a deep
network that uses convolutional operations. While filters
are hand-made in an edge-based technique, these filters can
learn from CNNs. These networks were originally devel-
oped for natural images and have been adapted for use in
remote sensing applications such as road extraction (Gao
et al., 2019), cloud detection (Chai et al., 2019), crop
identification (Wu et al., 2021), and river and water body
extraction (Guo et al., 2020). Thus, CNNs appear to be
particularly well-suited for extracting field borders based
on their size, though this has not been empirically
demonstrated.
The objective of this study is to use ResUNet-a, for field
border extraction. We also implemented the post-
processing methods, because CNNs can produce discon-
tinuous borders, that utilize their outputs to produce better
borders and find different field. By using a composite of
Sentinel-2 images with geometrical accuracy, we carried
out a series of experiments that show that overfitting has
been minimized without re-calibration to be implemented
in a variety of conditions.
Method and Materials
Field borders are extracted by labeling each pixel with
these classes: ’border’ or ’not ’border’. If groups of interest
are predictable, then training signals for similar learning
tasks that could contribute to the accuracy of the initial task
are ignored. Similar tasks can help a model better gener-
alize the initial task. In the last layer of the architecture,
these can also be combined to further restrict the inference
of the initially task stabilize graduated changes and thereby
increase model accuracy instead of forecasting all associ-
ated tasks concurrently and independently (Kosari et al.,
2020). Therefore, field borders extraction was developed as
a semantic segmentation problem in which several class
labels are predicted. The four related tasks are to map
areas, define borders, approximate the distance to a closest
border and reconstruct images in the fields. Extracting field
borders entails classifying each pixel in a multispectral
image into one of two categories: ‘‘border’’ or ‘‘not bor-
der.’’ Although single-tasking can produce satisfactory
results, it lacks training signals from similar learning tasks
(Waldner & Diakogiannis, 2020).
Model Architecture
The encoder compresses the contents of the knowledge of
an arbitrarily large image. The decoder increases the
encoded functionality progressively up to the main reso-
lution and locates the categories of interest specifically.
Two simple architectures have been introduced that vary in
the encoder–decoder’s layers. In the encoder there are six
rest blocks in the ResUNet-a D6, and then in the PSP
Pooling layer, then in the encoder there are seven building
blocks in the ResUNet-a D7.
Border Extraction
Both make use of multiple semantic segmentation outputs
and are data-driven, enabling the configuration of a small
sample of reference data to be automated. In the watershed
process, the three output masks are segmented using a
seeded watershed segmentation algorithm. When catch-
ment basins hit other catchments, they stop growing; their
borders are lines that divide neighboring catchment basins.
Journal of the Indian Society of Remote Sensing
123
By comparison, the centers of each field (seed) can be
identified and developed prior to them colliding with the
borders of other fields or pixels that do not belong to the
field (background). The seed is identified by the distance
mask’s threshold, and the topographical surface is identi-
fied by the extent mask’s threshold.
The post-processing system thresholds are installed in
two stages: the first stage optimizes the scope of fields
collected (the scale mask’s threshold); the second stage
optimizes the form and size of fields (thresholds on the
border and distance masks). This technique therefore needs
reference dataset for optimal threshold values to be avail-
able. In the first place, the degree threshold is defined by
the maximization of the correlation coefficient of the
Matthew (Ghaderizadeh et al., 2021). The MCC is calcu-
lated as below.
MCC ¼TP.TN FP.FN
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
TP þFNðÞTP þFPðÞTN þFPðÞTN þFNðÞ
p
ð1Þ
where true positive, true negative, false positive, and false
negative rates are denoted by TP, TN, FP, and FN,
respectively. MCC values range from -1 (complete dis-
agreement) to ?1 (complete agreement), with 0 indicating
that the matter is correct.
Second, border and distance mask borders are optimized
together to reduce the inappropriate separation into smaller
(oversegmentation) of larger objects and an improper
aggregation into larger neighboring small objects (under-
segmentation). In the reference fields (T) and in the
extracted field (E) data using Eqs. (2) and (3) (Wang &
Wang, 2012):
Soverseg ¼1Ti\Ej
Ti
jj ð2Þ
Sunderseg ¼1Ti\Ej
Ej
ð3Þ
Dataset
We have chosen Sentinel-2 images, which acquired from
some agricultural fields from Sarab County in East Azer-
baijan Province, Iran. This dataset increases the probability
of cloud-free images acquired while also data which sup-
port multi-temporal composites with 10-m resolution. The
sentinel- 2 images are given for large-scale applications
free of charge. Additionally, we obtained Landsat-8 images
in the visible and near-infrared bands that were cloud-free.
Preprocessing consisted of two steps: subtraction of the
blue, green, red, and near infrared bands from the satellite
images and standardization of pixel values to ensure that
each band had a mean of zero and a standard deviation of
one. Standardization is important because the input vari-
able size of the neural network is responsive and gradient
optimization approaches (Mancino et al., 2020).
Also, field borders were generated by digitizing manu-
ally all fields across the study area using google earth
images. Field polygons were rasterized at 10 m, matching
the grid of the Sentinel-2 and supplying a collection of
reference pixels for validation. Pixels of ‘‘non-field’’ have
been set to 0. Each image examined (representing various
study areas) was divided into a training and testing area in a
3:2 ratio. Due to the limited GPU memory, each Sentinel-2
tile has been partitioned into a series of smaller images
(256,256 pixels) as input images. There was only one
preprocessing step: standardizing pixel values to have a
mean of zero and a standard deviation of one for each band.
Standardization is required because neural networks are
sensitive to the scale of their input variables and gradient
optimization methods converge faster when features have a
mean of zero and a variance of one.
Evaluation Method
A frequent problem in classification is determining the
appropriate threshold for class assignment given a contin-
uous classifier. The threshold value selected is highly
dependent on the user requirements and their tolerance for
false positives and negatives (Mahdipour et al., 2016). The
Receiver Operating Characteristic (ROC) curve is fre-
quently used to assess a continuous classifier’s ability to
correctly identify a binary array, in this case a map of
‘‘boundary’’ and ‘‘not boundary’’ pixels. The true positive
rate (the fraction of mapped boundary pixels that are cor-
rectly classified as ‘‘boundary’’) is plotted against the false
positive rate for a range of classifier threshold values (the
fraction of mapped not boundary pixels that are incorrectly
classified as ‘‘boundary’’). As the threshold is lowered from
a value where all pixels are classified as ‘‘not boundary’’ to
one where all pixels are classified as ‘‘boundary,’’ a good
classifier will identify true positives more quickly than it
accepts false positives. The ROC curve plots the true
positive rate against the false positive rate, with more
accurate classifiers resulting in a curve that is closer to the
plot’s upper left corner (Mahdipour et al., 2020). Thus, the
area under the curve can be used to quantify a classifier’s
overall performance (AUC). The ROC plots as a straight
line between (0,0) and (1,1) with an AUC of 0.5 for a
random classification surface where each pixel has a 50%
chance of being classified as boundary or not boundary.
The AUC of a classifier is expected to be between 0.5 and
1.0. (Fig. 1).
The performance of various methods was evaluated
using two widely used metrics, namely the overall accuracy
Journal of the Indian Society of Remote Sensing
123
(OA) and the mean average precision (mAP). For com-
parisons, four cutting-edge FCN models were used: SegNet
(Badrinarayanan, Kendall, and Cipolla 2017), DeconvNet
(Hyeonwoo Noh, Hong, et al., 2015; Noh, Seunghoon,
et al., 2015), FCN8s (Shelhamer et al., 2017), and U-Net
(Ronneberger et al., 2015a). These methods were chosen
because they have all been demonstrated to be effective in
semantic labeling for remote sensing images and are all
open source and simple to implement.
Result and Discussion
The model was implemented on Python platform using a
computer with a 1.2 GHz quad-core CPU, 6 GB of mem-
ory, and a Linux 4.4 based Raspbian operating system. We
trained ResUNet-a D6 and D7 models (it took 73 s); We
trained ResUnet-a D6 and D7 models using Adam as the
optimizer. Adam calculates individual adaptive learning
rates for different parameters using estimates of the gra-
dient’s first and second moments. Adam achieves faster
convergence than other alternatives, as demonstrated
empirically (Kingma & Ba, 2015). The weight decay (WD)
approach incorporates a weight decay parameter into the
learning rate in order to incrementally decrease the weight
and bias values of the neural network during each training
iteration. This technique is equivalent to adding an L2
regularization term to the loss function and can accelerate
convergence. For 200 epochs, we trained models with
different weight decay parameters (10
–4
,10
–5
, and 10
–6
.).
For 200 epochs, the network is trained. We chose this large
number of epochs to ensure that each network performs
optimally and to save the optimal set of weights for each
network to avoid overfitting. The first three iterations were
parameterized with weight decay (WD) values, while the
final iteration was configured interactively. After each
epoch, the loss function and MCC were calculated for the
test set. We concluded that it was unlikely that models
would increase accuracy over more than 200 epochs
because training curves showed overfit signs in seventy
epochs.
Ground truth map were created using digitizing of
Google Earth images. The maps were then converted to a
binary file including the farm border (one) and the fore-
ground (zero). Finally, using evaluation metrics, compar-
isons were made and numerical results were obtained
according to Table 1. The average accuracy is 85% and the
MCC is 74%. The cropland grade was marginally lower
than the non-cropland grade by 85 percent. Tuning the map
scale threshold to optimize MCC only resulted in small
deviations in comparison with a 50% defect threshold.
However, we maintained the optimized threshold because
we tried to reach the grading with the most balanced
accuracy for croplands and non-croplands. Also, the results
suggest that temporal data must also be taken into account
in certain situations.
As can be seen in Table 2, our model defined borders
with considerably higher sharpness and less noise than an
edge-based benchmark tool (more than 0.86 for hit rates).
This shows, by using hierarchical contextual knowledge,
that convolutional neural networks reduce the value of
temporal information. Current neural networks are invari-
ant of size. The model functions are object-based, which
means they don’t depend on pixel values but rather on the
fact that they all belong to the same object. This process is
not subject to a specific resolution.
We contrasted the mask found on the borders of the
Canny filter with our ResUNet-a (Fig. 2). The edge-based
approach has generated slightly less and noisier interfaces.
Inside pixel values were by average higher with traditional
rim detection and wider than those achieved with ResU-
Net-a. The retracted edges become clearer as the neural
network learns to be adaptive to specific edge shapes. As
compared to a monthly composite, feeding the method with
single data has no effect on the model’s performance. For
example, oversegmentation and undersegmentation
Fig. 1 The overall performance of models by the area under the curve
(AUC)
Table 1 Pixel-based assessment of proposed model for the single
image with different thresholds, multitemporal images, and resampled
Sentinel-2 image with specific thresholds
Threshold OA MCC F
C
F
NC
Single (50%) 85.40 74.12 78.34 87.15
Single (35%) 86.19 73.75 78.52 86.83
Multitemporal (35%) 78.21 69.09 78.96 81.59
Resamples 30-m (35%) 75.83 71.33 73.41 79.20
Journal of the Indian Society of Remote Sensing
123
premiums fell, while the compensation remained
unchanged.
The quantitative results obtained using various methods
are summarized in Table 3. ResUNet produced the best
result, with an overall accuracy of 85.60 percent and a
mean AP score of 88.05 percent, when compared to all
other methods utilizing encoder-decoder network
architectures. SegNet outperformed the ResUNet model in
terms of accuracy. ResUNet and SegNet are demonstrated
to perform admirably. UNet outperforms DeconvNet and
FCN8s in terms of stable classification performance across
a variety of scenes.
The statistically relevant variances were only for the
sub-segmentation rate, which means that it is a valid
alternative to derive field borders from single dated images.
The model has been trained on Sentinel-2 images and
attained a high degree of pixel and field accuracy. The
same model was correctly created, without retraining. The
same model was acquired by Sentinel-2 and Landsat-8 for
single images of the same site. In areas where cloud cov-
erage is continuous as in the tropics, the ability to remove
borders from a one-date image is cost-efficient. It shows
that by exploiting contextual knowledge at multiple levels,
convolutional neural networks minimize the impact of
spectral and temporal information.
Conclusion
This study demonstrates that solving the issue of the field
border extraction of satellite images with a convolutional
neural network as multiple semantic segmenting tasks
delivers excellent efficiency at pixel and object stage.
ResUNet-a as to extract field borders from rougher artifacts
can be associated, but this needs to be confirmed experi-
mentally, through the use of several convolutions to clas-
sify features on various scales. We observed a reduced hit
rate and geometric accuracy up to 11% while we were
applying the model to resampled Sentinel-2 to 30-m and
Landsat-8 images. The main reason was that smaller fields
in the 30-m image were not resolved. Differences between
the reference resolution and extracted maps may lead to
artificial distortions in the precise evaluation. The differ-
ence between the results of resampled Sentinel-2 to 30-m
and Landsat-8 images is lack of accuracy is partially due to
differences in spectral and spatial responses in these
satellite images. The results showed that semantic seg-
mentation in combination with multitasking is used to
increasing the potential to recover individual areas from
satellite image. One of the main limitation of this study was
the computational density of the proposed model, which
requires significantly more FLOPS than similar models
EfficientNets. ResNets are typically run on GPUs and they
are computationally heavy. Also, the maximum pixel-to-
spatial level accuracy of segmentation which our proposed
model can achieve was one of our limitation. Many sup-
pliers of automated agriculture services urgently require
the ability to automatically detect field borders from remote
sensing images. In order to forecast the likelihood for any
pixel to belong to the field, to the border of the field and to
Table 2 Object-based assessment of proposed model for the Sentinel-
2 image, resampled Sentinel-2 image, and Landsat-8 image
Image Sentinel-2 Sentinel-2 (30 m) Landsat-8
Hit rate 0.952 0.879 0.865
Oversegmentation 0.893 0.797 0.781
Undersegmentation 0.897 0.795 0.778
Eccentricity 0.941 0.915 0.903
Shift (m) 5 7 8
Fig. 2 (a) Agricultural fields in the study of area and comparison of
the field borders extracted using (b) proposed method, (c) references
borders, and (d) Canny filter
Table 3 The quantitative results
using the deep learning models Model mAP OA
FCN8s 63.88 75.09
DeconvNet 65.22 77.87
UNet 69.21 81.11
SegNet 72.04 83.80
ResUNet 75.62 85.60
Journal of the Indian Society of Remote Sensing
123
predict the distance to the closest border, our model was
based on multitasking and conditioning. These projections
were then post-processed such that individual fields were
segmented and extracted. The proposed method demon-
strated suitable efficiency in field border detection and
strong generalization skills over time, space and sensors.
Declarations
Conflict of interest The authors whose names are listed immediately
below certify that they have NO affiliations with or involvement in
any organization or entity with any financial interest (such as hono-
raria; educational grants; participation in speakers’ bureaus; mem-
bership, employment, consultancies, stock ownership, or other equity
interest; and expert testimony or patent-licensing arrangements), or
non-financial interest (such as personal or professional relationships,
affiliations, knowledge or beliefs) in the subject matter or materials
discussed in this manuscript.
1. Alireza Sharifi
2. Hadi Mahdipour
3. Elahe Moradi
4. Aqil Tariq
References
Badrinarayanan, V., Kendall, A., & Cipolla, R. (2017). SegNet: A
deep convolutional encoder-decoder architecture for image
segmentation. IEEE Transactions on Pattern Analysis and
Machine Intelligence, 39(12), 2481–2495.
Burdick, J., Marques, O., Weinthal, J., & Furht, B. (2018). Rethinking
skin lesion segmentation in a convolutional classifier. Journal of
Digital Imaging.https://doi.org/10.1007/s10278-017-0026-y
Chai, D., Newsam, S., Zhang, H. K., Qiu, Y., & Huang, J. (2019).
Cloud and cloud shadow detection in landsat imagery based on
deep convolutional neural networks. Remote Sensing of Envi-
ronment.https://doi.org/10.1016/j.rse.2019.03.007
Ciobanu, R. I., Negru, C., Pop, F., Dobre, C., Mavromoustakis, C. X.,
& Mastorakis, G. (2019). Drop computing: Ad-Hoc dynamic
collaborative computing. Future Generation Computer Systems.
https://doi.org/10.1016/j.future.2017.11.044
Gao, L., Song, W., Dai, J., & Chen, Y. (2019). Road extraction from
high-resolution remote sensing imagery using refined deep
residual convolutional neural network. Remote Sensing.https://
doi.org/10.3390/rs11050552
Ghaderizadeh, S., Abbasi-Moghadam, D., Sharifi, A., Zhao, Na., &
Tariq, A. (2021). Hyperspectral image classification using a
hybrid 3D–2D convolutional neural networks. IEEE Journal of
Selected Topics in Applied Earth Observations and Remote
Sensing.https://doi.org/10.1109/JSTARS.2021.3099118
Guo, H., He, G., Jiang, W., Yin, R., Yan, L., & Leng, W. (2020).
A Multi-scale water extraction convolutional neural network
(MWEN) method for GaoFen-1 remote sensing images. ISPRS
International Journal of Geo-Information.https://doi.org/10.
3390/ijgi9040189
Hadi, M., Khademi, M., & Yazdi, H. S. (2016). Efficient land-cover
segmentation using meta fusion. Journal of Information Systems
and Telecommunication.https://doi.org/10.7508/jist.2016.03.003
Hadi, M., Shabanian, M., & Kazerouni, I. A. (2020). Vectorized
kernel-based fuzzy C-means: A method to apply KFCM on crisp
and non-crisp numbers. International Journal of Uncertainty,
Fuzziness and Knowlege-Based Systems. https://doi.org/10.1142/
S0218488520500270
IdBenIdder, H., & Laachfoubi, N. (2015). Unsupervised multi-level
thresholding method for weather satellite cloud segmentation.
International Journal of Computer Applications.https://doi.org/
10.5120/20826-3544
Khan, S., Lee, D. H., Khan, M. A., Gilal, A. R., & Mujtaba, G. (2019).
Efficient edge-based image interpolation method using neigh-
boring slope information. IEEE Access.https://doi.org/10.1109/
ACCESS.2019.2942004
Kingma, D.P., and Ba, J. (2015). ‘‘Adam: A method for stochastic
optimization.’’ CoRR abs/1412.6980.
Kosari, A., Sharifi, A., Ahmadi, A., & Khoshsima, M. (2020). Remote
sensing satellite’s attitude control system: Rapid performance
sizing for passive scan imaging mode. Aircraft Engineering and
Aerospace Technology, 92(7), 1073–1083. https://doi.org/10.
1108/AEAT-02-2020-0030
Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2017). ImageNet
classification with deep convolutional neural networks. Commu-
nications of the ACM.https://doi.org/10.1145/3065386
Kumar, L., Afzal, M. S., & Afzal, M. M. (2020). Mapping shoreline
change using machine learning: A case study from the Eastern
Indian Coast. Acta Geophysica.https://doi.org/10.1007/s11600-
020-00454-9
Mancino, G., Ferrara, A., Padula, A., & Nole
`, A. (2020). Cross-
comparison between landsat 8 (OLI) and landsat 7 (ETM?)
derived vegetation indices in a mediterranean environment.
Remote Sensing.https://doi.org/10.3390/rs12020291
Meher, B., Agrawal, S., Panda, R., & Abraham, A. (2019). A survey
on region based image fusion methods. Information Fusion.
https://doi.org/10.1016/j.inffus.2018.07.010
Mohammadi, M., & Sharifi, A. (2021). Evaluation of convolutional
neural networks for urban mapping using satellite images.
Journal of the Indian Society of Remote Sensing.https://doi.org/
10.1007/s12524-021-01382-x
Noh, H, S Hong, and B Han. 2015. ‘‘Learning Deconvolution
Network for Semantic Segmentation.’’ In 2015 IEEE Interna-
tional Conference on Computer Vision (ICCV), 1520–28. https://
doi.org/10.1109/ICCV.2015.178.
Noh, H., Seunghoon, H., and Bohyung, H. (2015). ‘‘Learning
deconvolution network for semantic segmentation.’’ In Proceed-
ings of the IEEE International Conference on Computer Vision.
https://doi.org/10.1109/ICCV.2015.178
Park, K., Chae, M., & Cho, J. H. (2021). Image pre-processing
method of machine learning for edge detection with image signal
processor enhancement. Micromachines.https://doi.org/10.3390/
mi12010073
Paudel, D., Boogaard, H., de Wit, A., Janssen, S., Osinga, S.,
Pylianidis, C., & Athanasiadis, I. N. (2021). Machine learning
for large-scale crop yield forecasting. Agricultural Systems.
https://doi.org/10.1016/j.agsy.2020.103016
Rabbi, J., Ray, N., Schubert, M., Chowdhury, S., & Chao, D. (2020).
Small-object detection in remote sensing images with end-to-end
edge-enhanced GAN and object detector network. Remote
Sensing.https://doi.org/10.3390/RS12091432
Ronneberger, O., Fischer, P., & Brox, T. (2015a). U-Net: convolu-
tional networks for biomedical image segmentation. In Lecture
Notes in Computer Science (Including Subseries Lecture Notes
in Artificial Intelligence and Lecture Notes in Bioinformatics).
https://doi.org/10.1007/978-3-319-24574-4_28
Journal of the Indian Society of Remote Sensing
123
Sharifi, A. (2018). Estimation of biophysical parameters in wheat
crops in golestan province using ultra-high resolution images.
Remote Sensing Letters, 9(6), 559–568. https://doi.org/10.1080/
2150704X.2018.1452058
Sharifi, A. (2020). Using sentinel-2 data to predict nitrogen uptake in
maize crop. IEEE Journal of Selected Topics in Applied Earth
Observations and Remote Sensing, 13, 2656–2662. https://doi.
org/10.1109/JSTARS.2020.2998638
Shelhamer, E., Long, J., & Darrell, T. (2017). Fully convolutional
networks for semantic segmentation. IEEE Transactions on
Pattern Analysis and Machine Intelligence.https://doi.org/10.
1109/TPAMI.2016.2572683
Shi, J., Jin, H., & Xiao, Z. (2020). A novel hybrid edge detection
method for polarimetric SAR images. IEEE Access.https://doi.
org/10.1109/ACCESS.2020.2963989
Sudmanns, M., Tiede, D., Augustin, H., & Lang, S. (2020). Assessing
global sentinel-2 coverage dynamics and data availability for
operational earth observation (EO) applications using the EO-
compass. International Journal of Digital Earth.https://doi.org/
10.1080/17538947.2019.1572799
Tariq, A., Shu, H., Siddiqui, S., Munir, I., Sharifi, A., Li, Q., & Linlin,
Lu. (2021). Spatio-temporal analysis of forest fire events in the
margalla hills, islamabad, pakistan using socio-economic and
environmental variable data with machine learning methods.
Journal of Forestry Research.https://doi.org/10.1007/s11676-
021-01354-4
Waldner, F., & Diakogiannis, F. I. (2020). Deep learning on edge:
extracting field boundaries from satellite images with a convo-
lutional neural network. Remote Sensing of Environment.https://
doi.org/10.1016/j.rse.2020.111741
Wang, J., & Wang, X. (2012). VCells: simple and efficient
superpixels using edge-weighted centroidal voronoi tessellations.
IEEE Transactions on Pattern Analysis and Machine Intelli-
gence.https://doi.org/10.1109/TPAMI.2012.47
Wang, S., Tuor, T., Salonidis, T., Leung, K. K., Makaya, C., He, T., &
Chan, K. (2019). Adaptive federated learning in resource
constrained edge computing systems. IEEE Journal on Selected
Areas in Communications.https://doi.org/10.1109/JSAC.2019.
2904348
Wu, F., Bingfang, Wu., Zhang, M., Zeng, H., & Tian, F. (2021).
Identification of crop type in crowdsourced road view photos
with deep convolutional neural network. Sensors (switzerland).
https://doi.org/10.3390/s21041165
Publisher’s Note Springer Nature remains neutral with regard to
jurisdictional claims in published maps and institutional affiliations.
Journal of the Indian Society of Remote Sensing
123
