Content uploaded by Oytun Emre Sakici
Author content
All content in this area was uploaded by Oytun Emre Sakici on Aug 12, 2018
Content may be subject to copyright.
Sakici ‒ Günlü: Artificial intelligence applications using satellite data for stand parameters’ estimations
- 5269 -
APPLIED ECOLOGY AND ENVIRONMENTAL RESEARCH 16(4):5269-5285.
http://www.aloki.hu ● ISSN 1589 1623 (Print) ● ISSN 1785 0037 (Online)
DOI: http://dx.doi.org/10.15666/aeer/1604_52695285
2018, ALÖKI Kft., Budapest, Hungary
ARTIFICIAL INTELLIGENCE APPLICATIONS FOR
PREDICTING SOME STAND ATTRIBUTES USING LANDSAT 8
OLI SATELLITE DATA: A CASE STUDY FROM TURKEY
SAKICI, O. E.1* – GÜNLÜ, A.2
1Faculty of Forestry, Kastamonu University, Kastamonu, Turkey
2Faculty of Forestry, Çankırı Karatekin University, Çankırı, Turkey
*Corresponding author
e-mail: oesakici@kastamonu.edu.tr; phone: +90-366-280-1740; fax: +90-366-215-2316
(Received 13th Jun 2018; accepted 1st Aug 2018)
Abstract. Forest resources inventory is one of the essential parts of the sustainable forest management.
Remote sensing applications have broad usage areas for this aim, since field measurements are costly,
time consuming and laborious. Monitoring forest resources with various satellite images has found wide
usage areas in forestry. In this study, the relationships between some stand attributes (mean diameter,
basal area, stand volume and number of trees) and texture values obtained from Landsat 8 OLI satellite
image were investigated for Crimean pine (Pinus nigra J.F. Arnold subsp. pallasiana (Lamb.) Holmboe)
stands in Kastamonu region of Turkey. The multiple linear regression analysis and artificial neural
networks (ANN) were utilized to fit stand parameters using texture values. To form the ANN
architectures, various transfer functions in hidden and output layers and number of nodes ranged from 1
to 20 in hidden layer were used, and a total of 180 architectures were designed for each stand attribute.
The results indicated that the regression models had low R2 values (0.399 for mean diameter, 0.337 for
basal area, 0.332 for stand volume, and 0.183 for number of trees), and the most of the ANN models were
better than the regression models for predicting stand attributes. The model containing hyperbolic tangent
transfer functions in both hidden and output layers for mean diameter (R2 = 0.593), logistic transfer
function in hidden layer and hyperbolic tangent function in output layer for basal area and stand volume
(R2 = 0.632 and 0.650, respectively), and hyperbolic tangent function in hidden layer and linear function
in output layer for number of trees (R2 = 0.610) were the best ANN models. This study concluded that the
ANN models developed with Landsat 8 OLI were useful to predict stand parameters better than the
regression models in Crimean pine stands located in Kastamonu, Turkey.
Keywords: forest inventory, satellite image, artificial neural networks, multiple linear regression
Introduction
Forest ecosystems provide so many different economic, ecological and social
products and services such as wood and non-wood products, carbon sequestration,
biodiversity conservation, wildlife contribution, water and soil protection, and
recreation. However, these ecosystems face deforestation and desertification problems
because of human activities, unusual climatic conditions due to global warming, insect
attacks, erosion, wildfires, etc. In this negative situation, sustainable management of
forests becomes more important. The inventory and monitoring of forest resources are
the main parts of the sustainable forest management.
Traditional forest inventory based on ground measurements is very hard, time-
consuming and costly in a great forest area (Lu et al., 2004). In addition to ground
measurements, remote sensing data are also widely used for forest management
planning (Holmgren and Thuresson, 1998). Particularly for predicting stand-level
circumstances across great forest areas, remote sensing data have been utilized
efficiently to offer valuable information regarding forest constructions for forest
Sakici ‒ Günlü: Artificial intelligence applications using satellite data for stand parameters’ estimations
- 5270 -
APPLIED ECOLOGY AND ENVIRONMENTAL RESEARCH 16(4):5269-5285.
http://www.aloki.hu ● ISSN 1589 1623 (Print) ● ISSN 1785 0037 (Online)
DOI: http://dx.doi.org/10.15666/aeer/1604_52695285
2018, ALÖKI Kft., Budapest, Hungary
managers (Cohen et al., 1995). Since the early 21st century, Landsat satellite data have
had broad usage in many forestry studies such as land cover or land use change (Günlü
et al., 2008; Huang et al., 2009; Hu et al., 2016), stand parameters predictions
(Kahriman et al., 2014; Günlü and Başkent, 2015; Günlü and Kadıoğulları, 2018) and
aboveground biomass estimations (Zheng et al., 2004; Lu et al., 2012; Günlü et al.,
2014a).
Forest attributes such as mean diameter, stand basal area, stand volume and numbers
of trees are important for forest management planning activities. In studies on predicting
stand parameters using Landsat satellite images, the models were developed by using
reflectance and vegetation indices values obtained from Landsat satellite data (Hall et
al., 2006; Mohammadi et al., 2010; Günlü and Kadıoğulları, 2018). Besides, there are
several studies to predict the stand parameters (especially for predicting aboveground
biomass) by using texture values generated from Landsat satellite data (Kelsey and
Neff, 2014; Safari and Sohrabi, 2016). In general, regression analysis was used to
predict stand parameters with remote sensing data (Hyde et al., 2006; Gama et al.,
2010). Recently, some studies have been performed to estimate the stand parameters
using artificial neural networks (ANN) techniques (Ercanli et al., 2016; Reis et al.,
2018).
The aims of this research are (i) to generate the ANN models estimating relationships
between the stand parameters (mean diameter, stand basal area, stand volume and
numbers of trees) obtained from ground measurements and texture values generated
from Landsat 8 OLI satellite image, (ii) to assess the utilization of the ANN models for
attaining the estimation of stand parameters by matching the regression analysis
outcomes in pure Crimean pine stands of Kastamonu Regional Directorate of Forestry.
Materials and methods
Study area
This study was carried out on pure Crimean pine stands within the boundaries of
Kastamonu Regional Directorate of Forestry located in the Black Sea Region of Turkey
(Fig. 1). This directorate is the first among 28 regional directorates of Turkey in terms
of growing stock (201.4 million m3) and annual volume increment (4.3 million m3), and
these amounts equal to about 13% of the whole country. Forests cover a total of 1.26
million ha, which is about 66% of the total area of the region (General Directorate of
Forestry, 2015). Crimean pine is also the most common tree species with 0.38 million
ha distribution area in the region (General Directorate of Forestry, 2006).
Mean annual temperature and precipitation of the study area are 9.8 °C and 480 mm,
respectively. The slope varies from 0% to 80%, and the elevation ranges between 604-
1579 m above sea level, with an average of 1149 m. All sampled areas consist of
naturally regenerated pure Crimean pine stands.
Field measurements
Field measurements for this study were conducted in 184 circular sample plots
during summer seasons of 2015 and 2016. Sample plot sizes were determined
considering stand crown closures, which is a key parameter to decide the sample plot
sizes for forest inventory activities in Turkey. According to the crown closure classes of
the stands (i.e., 11-40%, 41-70% and >71%), the sizes of circular sample plots were set
Sakici ‒ Günlü: Artificial intelligence applications using satellite data for stand parameters’ estimations
- 5271 -
APPLIED ECOLOGY AND ENVIRONMENTAL RESEARCH 16(4):5269-5285.
http://www.aloki.hu ● ISSN 1589 1623 (Print) ● ISSN 1785 0037 (Online)
DOI: http://dx.doi.org/10.15666/aeer/1604_52695285
2018, ALÖKI Kft., Budapest, Hungary
as 800, 600 or 400 m2 with the radii of 15.96, 13.82 or 11.28 m, respectively. For some
sample plots including excessive number of trees, the radii of the plots were reduced to
7.98 m (i.e., 200 m2 in size). Sample plots were randomly selected to represent different
stand characteristics, which directly affect the growth rate, such as stand densities,
diameter classes and site qualities.
Figure 1. Landsat 8 OLI satellite image of study area
Measurements in the sample plots were initiated by recording UTM coordinates
using a Global Positioning Systems (GPS) receiver placed at the center of every sample
plot. In each sample plot, the diameter at breast height (dbh) over-bark was measured to
the nearest 0.1 cm using calipers for each trees greater than 7.9 cm dbh, and the number
of trees measured were counted. In total, 5757 trees were measured for dbh, and the
number of trees measured in sample plots were ranged from 7 to 92 trees. Basal areas of
each tree in sample plots were calculated. Volumes of the trees were predicted using the
single-entry volume equation developed by Sakici et al. (2018) for Crimean pine stands
located in the study area. After measuring and calculating the sample trees’
dendrometric values, total basal area and total volume of each sampling plot were
calculated by summing basal areas and volumes of sample trees in sample plot,
respectively. The total basal area and total volume of sample plots were ranged between
0.341-4.371 m2 and 0.848-45.203 m3, respectively. Then, stand parameters such as
mean diameter (dq), stand basal area (G), stand volume (V) and number of trees per
hectare (N) were calculated using Equations 1, 2, 3 and 4 for each sample plots.
(Eq.1)
(Eq.2)
Sakici ‒ Günlü: Artificial intelligence applications using satellite data for stand parameters’ estimations
- 5272 -
APPLIED ECOLOGY AND ENVIRONMENTAL RESEARCH 16(4):5269-5285.
http://www.aloki.hu ● ISSN 1589 1623 (Print) ● ISSN 1785 0037 (Online)
DOI: http://dx.doi.org/10.15666/aeer/1604_52695285
2018, ALÖKI Kft., Budapest, Hungary
(Eq.3)
(Eq.4)
where, dq is the quadratic mean diameter (cm), dbh is the tree diameter measured at
1.30 m from the ground, G is the stand basal area (m2 ha-1), g is the total basal area of
sample plot (m2), V is the stand volume (m3 ha-1), v is the total volume of sample plot
(m3), N is the number of trees per hectare, n is the number of trees in the sample plot,
and A is the sample plot area (m2).
The 184 sample plots were randomly divided into two groups to generate the model
development and validation data sets. The data from 138 sample plots (75% of total
data) were used to develop the models. As independent data set, the data from
remaining 46 sample plots (25% of total data) were reserved for validation process of
the developed models. The summary statistics for the sample plots were given in
Table 1 for modeling and validation data groups, separately.
Table 1. Summary statistics of sample plots
Mean diameter
(cm)
Basal area
(m2 ha-1)
Stand volume
(m3 ha-1)
Number of trees
(ha-1)
Modeling data (n = 138)
Minimum
10.4
5.685
14.128
88
Maximum
51.7
75.873
882.648
1800
Mean
28.3
31.160
273.305
570
Standard deviation
10.3
16.172
177.385
345.6
Validation data (n = 46)
Minimum
12.9
7.224
19.976
88
Maximum
53.7
66.113
639.660
1850
Mean
29.7
36.288
323.589
634
Standard deviation
10.8
14.202
150.482
380.6
Remote sensing data and processing
The Landsat 8 Operational Land Imager (OLI) satellite image used in this research
was free downloaded from United States Geological Survey Global Visualization
Viewer (URL-1, 2014). The Landsat 8 OLI satellite image has already included
atmospheric and geometric corrections, and radiometric calibration. In this study, five
bands (Band 2, 3, 4, 5 and 7) of Landsat 8 OLI satellite image with a spatial resolution
of 30 meters were used. Eight different texture metrics (contrast, correlation,
dissimilarity, entropy, homogeneity, mean, second moment and variance) for each band
were produced using four different window sizes (3 x 3, 5 x 5, 7 x 7 and 9 x 9). To
produce texture values, ENVI software was used. After that, depending on the size of
the sample areas, buffer zones (with radius of 7.98, 11.28, 13.82 or 15.96 m) were
composed around the sample plots’ centers in accordance with UTM coordinates
recorded by GPS receiver. The texture images produced for each band were overlaid
using GIS with the sample plots. The texture values of each sample plot were calculated
Sakici ‒ Günlü: Artificial intelligence applications using satellite data for stand parameters’ estimations
- 5273 -
APPLIED ECOLOGY AND ENVIRONMENTAL RESEARCH 16(4):5269-5285.
http://www.aloki.hu ● ISSN 1589 1623 (Print) ● ISSN 1785 0037 (Online)
DOI: http://dx.doi.org/10.15666/aeer/1604_52695285
2018, ALÖKI Kft., Budapest, Hungary
by two different methods. If the sample plot centers were at or near the center of a pixel,
the texture values of the sample plots were calculated as the value of a single pixel for
200, 400 and 600 m2 sample plots, while calculated by taking the average of some pixel
values in the buffer zone for 800 m2 sample plots (Fig. 2a). However, in the second
method, if the sample plot centers were far from the pixel centers, the texture values of
each sample plot were computed by taking the average of pixel values in the buffer zone
for all sample plot sizes (Fig. 2b). In this way, the texture values for five bands, four
window sizes and eight different texture metrics were calculated. Thus, a total of 160
texture values were obtained for each sample plot.
Figure 2. Texture value calculation of sample plots according to buffer zones on satellite image
Regression models
The multiple linear regression analysis was used to develop the equations modeling
the interactions between stand parameters (mean diameter, stand basal area, stand
volume and number of trees) and texture values obtained from Landsat 8 OLI satellite
image. To obtain multiple linear regression models based on stepwise variable selection
method, the ordinary least squares technique was used. The dependent variables in these
models were quadratic mean diameter, stand basal area, stand volume and number of
trees per hectare, and independent variables were 160 texture values of modeling data
group containing 138 sample plots. The dependent variables were observed values from
field works, while the independent ones were produced values from satellite image. The
stepwise regression procedure in IBM SPSS 23 software was used to select the
statistically significant (p < 0.05) texture values as predictor variables to estimate stand
parameters. The relationships between stand parameters and texture values were
assumed linear as given formula below (Eq. 5):
(Eq.5)
where Stand Parameter is quadratic mean diameter, stand basal area, stand volume or
number of trees per hectare, Xi (i = 1 to n) are texture values, bi represent the model
coefficients, and e is the error term.
800 m2
600 m2
400 m2
200 m2
Sample plot center
Sakici ‒ Günlü: Artificial intelligence applications using satellite data for stand parameters’ estimations
- 5274 -
APPLIED ECOLOGY AND ENVIRONMENTAL RESEARCH 16(4):5269-5285.
http://www.aloki.hu ● ISSN 1589 1623 (Print) ● ISSN 1785 0037 (Online)
DOI: http://dx.doi.org/10.15666/aeer/1604_52695285
2018, ALÖKI Kft., Budapest, Hungary
Artificial neural networks
Artificial intelligence techniques were also employed to model stand parameters
beside multiple linear regression analysis to determine prediction success of these
techniques and to compare with regression modeling. For this purpose, artificial neural
network (ANN) models were developed for each stand parameter examined in this
study. For defining the architecture of neural networks, there are several criteria such as
number of layers, learning algorithms, transfer function forms, number of nodes in
hidden layer and determination of data sizes for training, verification and test processes.
The ANN models developed in this study consist of three layers: input, hidden and
output layers. The feed-forward back-propagation network structure was selected
because of its success popularity in forestry literature (e.g., Özçelik et al., 2014;
Diamantopoulou et al., 2015). The learning algorithm used in ANN models was the
Levenberg-Marquardt algorithm. Three transfer function forms (linear, logistic and
hyperbolic tangent) were examined in hidden and output layers, separately (Eqs. 6, 7
and 8).
(Linear function)
(Eq.6)
(Logistic function)
(Eq.7)
(Hyperbolic tangent function)
(Eq.8)
where s = ∑wixi, wi are weights and xi are the input variables.
To determine the most predictive alternatives, the number of nodes in hidden layers
ranged from 1 to 20 adding one by one in training process of ANN models. Thus, a total
of 180 ANN model architectures were created for each stand parameter (Fig. 3). The
ANN models were built using the neural network toolbox in R2015a version of
MATLAB software. The modeling data obtained from 138 sample plots were used for
ANNs’ fitting. Input variables of ANN models were texture values of satellite image
determined as the best predictor for each stand parameter in multiple linear regression
analyses, and output (target) variable was observed mean diameter, stand basal area,
stand volume or number of trees according to the stand attributes fitted.
Figure 3. ANN models architecture for stand parameters
Sakici ‒ Günlü: Artificial intelligence applications using satellite data for stand parameters’ estimations
- 5275 -
APPLIED ECOLOGY AND ENVIRONMENTAL RESEARCH 16(4):5269-5285.
http://www.aloki.hu ● ISSN 1589 1623 (Print) ● ISSN 1785 0037 (Online)
DOI: http://dx.doi.org/10.15666/aeer/1604_52695285
2018, ALÖKI Kft., Budapest, Hungary
Model comparisons and validation tests
The regression and ANN models were evaluated based on four goodness-of-fit
statistics; the coefficient of determination (R2), root mean square error (RMSE), bias and
mean absolute error (MAE). Corresponding mathematical forms of statistical criteria
utilized were defined as Equations 9, 10, 11 and 12.
(Eq.9)
(Eq.10)
(Eq.11)
(Eq.12)
In these equations; and are observed and estimated values of corresponding
stand attribute, and n is sample size.
When comparing alternative models, it is desirable that the R2 values are high while
the others (RMSE, Bias and MAE) are low. For ranking of models, taking into account
of all goodness-of-fit statistics together is better than the ranking of each criteria
separately. The relative ranking method proposed by Poudel and Cao (2013) was used
for model comparisons. In this method, the relative rank of model i for a statistical
criterion is defined using Equation 13.
(Eq.13)
where Ri is the relative rank of model i (i = 1, 2, …, m), Si is the goodness-of-fit statistic
of model i, Smin and Smax are the minimum and maximum values of Si, respectively.
For each stand parameter, relative rankings of ANN models were first implemented
according to number of nodes in hidden layer for transfer function pairs of hidden and
output layers, separately, for each statistical criterion. So, four rankings with 20 ANN
models were formed for nine transfer function pairs (i.e., linear, logistic and hyperbolic
tangent functions were used in both hidden and output layers) for every stand parameter
fitting. The model with the highest R2 was ranked as 1 and the lowest R2 was ranked as
20 for coefficient of determination, while the model with the lowest value was ranked as
1 and the highest value was ranked as 20 for RMSE, Bias and MAE. Then, four relative
ranks of each model according to statistical criteria were summed. The second relative
ranking was generated using total relative ranks of ANN models based on number of
nodes ranged from 1 to 20 for each transfer function pairs. Thus, the most successful
ANN models were specified for every transfer function pairs of each stand parameter
fitting.
Sakici ‒ Günlü: Artificial intelligence applications using satellite data for stand parameters’ estimations
- 5276 -
APPLIED ECOLOGY AND ENVIRONMENTAL RESEARCH 16(4):5269-5285.
http://www.aloki.hu ● ISSN 1589 1623 (Print) ● ISSN 1785 0037 (Online)
DOI: http://dx.doi.org/10.15666/aeer/1604_52695285
2018, ALÖKI Kft., Budapest, Hungary
To compare estimation success of regression and ANN models, the relative ranks of
10 models (a regression model and the best ANN models for nine transfer function
pairs) were rebuilt for each stand parameter. So, the most successful models were
identified for predicting stand parameters evaluated in this study. The validities of
successful models were tested by comparing the validation data observed from 46
sample plots and models’ results by Student’s t-test (also called paired samples t-test)
procedure in IBM SPSS Statistics 23 software. To exhibit the prediction abilities of the
models, residual graphs based on observed vs. predicted stand parameters were also
prepared.
Results
In this study, multiple regression and artificial intelligence techniques were
conducted for predicting some stand parameters such as mean diameter, basal area,
stand volume and number of trees using texture values obtained from Landsat 8 OLI
satellite image, and regression and ANN models were developed with these techniques.
To determine the best predictor texture values for predicting stand attributes and to
develop linear regression models for each stand parameter, multiple linear regression
technique was used. In this analyses, all stand parameters were tried to fit using 160
texture values as candidate independent variables. These variables were produced by
combining five bands of Landsat 8 OLI satellite image, four window sizes and eight
texture metrics. The results of regression analyses are given in Table 2. As seen in this
table, six, seven, six and three variables included in regression models of mean
diameter, basal area, stand volume and number of trees, respectively. All coefficients of
these variables were statistically significant as well as constants of models except for
number of trees (p < 0.05). It might be said that all stand parameters were fitted poorly
with regression models due to the low R2 values. The maximum coefficient of
determination value was obtained for mean diameter (R2 = 0.399), while the minimum
was for number of trees (R2 = 0.183). The R2 values of basal area and stand volume
models were similar, which were 0.337 and 0.332, respectively.
In regression models developed for stand parameters, texture value combinations
containing Band 2 were not statistically significant for number of trees as well as Band
4 for mean diameter and number of trees and Band 7 for all stand parameters except
mean diameter, while some combinations with Band 3 and Band 5 were significant for
all. The combinations including 7 x 7 window sizes were found insignificant for number
of trees, and 9 x 9 ones were for the parameters except basal area. Some combinations
with 3 x 3 and 5 x 5 window sizes were significant for all stand parameters. Texture
values comprising contrast for mean diameter, dissimilarity for all, entropy and
homogeneity for the parameters except mean diameter, mean for number of trees,
second moment for all except number of trees, and variance for basal area and number
of trees were not statistically significant. Some combinations with correlation were
significant for all stand parameters.
ANN models were also fitted for stand parameters using input variables found
statistically significant in regression models. Totally 180 ANN architectures including
various transfer functions in hidden and output layers and number of nodes from 1 to 20
in hidden layer were designed for every stand parameter. Thus, a total of 720 ANN
procedures were implemented in this study. For all stand parameters evaluated, all ANN
models were successfully completed the training process except architectures
Sakici ‒ Günlü: Artificial intelligence applications using satellite data for stand parameters’ estimations
- 5277 -
APPLIED ECOLOGY AND ENVIRONMENTAL RESEARCH 16(4):5269-5285.
http://www.aloki.hu ● ISSN 1589 1623 (Print) ● ISSN 1785 0037 (Online)
DOI: http://dx.doi.org/10.15666/aeer/1604_52695285
2018, ALÖKI Kft., Budapest, Hungary
comprising logistic transfer function in output layer. The ANN models with logistic
transfer function in output layer could not give any result. It means that 120 ANN
models could be ranked instead of 180 models for each stand parameter.
Table 2. Multiple linear regression results for stand parameters
Stand
parameter
Independent
variable
Variable combination (Texture value)
Coefficients
(bi)
R2
SEEa
Band
Window size
Texture metric
Mean diameter
(dq, cm)
Constant
49.531*
0.399
8.158
X1
2
3x3
Correlation
-0.026*
X2
3
7x7
Entropy
-0.149*
X3
5
5x5
Homogeneity
-0.086*
X4
5
5x5
Mean
0.072*
X5
5
5x5
Variance
-0.133*
X6
7
3x3
Variance
0.073*
Basal area
(G, m2 ha-1)
Constant
31.709***
0.337
15.348
X1
2
9x9
Mean
-0.774**
X2
3
9x9
Mean
0.556**
X3
4
7x7
Correlation
0.066**
X4
5
3x3
Contrast
-0.195***
X5
5
5x5
Contrast
0.583***
X6
5
9x9
Contrast
-0.196*
X7
5
9x9
Correlation
-0.078**
Stand volume
(V, m3 ha-1)
Constant
225.797***
0.332
172.487
X1
2
3x3
Mean
-6.949**
X2
3
3x3
Mean
4.149*
X3
4
7x7
Correlation
0.640**
X4
5
3x3
Contrast
-1.514**
X5
5
5x5
Contrast
5.536***
X6
5
5x5
Variance
-2.619**
Number of trees
(N, ha-1)
Constant
-21.209ns
0.183
425.452
X1
3
5x5
Contrast
20.264***
X2
5
3x3
Correlation
1.693**
X3
5
3x3
Second moment
1.402**
aStandard error of estimate
*p < 0.05, **p < 0.01, ***p < 0.001, nsnon-significant (p > 0.05)
Four statistical criteria (R2, RMSE, Bias and MAE) were used to rank the ANN
models. These rankings were applied within six transfer function pairs, i.e. linear,
logistic and hyperbolic tangent functions in hidden layer combined with linear and
hyperbolic tangent functions in output layer, separately. Hence, a total of 24 ranking
lists were produced due to the presence of four stand attributes and six transfer function
pairs. Finally, the best ANN architectures within each ranking list were determined to
compare with each other and regression models.
The regression models and all of the best ANN models for each stand parameter
were re-ranked to decide the most successful models for predicting stand parameters.
Sakici ‒ Günlü: Artificial intelligence applications using satellite data for stand parameters’ estimations
- 5278 -
APPLIED ECOLOGY AND ENVIRONMENTAL RESEARCH 16(4):5269-5285.
http://www.aloki.hu ● ISSN 1589 1623 (Print) ● ISSN 1785 0037 (Online)
DOI: http://dx.doi.org/10.15666/aeer/1604_52695285
2018, ALÖKI Kft., Budapest, Hungary
The final ranking lists are given in Table 3. In this table, it is seen that the regression
models and the ANN models including linear transfer function in hidden layer were
unsuccessful than the other ANN models. The ANN4 models (i.e. the logistic function
in hidden layer and hyperbolic tangent function in output layer) were the most
successful for predicting basal area (with 12 nodes in hidden layer) and stand volume
(with 9 nodes in hidden layer) having high R2 values as 0.632 and 0.650, respectively.
The ANN6 model (i.e. the hyperbolic tangent functions in both hidden and output
layers, and 10 nodes in hidden layer) was the best for mean diameter prediction
(R2 = 0.593), while ANN5 (i.e. the hyperbolic tangent function in hidden layer, the linear
function in output layer, and 9 nodes in hidden layer) was the best to predict number of
trees (R2 = 0.610). The other goodness-of-fit statistics had similar trends to the
coefficient of determinations among the models compared. In the other words, the ANN
models having higher R2 values had lower RMSE, Bias and MAE values. As it can be
understood from these conclusions, the powers of stand parameter predictions were
considerably increased with artificial intelligence applications. These increases of
prediction powers are approximately 50% for mean diameter, nearly 100% for basal
area and stand volume, and more than 200% for number of trees.
If the ranking lists in Table 3 are examined in detail, the regression models were
weaker than the ANN models except the architectures containing linear transfer
functions in both hidden and output layers (ANN1). For number of trees prediction, the
regression model was also stronger than the ANN model comprising the linear transfer
function in hidden layer and the hyperbolic tangent transfer function in output layer
(ANN2). According to the transfer function in hidden layer, the ANN models with
hyperbolic tangent function were more successful for mean diameter and number of
trees predictions, while the models with logistic function for basal area and stand
volume. Besides, according to the transfer function in output layer, the ANN model
with linear function were more successful for number of trees prediction, while the
models with hyperbolic tangent function for the other stand parameters. There was not
any general trend for the ANN models with respect to the number of nodes in hidden
layer. However, the node numbers were around 10 in the best models for all parameters.
The validities of the developed regression models and the best ANN models were
tested with Student’s t-test using independent data set obtained from 46 sample plots.
For all the models tested, there were no significant differences between observed and
predicted values (p > 0.05) and consequently it was decided that the models were
statistically usable for predicting the aforementioned stand parameters (Table 4). Thus,
because of their statistical successes, the ANN6 model for mean diameter, the ANN4
models for basal area and stand volume, and the ANN5 model for number of trees
estimations can be used.
The residual distributions of predicted stand parameters obtained by the best ANN
models and regression models using all data from 184 sample plots are shown in
Figure 4. When the residual distributions are examined, it is seen that the residuals were
distributed randomly in all regression and ANN models. The regression and ANN
models had similar residuals and the mean residuals were close to zero for each stand
parameter. Mean residual values for mean diameter, basal area, stand volume and
number of trees estimates were -0.4 cm, -0.37 m2 ha-1, -3.96 m3 ha-1 and 18.6 ha-1 for
regression models (Fig. 4a-d), and -0.1 cm, 0.64 m2 ha-1, -10.04 m3 ha-1 and 5.5 ha-1 for
the best ANN models (Fig. 4e-h), respectively.
Sakici ‒ Günlü: Artificial intelligence applications using satellite data for stand parameters’ estimations
- 5279 -
APPLIED ECOLOGY AND ENVIRONMENTAL RESEARCH 16(4):5269-5285.
http://www.aloki.hu ● ISSN 1589 1623 (Print) ● ISSN 1785 0037 (Online)
DOI: http://dx.doi.org/10.15666/aeer/1604_52695285
2018, ALÖKI Kft., Budapest, Hungary
Table 3. Relative ranks of models based on goodness-of-fit statistics1,2
Stand
parameter
Model
ANN architecture
R2
(Ri)
RMSE
(Ri)
Bias
(Ri)
MAE
(Ri)
Total
Ri
General
rank
Transfer
function in
hidden
layer
Transfer
function in
output
layer
Number of
nodes in
hidden
layer
Mean
diameter
(dq, cm)
Regression
0.399
(6.94)
7.950
(6.93)
-0.13
(3.86)
6.49
(7.00)
24.72
6.04
ANN1
Linear
Linear
6
0.397
(7.00)
7.967
(7.00)
-0.24
(7.00)
6.49
(7.00)
28.00
7.00
ANN2
Linear
Hyp.
Tangent
16
0.398
(6.97)
7.956
(6.95)
-0.03
(1.00)
6.48
(6.96)
21.89
5.22
ANN3
Logistic
Linear
15
0.550
(2.32)
6.879
(2.41)
-0.03
(1.00)
5.37
(2.80)
8.52
1.32
ANN4
Logistic
Hyp.
Tangent
8
0.497
(3.94)
7.273
(4.07)
0.09
(2.71)
5.86
(4.64)
15.36
3.31
ANN5
Hyp.
Tangent
Linear
9
0.538
(2.68)
6.971
(2.79)
0.09
(2.71)
5.10
(1.79)
9.98
1.74
ANN6
Hyp.
Tangent
Hyp.
Tangent
10
0.593
(1.00)
6.546
(1.00)
0.15
(4.43)
4.89
(1.00)
7.43
1.00
Basal area
(G, m2 ha-1)
Regression
0.337
(6.80)
14.897
(6.83)
-0.0873
(1.00)
12.0117
(7.00)
21.63
6.92
ANN1
Linear
Linear
5
0.327
(7.00)
15.011
(7.00)
-0.0977
(1.07)
11.8592
(6.76)
21.83
7.00
ANN2
Linear
Hyp.
Tangent
11
0.359
(6.37)
14.655
(6.45)
-0.1442
(1.36)
11.8403
(6.73)
20.92
6.63
ANN3
Logistic
Linear
16
0.544
(2.73)
12.354
(2.92)
-0.2436
(2.00)
9.3266
(2.79)
10.44
2.34
ANN4
Logistic
Hyp.
Tangent
12
0.632
(1.00)
11.100
(1.00)
0.5824
(4.15)
8.1835
(1.00)
7.15
1.00
ANN5
Hyp.
Tangent
Linear
8
0.465
(4.29)
13.387
(4.51)
1.0289
(7.00)
10.2516
(4.24)
20.04
6.27
ANN6
Hyp.
Tangent
Hyp.
Tangent
18
0.491
(3.77)
13.057
(4.00)
0.6217
(4.41)
9.8674
(3.64)
15.82
4.54
Stand volume
(V, m3 ha-1)
Regression
0.332
(6.94)
168.055
(6.95)
0.0467
(1.00)
130.3944
(6.93)
21.82
6.23
ANN1
Linear
Linear
11
0.329
(7.00)
168.471
(7.00)
-2.9128
(2.84)
130.8383
(7.00)
23.84
7.00
ANN2
Linear
Hyp.
Tangent
4
0.392
(5.82)
160.306
(5.95)
-3.2678
(3.07)
125.7314
(6.16)
21.01
5.92
ANN3
Logistic
Linear
8
0.540
(3.06)
139.422
(3.28)
-0.1988
(1.10)
101.8162
(2.24)
9.68
1.62
ANN4
Logistic
Hyp.
Tangent
9
0.650
(1.00)
121.599
(1.00)
-6.3624
(5.05)
94.2181
(1.00)
8.05
1.00
ANN5
Hyp.
Tangent
Linear
8
0.585
(2.21)
132.460
(2.39)
-5.9340
(4.78)
101.5374
(2.20)
11.58
2.34
ANN6
Hyp.
Tangent
Hyp.
Tangent
5
0.513
(3.56)
143.435
(3.80)
-9.4027
(7.00)
103.9747
(2.60)
16.95
4.38
Number of
trees
(N, ha-1)
Regression
0.183
(6.91)
419.241
(6.92)
0.03
(1.00)
289.77
(7.00)
21.82
6.62
ANN1
Linear
Linear
14
0.176
(7.00)
421.113
(7.00)
-9.25
(2.00)
284.57
(6.61)
22.61
6.88
ANN2
Linear
Hyp.
Tangent
7
0.216
(6.46)
410.800
(6.54)
-22.41
(3.44)
283.41
(6.52)
22.97
7.00
ANN3
Logistic
Linear
10
0.599
(1.30)
293.818
(1.37)
-12.20
(2.33)
209.70
(1.00)
5.99
1.30
ANN4
Logistic
Hyp.
Tangent
19
0.621
(1.00)
285.546
(1.00)
-55.11
(7.00)
216.10
(1.48)
10.48
2.81
ANN5
Hyp.
Tangent
Linear
9
0.610
(1.15)
289.685
(1.18)
6.92
(1.75)
209.71
(1.00)
5.08
1.00
ANN6
Hyp.
Tangent
Hyp.
Tangent
17
0.551
(1.94)
310.820
(2.12)
-8.17
(1.89)
217.52
(1.59)
7.54
1.82
1Values in parenthesis indicate the relative ranks of the models based on statistical criteria
2The best models are marked in bold for each stand parameters
Sakici ‒ Günlü: Artificial intelligence applications using satellite data for stand parameters’ estimations
- 5280 -
APPLIED ECOLOGY AND ENVIRONMENTAL RESEARCH 16(4):5269-5285.
http://www.aloki.hu ● ISSN 1589 1623 (Print) ● ISSN 1785 0037 (Online)
DOI: http://dx.doi.org/10.15666/aeer/1604_52695285
2018, ALÖKI Kft., Budapest, Hungary
Regression models
ANN models
Mean diameter
Basal area
Stand volume
Number of trees
Figure 4. Residual distributions of the regression models (a-d) and selected ANN models (e-h)
When the observed stand parameters and predicted values of these parameters with
two modeling approaches were compared considering Table 4, the mean diameter
predictions of the regression and ANN6 models were similar to observed values with
mean differences of 1.4 and 1.5 cm, respectively. The basal area predictions of the
ANN4 model were very close to the observed values with mean difference of 0.860 m2,
while the regression model’s predictions were quite different with of 5.221 m2 per
hectare. Since mean differences of the regression and ANN4 models were 65.586 and
(a)
(e)
(b)
(f)
(c)
(d)
(g)
(h)
Sakici ‒ Günlü: Artificial intelligence applications using satellite data for stand parameters’ estimations
- 5281 -
APPLIED ECOLOGY AND ENVIRONMENTAL RESEARCH 16(4):5269-5285.
http://www.aloki.hu ● ISSN 1589 1623 (Print) ● ISSN 1785 0037 (Online)
DOI: http://dx.doi.org/10.15666/aeer/1604_52695285
2018, ALÖKI Kft., Budapest, Hungary
44.559 m3 per hectare, the ANN4 model was more prosperous for the stand volume
estimations. The comparison for the number of trees was resulted different from the
other stand parameters, since the mean difference of the regression model was
positively 45.4 trees while of the ANN5 model was negatively 63.5 trees per hectare.
However, the ANN5 model was superior than the regression model for the number of
trees prediction when the mean residuals of the models took into consideration.
Table 4. Student’s t-test results for the regression and ANN models selected
Stand
parameter
Data
Model
Mean
Paired differences
t
p
Mean
SDa
SEEb
Mean diameter
(dq, cm)
Observed
29.7
Predicted
Regression
28.3
1.4
10.5
1.55
0.894
0.376
ANN6
28.1
1.5
10.8
1.60
0.948
0.348
Basal area
(G, m2 ha-1)
Observed
36.288
Predicted
Regression
31.067
5.221
22.949
3.384
1.543
0.130
ANN4
35.427
0.860
23.279
3.432
0.251
0.803
Stand volume
(V, m3 ha-1)
Observed
323.589
Predicted
Regression
258.003
65.586
238.683
35.192
1.864
0.069
ANN4
279.030
44.559
312.397
46.060
0.967
0.339
Number of
trees
(N, ha-1)
Observed
633.7
Predicted
Regression
588.3
45.4
469.7
69.26
0.655
0.516
ANN5
697.2
-63.5
673.3
99.27
-0.640
0.525
aStandard deviation
bStandard error of estimate
Discussion and conclusions
In this study, the relationships between some stand attributes (mean diameter, basal
area, stand volume and number of trees) and texture values obtained from Landsat 8
OLI satellite image were predicted for Crimean pine stands in Kastamonu region of
Turkey. The multiple linear regression analysis and ANN technique were performed to
fit stand attributes using texture values. The stand attributes estimation models
developed by ANN models performed better than multiple linear regression models
except ANN1 for basal area, stand volume and number of trees and except ANN1 and
ANN2 for mean diameter.
The band reflectance and vegetation indices were used as independent variables in
some published studies performed to estimate the stand parameters using the Landsat
satellite images (Mohammadi et al., 2010; Kahriman et al., 2014; Günlü and
Kadıoğulları, 2018). However, in the literature no studies have been found to predict
mean diameter, basal area, stand volume and number of trees using texture values
obtained from Landsat satellite images. On the other hand, there are some studies on
estimating stand parameters such as aboveground biomass and carbon stock using
textures values obtained from Landsat satellite data images (Kelsey and Neff, 2014;
Zhao et al., 2016; Safari and Sohrabi, 2016 In addition, there are many studies on
estimating stand parameters such as stand volume, basal area, number of trees,
aboveground biomass and carbon stock using texture values obtained from the other
satellite data images (Castillo-Santiago et al., 2010; Özdemir and Karnieli, 2011;
Sakici ‒ Günlü: Artificial intelligence applications using satellite data for stand parameters’ estimations
- 5282 -
APPLIED ECOLOGY AND ENVIRONMENTAL RESEARCH 16(4):5269-5285.
http://www.aloki.hu ● ISSN 1589 1623 (Print) ● ISSN 1785 0037 (Online)
DOI: http://dx.doi.org/10.15666/aeer/1604_52695285
2018, ALÖKI Kft., Budapest, Hungary
Eckert, 2012; Pu and Cheng, 2015; Wallner et al., 2015; Xie et al., 2017). In general,
regression analysis is used in the studies to estimate stand parameters using remote
sensing data images (Castillo-Santiago et al., 2010; Mohammadi et al., 2011; Günlü et
al., 2014b; Kahriman et al., 2014).
The reflectance values, vegetation indices and texture values were utilized as
independent variables of a multiple linear regression model in some studies on the
subject. For instance, the models using vegetation indices obtained from Landsat ETM+
data had the R2 values as 0.43 for stand volume and 0.73 for tree density (Mohammadi
et al., 2010), while the texture variables of IKONOS-2 explained 35% of the variation in
basal area of even-aged stands in temperate forests (Kayitakire et al., 2006). In other
studies, the reflectance and vegetation indices values obtained from Landsat TM
satellite data explained 61% and 70% of tree density, respectively, in mixed forest
areas, and generated from pan-sharpened IKONOS satellite data explained 41% and
43% of stand volume, and 55% and 59% of basal area, respectively (Kahriman et al.,
2014; Günlü et al., 2014b). In the multiple linear regression models with digital number
values of Landsat TM, the R2 values for mean diameter, basal area, stand volume and
number of trees were founded as 0.25, 0.32, 0.37 and 0.44, respectively (Günlü and
Kadıoğulları, 2018). In the same study, the R2 values of the multiple linear regression
analyses obtained from vegetation indices values for mean diameter, basal area, stand
volume and number of trees were founded as 0.17, 0.34, 0.36 and 0.28, respectively.
The results of our study indicated that the R2 values of the multiple linear regression
analyses generated with texture values from Landsat 8 OLI satellite image for mean
diameter, basal area, stand volume and number of trees were founded as 0.40, 0.34, 0.33
and 0.18, respectively. When the results obtained from our study are compared with
some other studies mentioned above (Mohammadi et al., 2010: Kahriman et al., 2014;
Günlü and Kadıoğulları, 2018), the regression model generated from texture values is
not suitable for estimating mean diameter, basal area, stand volume and number of trees
due to the low R2 values.
In recent years, the ANN method used in some published studies on predicting leaf
area index (Shoemaker and Cropper, 2010), stand volume and basal area (Shataee,
2013; Santi et al., 2015), stand carbon stock (Ercanli et al., 2016), and aboveground
biomass (Ni et al., 2017; Yan et al., 2018). We also used the ANN method to predict the
stand parameters in this study. The comparisons between multiple linear regression and
ANN models for estimating stand parameters show that the abilities of the ANN models
are higher. In other words, the R2 values of the best ANN models increased between
48% and 239% for the stand parameters compared to the regression models. Similar
results were published in some rare studies. For instance, Ercanli et al. (2016) modelled
aboveground stand carbon stock using multiple linear regression analysis and ANN with
vegetation indices obtained from Landsat TM satellite data and founded that the R2
values of 0.43 and 0.65, respectively. In another study, a comparative analysis was
conducted for different satellite data and modeling techniques (e.g. ANN, random
forest, linear regression, k-nearest neighbor, support vector regression) for aboveground
biomass prediction in different forest types (Gao et al., 2018). The results obtained from
this study showed that the ANN models were more accurate for mixed, broadleaf and
conifer forest types (RMSE values ranged 30.0-36.0 Mg/ha, 24.0-26.4 Mg/ha and 28.2-
30.9, respectively) than linear regression model (RMSE values ranged 32.6-37.0 Mg/ha,
24.4-31.0 Mg/ha and 28.2-32.4, respectively). It is seen that there are few studies to
predict the stand parameters using both regression analysis and ANN method. Thus, our
Sakici ‒ Günlü: Artificial intelligence applications using satellite data for stand parameters’ estimations
- 5283 -
APPLIED ECOLOGY AND ENVIRONMENTAL RESEARCH 16(4):5269-5285.
http://www.aloki.hu ● ISSN 1589 1623 (Print) ● ISSN 1785 0037 (Online)
DOI: http://dx.doi.org/10.15666/aeer/1604_52695285
2018, ALÖKI Kft., Budapest, Hungary
study is one of the rare studies that use both the multiple linear regression and ANN
models in estimating of stand parameters using remote sensing data, especially texture
values. Therefore, investigating the estimation success of the ANN models using
satellite data for stand attributes is essential for further studies.
The ANN models can be suggested as more suitable than regression models for
predicting stand attributes such as mean diameter, basal area, stand volume and number
of trees using remote sensing data, since the estimation power of the ANN models are
higher than the regression models. The texture values might be used successfully as
predictor variables in these models. The utilization of different model techniques such
as random forest, k-nearest neighbor, support vector regression and mixed effect
modeling and various satellite data such as optical, active and combined can improve
the model achievement criteria in various forest ecosystems.
Acknowledgements. The field data of this study were obtained from a research project supported by The
Scientific and Technological Research Council of Turkey (TUBITAK), Project no: TOVAG-214O217.
REFERENCES
[1] Castillo-Santiago, M. A., Ricker, M., de Jong, B. H. J. (2010): Estimation of tropical
forest structure from SPOT-5 satellite images. – International Journal of Remote Sensing
31(10): 2767-2782.
[2] Cohen, W. B., Spies, T. A., Fiorella, M. (1995): Estimating the age and structure of
forests in a multi-ownership landscape of western Oregon, USA. – International Journal
of Remote Sensing 16(4): 721-746.
[3] Eckert, S. (2012): Improved forest biomass and carbon estimations using texture
measures from WorldView-2 satellite data. – Remote Sensing 4(4): 810-829.
[4] Ercanli, İ., Günlü, A., Şenyurt, M., Bolat, F., Kahriman, A. (2016): Artificial neural
network for predicting stand carbon stock from remote sensing data for even-aged scots
pine (Pinus sylvestris L.) stands in the Taşköprü-Çiftlik forests. – Proceedings of the 1st
International Symposium of Forest Engineering and Technologies (FETEC 2016), 02-04
June, Bursa, Turkey.
[5] Diamantopoulou, M. J., Özçelik, R., Crecente-Campo, F., Eler, Ü. (2015): Estimation of
Weibull function parameters for modelling tree diameter distribution using least squares
and artificial neural networks methods. – Biosystems Engineering 133: 33-45.
[6] Gama, F. F., Santos, J. R., Mura, J. C. (2010): Eucalyptus biomass and volume estimation
using interferometric and polarimetric SAR data. – Remote Sensing 2(4): 939-956.
[7] Gao, Y., Lu, D., Li, G., Wang, G., Chen, Q., Liu, L., Li, D. (2018): Comparative analysis
of modeling algorithms for forest aboveground biomass estimation in a subtropical
region. – Remote Sensing 10(4): 627.
[8] General Directorate of Forestry (2006): Orman Varlığımız. – General Directorate of
Forestry Publications, Ankara.
[9] General Directorate of Forestry (2015): Türkiye Orman Varlığı 2015. – General
Directorate of Forestry Publications, Ankara.
[10] Günlü, A., Başkent, E. Z. (2015): Estimating crown closure of forest stands using Landsat
TM data: A case study from Turkey. – Environmental Engineering and Management
Journal 14(1): 183-193.
[11] Günlü, A., Kadıoğulları, A. İ. (2018): Modeling forest stand attributes using Landsat
ETM+ and QuickBird satellite images in western Turkey. – Bosque 39(1): 49-59.
Sakici ‒ Günlü: Artificial intelligence applications using satellite data for stand parameters’ estimations
- 5284 -
APPLIED ECOLOGY AND ENVIRONMENTAL RESEARCH 16(4):5269-5285.
http://www.aloki.hu ● ISSN 1589 1623 (Print) ● ISSN 1785 0037 (Online)
DOI: http://dx.doi.org/10.15666/aeer/1604_52695285
2018, ALÖKI Kft., Budapest, Hungary
[12] Günlü, A., Sivrikaya, F., Baskent, E. Z., Keles, S., Çakir, G., Kadiogullari, A. İ. (2008):
Estimation of stand type parameters and land cover using Landsat-7 ETM image: A case
study from Turkey. – Sensors 8(4): 2509-2525.
[13] Günlü, A., Ercanlı, İ., Başkent, E. Z., Çakır, G. (2014a): Estimating aboveground biomass
using Landsat TM imagery: A case study of Anatolian Crimean pine forests in Turkey. –
Annals of Forest Research 57(2): 289-298.
[14] Günlü, A., Ercanlı, İ., Sönmez, T., Başkent, E. Z. (2014b) Prediction of some stand
parameters using pan-sharpened IKONOS satellite image. – European Journal of Remote
Sensing 47(1): 329-342.
[15] Hall, R. J., Skakun, R. S., Arsenault, E. J., Case, B. S. (2006): Modeling forest stand
structure attributes using Landsat ETM+ data: Application to mapping of aboveground
biomass and stand volume. – Forest Ecology and Management 225(1-3): 378-390.
[16] Holmgren, P., Thuresson, T. (1998): Satellite remote sensing for forestry planning-A
review. – Scandinavian Journal of Forest Research 13(1-4): 90-110.
[17] Hu, T., Yang, J., Li, X., Gong, P. (2016): Mapping urban land use by using Landsat
images and open social data. – Remote Sensing 8(2): 151.
[18] Huang, C., Goward, S. N., Schleeweis, K., Thomas, N., Masek, J. G., Zhu, Z. (2009):
Dynamics of national forests assessed using the Landsat record: Case studies in eastern
United States. – Remote Sensing of Environment 113(7): 1430−1442.
[19] Hyde, P., Dubayah, R., Walker, W., Blair, B., Hofton, M., Hunsaker, C. (2006): Mapping
forest structure for wildlife habitat analysis using multi-sensor (LiDAR, SAR/InSAR,
ETM+, Quickbird) synergy. – Remote Sensing of Environment 102(1-2): 63-73.
[20] Kahriman, A., Günlü, A., Karahalil, U. (2014): Estimation of crown closure and tree
density using Landsat TM satellite images in mixed forest stands. – Journal of the Indian
Society Remote Sensing 42(3): 559-567.
[21] Kayitakire, F., Hamel, C., Defourny, P. (2006): Retrieving forest structure variables
based on image texture analysis and IKONOS-2 imagery. – Remote Sensing of
Environment 102(3-4): 390-401.
[22] Kelsey, K. C., Neff, J. C. (2014): Estimates of aboveground biomass from texture
analysis of Landsat imagery. – Remote Sensing 6(7): 6407-6422.
[23] Lu, D., Mausel, P., Brondizio, E., Moran, E. (2004): Relationships between forest stand
parameters and Landsat TM spectral responses in the Brazilian Amazon Basin. – Forest
Ecology and Management 198(1-3): 149-167.
[24] Lu, D., Chen, Q., Wang, G., Moran, E., Batistella, M., Zhang, M., Laurin, G. V., Saah, D.
(2012): Aboveground forest biomass estimation with Landsat and LIDAR date and
uncertainty analysis of the estimates. – International Journal of Forestry Research Article
ID: 436537.
[25] Mohammadi, J., Shataee, S., Yaghmaee, F., Mahiny, A. S. (2010): Modelling forest stand
volume and tree density using Landsat ETM data. – International Journal of Remote
Sensing 31(11): 2959-2975.
[26] Mohammadi, J., Shataee, S., Babanezhad, M. (2011): Estimation of forest stand volume,
tree density and biodiversity using Landsat ETM + Data, comparison of linear and
regression tree analyses. – Procedia Environmental Sciences 7: 299-304.
[27] Ni, X., Cao, C., Zhou, Y., Ding, L., Choi, S., Shi, Y., Park, T., Fu, X., Hu, H., Wang, X.
(2017): Estimation of forest biomass patterns across Northeast China based on allometric
scale relationship. – Forests 8(8): 288.
[28] Özçelik, R., Diamantopoulou, M. J., Brooks, J. R. (2014): The use of tree crown variables
in over-bark diameter and volume prediction models. – iForest-Biogeosciences and
Forestry 7(3): 132-139.
[29] Ozdemir, I., Karnieli, A. (2011): Predicting forest structural parameters using the image
texture derived from WorldView-2 multispectral imagery in a dryland forest, Israel. –
International Journal of Applied Earth Observation and Geoinformation 13(5): 701-710.
Sakici ‒ Günlü: Artificial intelligence applications using satellite data for stand parameters’ estimations
- 5285 -
APPLIED ECOLOGY AND ENVIRONMENTAL RESEARCH 16(4):5269-5285.
http://www.aloki.hu ● ISSN 1589 1623 (Print) ● ISSN 1785 0037 (Online)
DOI: http://dx.doi.org/10.15666/aeer/1604_52695285
2018, ALÖKI Kft., Budapest, Hungary
[30] Poudel, K. P., Cao, Q. V. (2013): Evaluation of methods to predict Weibull parameters
for characterizing diameter distributions. – Forest Science 59(2): 243-252.
[31] Pu, R., Cheng, J. (2015): Mapping forest leaf area index using reflectance and textural
information derived from WorldView-2 imagery in a mixed natural forest area in Florida,
US. – International Journal of Applied Earth Observation and Geoinformation 42: 11-23.
[32] Reis, A. A., Acerbi, F. W., Mello, J. M., Carvalho, L. M. T., Gomide, L. R. (2018):
Relationship between spectral data and dendrometric variables in Eucalyptus sp. stands. –
Floresta e Ambiente 25: e20150170.
[33] Safari, A., Sohrabi, H. (2016): Ability of Landsat-8 OLI derived texture metrics in
estimating aboveground carbon stocks of Coppice oak forests. – Proceedings of the XXIII
ISPRS Congress, 12-19 July, Prague, Czech Republic.
[34] Sakici, O. E., Sağlam, F., Seki, M. (2018): Single- and double-entry volume equations for
Crimean pine stands in Kastamonu Regional Directorate of Forestry. – Turkish Journal of
Forestry 19(1): 20-29.
[35] Santi, E., Paloscia, S., Pettinato, S., Chirici, G., Mura, M., Maselli, F. (2015): Application
of Neural Networks for the retrieval of forest woody volume from SAR multifrequency
data at L and C bands. – European Journal of Remote Sensing 48(1): 673-687.
[36] Shataee, S. (2013): Forest attributes estimation using aerial laser scanner and TM data. –
Forest Systems 22(3): 484-496.
[37] Shoemaker, D. A., Cropper, W. P. (2010): Application of remote sensing, an artificial
neural network leaf area model, and a process-based simulation model to estimate carbon
storage in Florida slash pine plantations. – Journal of Forestry Research 21(2): 171-176.
[38] URL-1 (2014): USGS Global Visualization Viewer. – http://glovis.usgs.gov (accessed on
30 September 2014).
[39] Wallner, A., Elatawneh, A., Schneider, T., Knoke, T. (2015): Estimation of forest
structural information using RapidEye satellite data. – Forestry: An International Journal
of Forest Research 88(1): 96-107.
[40] Xie, S., Wang, W., Liu, Q., Meng, J., Zhao, T., Huang, G. (2017): Estimation of forest
stand parameters using SPOT-5 satellite images and topographic information. – Preprints
2017100017.
[41] Yang, S., Feng, Q., Liang, T., Liu, B., Zhang, W., Xie, H. (2018): Modeling grassland
above-ground biomass based on artificial neural network and remote sensing in the
Three-River Headwaters Region. – Remote Sensing of Environment 204: 448-455.
[42] Zhao, P., Lu, D., Wang, G., Wu, C., Huang, Y., Yu, S. (2016): Examining spectral
reflectance saturation in Landsat imagery and corresponding solutions to improve forest
aboveground biomass estimation. – Remote Sensing 8(6): 469.
[43] Zheng, D., Rademacher, J., Chen, J., Crow, T., Bressee, M., Le Moine, J., Ryu, S. R.
(2004): Estimating aboveground biomass using Landsat ETM+ data across a managed
landscape in northern Wisconsin, USA. – Remote Sensing of Environment 93(3): 402-
411.