Evaluation of models and drought-wetness factors contributing to predicting the vegetation health index in Dak Nong Province, Vietnam

Monitoring and predicting vegetation health are essential for agricultural activities and food security. This study aimed to select a model and evaluate the factors contributing to predicting the vegetation health index (VHI) in the Dak Nong Province, Vietnam. Machine learning algorithms were evaluated, including multiple linear regression, xGBoost, and artificial neural networks (ANN). The input variables of the models included the standardized precipitation evapotranspiration index (SPEI), soil moisture (SM), and VHI in the previous periods. Research results showed that the ANN model gave the best prediction results. The accuracy of prediction results depended on the season of the year, in which the dry season gave a result with high accuracy. The results also indicated that SM from one to two previous months, SPEI1 from one to three previous months, SPEI3 and SPEI5 from three to six previous months, and VHI from one previous month contributed mainly to the prediction model. The relative contribution of SM and SPEI ranged from 42% to 52% in the last 4 months of the dry season. In addition, land use type also affected prediction quality.


Introduction
The health of vegetation reflects the severity of drought and the condition of the land cover.They are also used in forecasting biomass and yield growth.It can also be seen that the drought-wetness conditions in the previous period had an apparent effect on the current vegetation health status, and the level of influence varied between different regions [1,2].Therefore, monitoring and predicting vegetation health status is necessary.It provides early warnings to help provide timely actions to ensure agricultural activities and food security, particularly in the context of climate change [3].This is especially important for agricultural areas in developing countries, such as the Dak Nong Province of Vietnam, where the level of farming techniques is still backward and agricultural production still largely depends on natural conditions such as soil type, precipitation, soil moisture (SM), extreme weather events, etc.
The vegetation health status can be expressed through vegetation indices, which can be extracted from remote sensing data.Many remote sensing-based vegetation indices have been developed and widely used in research related to drought assessment, crop health assessment, and crop yield forecasting, such as the Normalized Difference Vegetation Index (NDVI), Vegetation Condition Index (VCI), Vegetation Health Index (VHI), Enhanced Vegetation Index (EVI) [4][5][6][7][8], which VHI provides good support for the assessment of drought and crop responses [3].Because VHI is calculated from Vegetation Condition Index (VCI) and Temperature Condition Index (TCI), it reflects the state of vegetation and the effects of drought and surface temperature on its growth [9].VHI is closely related to crop growth and yield, which are associated with drought stress [10,11].According to Hiep et al [12], Van Viet and Thuy [13], Luong and Bui [14], VHI is the most suitable index for monitoring vegetation health status in Dak Nong province, which was our study area.Thus, VHI was chosen as the predicted variable in this study.
Machine learning, a state-of-the-art approach, has been widely applied to build prediction models.Many parametric and non-parametric machine learning algorithms, from simple to complex, such as Multiple Linear Regression (MLR), eXtreme Gradient Boosting (xGBoost), Artificial Neural Networks (ANN), Support Vector Machine (SVM), and Random Forest (RF) have been applied and proven to provide results with high accuracy [15][16][17][18][19].The performance of these algorithms is competitive, depending on the dataset.Therefore, choosing an optimal model for the available dataset of the study area is necessary.
With these issues in mind, this study aimed to (1) build and select an optimal VHI prediction model for Dak Nong province of Vietnam and (2) evaluate the factors contributing to the prediction.Machine learning algorithms were evaluated, including MLR, xGBoost, and ANN.Factors considered to build the prediction model included meteorological drought indices, SM, and VHI in previous periods.The study area is located in the tropical monsoon region with a 6-month dry season and is often affected by drought.The questions for this region are: (1) How do drought-wetness conditions in the previous periods affect the VHI by month of the year?(2) At what time of year is the VHI significantly affected by drought-wetness conditions?(3) Can drought indices from previous periods be used in VHI prediction?Moreover, how? (4) How much do soil moisture and precipitation from previous periods contribute to the vegetation health index?and (5) Which soil and land use types are sensitive to drought-wetness conditions in previous periods?This study was conducted to solve these questions and to meet the requirements of crop management.

Study area
Dak Nong province is located in the southwest of the Central Highlands, with geographical coordinates from 11°45′ to 12°50′ N, 107°13′ to 108°10′ E (figure 1).Its average altitude is about 600 m to 700 m and, in some places, up to 1982 m above mean sea level.The province has a diverse terrain alternating between high mountains, large and flat plateaus, and low-lying plains.Dak Nong's climate is typical of a sub-equatorial tropical monsoon climate.However, with terrain elevation, it is characterized by a humid tropical highland climate and is also influenced by the hot and dry southwest monsoon.There are two seasons in a year: The rainy season is from about April to the end of November (accounting for over 80% of the yearly rainfall), and the dry season is from December to the end of March next year.The average annual rainfall is about 2513 mm.For soil, the study area mainly consists of Rhodic Ferralsols, Plinthic Ferralsols, Xanthic Ferralsols, Chromic Luvisols, Haplic Arenosols, Arenic Acrisols, and Gleyic Luvisols, which are favorable for cultivating coffee, rubber, tea, pepper, cashews, rice, corn, and some short-term industrial crops [20].

Data
The overall data and workflow are shown in figure 2. The prediction models' input data included Standardized Precipitation Evapotranspiration Index (SPEI), SM, and VHI from 2000 to 2022.The details for obtaining these input data are as follows.
VHI was calculated from VCI and Temperature Condition Index (TCI).Meanwhile, VCI and TCI were calculated from NDVI and Land Surface Temperature (LST), respectively, which were obtained from the Moderate Resolution Imaging Spectroradiometer (MODIS) data from the United States Geological Survey (USGS) website via https://earthexplorer.usgs.gov/.Specifically, NDVI was obtained from the MODIS/Terra Vegetation Indices 16-Day L3 Global 500 m version 6 (MOD13A1 v006) product with a spatial resolution of 463 m × 463 m and a temporal resolution of 16 days.Meanwhile, LST was obtained from the MODIS/Terra Land Surface Temperature/Emissivity 8-Day L3 Global 1 km version 6 (MOD11A2 v006) product with a spatial resolution of 926.6 m × 926.6 m and a temporal resolution of 8 days.For analyzing the effects of drought on crops, input variables were needed to have a common temporal and spatial resolution.In terms of time, monthly data were chosen for calculation.To reduce the influences of atmospheric conditions, the monthly NDVI data were obtained by applying the Maximum Value Composite method [21,22].The monthly average LST data were calculated and then resampled using the Bilinear method to match the spatial resolution of NDVI (463 m × 463 m).
Vegetation was susceptible to precipitation and soil moisture [23][24][25], so the SPEI and SM were chosen as predictor variables.SPEI was a meteorological drought index commonly used in drought assessment [26,27].SPEI was calculated based on precipitation, temperature, and sunshine duration.These data were from 10 meteorological stations, as shown in figure 1.These were stations with relatively complete data.The missing data were only about 6% and is supplemented by linear regression based on data from neighboring stations.SM was collected from the Land Data Assimilation System (LDAS) website via https://ldas.gsfc.nasa.gov/gldas.The data had a spatial resolution of 0.25°× 0.25°and a temporal resolution of one month.The SM data were then resampled to 463 m × 463 m using the bilinear method.

Methods
In this study, machine learning methods were used for the forecasting process.Details are presented as follows.

Input variables
• SPEI SPEI was a drought index calculated based on the difference between precipitation and potential evapotranspiration (PET) [28].For any timescale, given P i and PET i be respectively the amount of precipitation and potential evapotranspiration during time i, the difference D i was determined as follows: After the D i series were calculated, the best-fit distribution function for the series was estimated.The probability of each D i value was calculated using the cumulative distribution function as equation (2).It was called the SPEI [28].
where W = (−2ln P) 0,5 for P 0.5.If P > 0.5, P was replaced by 1-P, and the sign of the SPEI was reversed.
PET is determined according to the Thornthwaite function [29] as follows: where: o T was the monthly average temperature (°C).
o K was a coefficient based on latitude and month of the year, and it was calculated as follows: where NDM was the number of days in the month, and N was the number of sunshine hours calculated based on the astronomical formula.
The SPEI was calculated for each meteorological station and then was interpolated using the Inverse Distance Weighted (IDW) method to obtain a raster layer with a resolution of 463 m × 463 m.This variable was built for a timescale from 1 to 12 months.

• SM
As mentioned, the monthly SM data at a spatial resolution of 0.25°× 0.25°were resampled to 463 m × 463 m using the bilinear method to match the spatial resolution of other variables.
• VHI VHI was developed to evaluate the health of vegetation based on a combination of VCI and TCI [30].A VHI value larger than 60 reveals that the vegetation was in good condition, while the value lower than 40 indicates that the vegetation was in a stress status.It was calculated as follows: where the weight of α typically was set to 0.5; meanwhile, VCI and TCI were the normalized values of NDVI and LST over time, respectively [31].They were calculated as follows: where NDVI i was the NDVI value of a particular pixel in a specific year at time i, while NDVI max and NDVI min were the maximum and minimum NDVI values over a period of analysis, respectively.
where LST i was the LST value of a particular pixel in a specific year at time i, while LST max and LST min were the maximum and minimum LST values over a period of analysis, respectively.
In addition to being a predicted variable in the model, the VHIs in previous periods were also used as the predictor variables.

Machine learning
VHI prediction models were constructed based on the predictor variables SPEI1, SPEI3, SPEI5, SM, and VHI, which were taken from the previous months of the predicted month.It should be noted that SPEI was calculated on a range of timescales from 1-12 months, and the number following SPEI indicated the timescale.The study area had six months of dry season and six months of rainy season, and this study focused on analyzing variability in drought-wetness conditions in short periods that affect crops.Therefore, a timescale of one, three, and five months of SPEI was selected.In addition, the time variable (t) was included in the model to take into account the effects of the trend.This variable is calculated as the number of months since the first month of available data.
Models were constructed for each month to eliminate the effects of the growing season.500 grid cells were selected with an even distribution.50% of cells are used for training, 20% for validation, and 20% for testing.The data were standardized (a mean of zero and a standard deviation of one) and were reassembled into a series at each grid cell.

• Multiple Linear Regression
The VHI was predicted using stepwise multiple regression.After each iteration, a potential predictor was added to or removed from the model based on testing statistical significance.The equation for predicting VHI was as follows: where Ys was the simulation value of VHI, t was a time variable, and x i was the predictor, which was SPEI, SM and VHI in the previous periods.
• Extreme Gradient Boosting Extreme Gradient Boosting (XGBoost) [32] is based on the Gradient Boosting Machine (GBM) algorithm [33][34][35].XGBoost efficiently constructs boosted trees and solves regression and classification problems.The basic principle of this algorithm is to minimize an objective function.It implements machine learning algorithms under the Gradient Boosting framework.In this study, the XGBoost regression model was used with the selection of booster type and adjustment of model parameters.

• Artificial neural networks
The ANN is a deep learning model that includes hidden layers with neurons in each layer.This model uses backpropagation for the training process [36].To build the model in this study, the trial-and-error method was used to choose the activation function, the number of hidden layers, the number of neurons in each hidden layer, and the values of epoch and batch size parameters.
In this study, the xGBoost model was implemented using the XGBoost Python Package; meanwhile, splitting the dataset and evaluating the simulation results were conducted using scikit-learn library in the Python environment.The other machine-learning models were built in the Fortran environment by authors.

Evaluating the prediction results and the contribution of variables
For validating the accuracy of the VHI prediction and identifying the statistical significance of predictor variables in the model, the parameters of the correlation coefficient (R), the Root Mean Square Error of Prediction (RMSE), and Willmott's index of agreement (d) were used.These statistical parameters were calculated as follows: where n was the length of the series, Ys was the simulation value of VHI, Y was the actual value of VHI, and Y ¯was the average of Y.These values could be normalized or transformed into a standardized form.
The contribution of a variable in the model was assessed by the decrease in the adjusted R square when that variable was removed from the model.

The relationship between the predicted variable and the predictors
The purpose of this analysis was to select predictor variables.As mentioned, the predictors in the analysis for the VHI prediction included SPEI1, SPEI3, SPEI5, SM, and VHI in the previous months.In other words, they needed to be taken before the predicted month by a period from one to several months.This period was called the lag time, and it needed to be determined to ensure the accuracy of the prediction.
From data derived at grid cells, the correlation coefficient between VHI and predictors in two seasons was calculated (figure 3).The result showed that the correlation coefficient in the rainy season was lower than in the dry season.In the dry season, VHI at a lag time from one to two months gained a good relationship with VHI of the predicted month, and SM and SPEI1 at a lag time from one to four months had a pretty good relationship with the predicted variable.In the rainy season, only the VHI of the month immediately preceding the predicted month had a fairly good relationship with the VHI of the forecast month.It showed that the ability to predict VHI in the rainy season months was not good.
The last dry season months, February to April, gained the highest correlation coefficients between VHI and predictors (figure 4).In addition, figures 3(a) and 4(b) indicated that a 1-month lag gave the highest correlation coefficient between SM and VHI.They also showed that soil moisture in a given month affected crop health several months later.
Figures 4(c) and (d) revealed that the lag time corresponding to the highest correlation coefficient depended on the timescale of SPEI.Because the SPEI value was calculated and recorded for the last month of the timescale, figures 3(a), 4(c), and (d) also showed that the precipitation in the last month of the rainy season and the first month of the dry season, from September to November, contributed to the development of vegetation.Due to the positive correlation coefficient, drought in these months may negatively affect vegetation and vice versa.
It can be seen that the highest correlation coefficients between some predictors and VHI occurred at different lag times, and there were many differences in the magnitude of correlation coefficients by season and month.Therefore, constructing a VHI prediction model by months may be more appropriate than by seasons and years, even though this may reduce the length of the input series for machine learning.In addition, this result also showed that it was possible to choose the lag time of the predictors VHI and SM from one to two months, SPEI1 from one to three months, and SPEI3 and SPEI5 from three to six months.

VHI prediction models
As mentioned, the selected predictor variables included VHI and SM from one to two months, SPEI1 from one to three months, and SPEI3 and SPEI5 from three to six months before the predicted month.Thus, there were 16 predictor variables, including the time variable.The VHI prediction models below were built for each month.Because VHI prediction results for a period of two months or more did not show high accuracy, below is only presented with a prediction period of one month.

Models using MLR and XGBoost
For models using MLR, the variables selected in the VHI prediction model based on the stepwise regression method are presented in table 1.In this table, the variables are arranged in order of selection, and the numbers after the variable name indicate the lag time.It can be seen that only two to three variables were included in each model.VHI.1, SM.1, SPEI1.1,SPEI3.4,and SPEI3.5 were variables participating in the models, depending on each case.It means that the variables included in these models were VHI, SM, and SPEI1 of one month and SPEI3 of four and five months before the predicted month.It can be seen that VHI.1 was always the first selected variable.The following essential variables were SPEI1.1,SM.1, SPEI3.5, and SPEI3.4.Two variables that did not participate in the models were time (t) and SPEI5.Table 2 illustrates the coefficients of the regression model for March.The obtained coefficients showed that the VHI prediction models for all months were dependable.The accuracy of prediction results was uneven across months.It was revealed by the value of the correlation coefficient and the Willmott index (figure 5).It can be seen that from May to October, the accuracy of the prediction results was lower than in the remaining months.It was when rainfall accounted for over 82% of the year.The remaining months had pretty high d and R.They were also the dry season months in the study area.It means that during this period, VHI was more sensitive to drought, and the ability to predict VHI may be better.In addition, figure 5 also showed that there was a clear relationship between the VHI prediction quality and its standard deviation.The months with high VHI prediction ability were the months with strong VHI fluctuations in the dry season.
For models using xGBoost, the gblinear booster was chosen for our data because it gave significantly better results than the gbtree booster.Using gblinear booster, the xGBoost model input parameters were built using the trial-and-error technique so that RMSE was the smallest and R and d were the highest.Since the booster was gblinear, the results were not significantly different from MLR.Compared to MLR, the quality of xGBoost had a slight decrease if all variables were used, the quality was the same if using the input variables shown in table 1, and the quality had a negligible increase if removing variables with small contributions.In the dry season, the predictor variables retained in this model were VHI, SM, and SPEI1, with a lag time of one to two months, and SPEI3 and SPEI5, with a lag time of four to five months.In the rainy season, the retained predictor variables were VHI, SM, and SPEI1 for a 1-month lag time and SPEI3 for a 4-to 5-month lag time.SPEI5 was not used.

Models using ANN
The ANN model was constructed using the trial-and-error method so that RMSE was the smallest and R, d was the highest.The result showed that the sigmoid activation function was appropriate.The model had two hidden layers: the first had 16 neurons, and the second had 8 neurons.Because there were too many parameters about the weights of neurons during the propagation and the bias of each neuron, they are not presented here.The characteristics of dispersion and forecast error were similar across months.Figure 6 illustrates these characteristics for April.The scatter plots showed that when actual VHI was small, predicted VHI tended to be positively skewed, and vice versa.VHI values around 50% had the most minor predicted error.Thus, the ANN model also tended to reduce the amplitude of the predicted value compared to reality, which often happened in the MLR model.
The results of evaluating VHI prediction accuracy for each month based on the validation dataset are presented in figure 7. Similar to MLR models, ANN models for the dry months brought greater accuracy.Compared with MLR and xGBoost, the prediction quality yielded a significant improvement (figure 7(a)).Therefore, ANN was the most suitable choice in this case study.According to figure 7(b), although R and d were lower than the other months from June to October, RMSE was relatively low, which was related to the low standard deviation of VHI in these months (as shown in figure 5).
Since this method provides higher accuracy than MLR and xGBoost, the contribution of predictors to the variation of VHI was analyzed.The relative contribution of key predictors to VHI prediction accuracy is shown in figure 8.The results showed that the main contributing variable was the VHI of 1-month lag time, which was about 48% to 58% from January to May and from 65% to 78% in the remaining months.In short, in the rainy early dry seasons, the VHI of the predicted month mainly depended on the VHI of the immediately preceding month.It also means that the contribution of the remaining variables was from 42% to 52% from January to May and was from 22% to 35% in the remaining months.Among the remaining variables, the three main contributing variables were SM.1, SPEI1.1, and SPEI3.5, in descending order.According to figure 8, the contribution of drought-wetness conditions to VHI forecast quality was most obvious in the last 4 months of the dry season (January to April).Furthermore, the soil moisture from the previous month played an important role in VHI simulation during this period.These results were consistent with the study of Hammad and Falchetta [3] to some extent, although they used a probabilistic forecasting approach.The predictors in their model were VHI and separate climate variables such as precipitation, temperature, humidity, percent of cloud, and solar radiation in previous periods.Their results showed that VHI and climate variables from two weeks to one month in advance gave the best prediction results.They did not use the SM variable in the model.
In this study, the raw SM data had a spatial resolution of 0.25°× 0.25°, and they were resampled to match a common spatial resolution of 463 m × 463 m using the bilinear method.The resampled SM data gained a pretty good relationship with VHI (figure 3(a)), and the contribution of SM to the VHI forecast quality was relatively high from December to April (figure 8).This result demonstrated that although the input SM data did not have a high spatial resolution, they still provided enough information for this study.
The limitations of this study consist of low prediction quality in some months and short prediction time.Adding additional predictor variables to increase prediction quality may be necessary.In addition, the contribution of TCI and VCI at different weights to VHI has not been considered, although their weights vary depending on the aridity of the study region [37].Furthermore, other prediction approaches, such as other machine learning techniques [11,38] and probabilistic forecasting [3], also should be experimented with.These issues need to be taken into account in future studies.

Other factors related to prediction quality
Because the prediction quality in the rainy season was not good, the following analysis focused on the results in the dry season.The evaluation results were based on the ANN model, which was considered the best prediction model in this study.Factors related to the quality of VHI prediction considered in this analysis consisted of soil and land use types.Figures 9(a To detect the difference in prediction quality between different land uses and soil types, a t-test for Willmott's index of agreement at p = 0.01 was used.The results of this analysis are shown in tables 3 and 4. In these tables, the symbol b (bottom) means no difference in the average value of d of the land use type (or soil type) in the row noted compared to the land use type in the lower row.Similarly, the symbol t (top) is used to compare to the upper row.The analysis results revealed that only land use type clearly showed a difference in prediction quality.
Regarding land use (table 3), the prediction quality mostly had apparent differences between land use types except for the pair AnC and PeA.The RMSE, R, and d showed that except for paddy rice, the forecast accuracy gradually decreased in the order of annual crops, perennial crops, planted forests, and natural forests.In this study area, annual crops were mainly corn and cassava (60% of the area), beans and vegetables (18%), soybeans, peanuts, sweet potatoes, and sugarcane.Perennial crops included coffee, cashew, pepper, and macadamia, of which Robusta coffee is the main.These annual and perennial plants often had shallow rooting systems that may be susceptible to drought.With planted forests, although they had deep root systems, they were still more  susceptible to the effects of drought than natural forests because they were planted on steep terrain and low density of rivers and streams.For wet rice, because it was prioritized for growing in places with favorable water sources, the impact of drought was less than other crops.Thus, the more sensitive the plants were to changes in soil moisture, the higher the predictability.This was consistent with the contribution of SM and SPEI variables in the ANN model.4), prediction accuracy did not differ significantly between soil types.The maximum difference in d between soil types was only 0.08, while this value was 0.17 for land use.The accuracy only had a clear difference between two soil groups: FRx, LVx, LVg, FRr (group 1) and FRp, Ach, ACa (group 2).d of group 1 was between 0.88 and 0.90, and the one of group 2 was between 0.82 and 0.84.
Table 5 shows the percentage of area by land use type and soil type in the study area.For the soil types in group 1, it can be seen that the area belonging to the land use types AnC, PeA, and PdR was always larger than the area belonging to the land use types PdF, SpF, and PtF.For soil types in group 2, this happened in reverse.Because AnC, PeA, and PdR always had higher accuracy in predicting VHI than the remaining land uses (as shown in table 3), the accuracy in VHI prediction on soils in group 1 was higher than the other soil group partly due to the plants grown on the soil.
In addition to crop factors, according to statistics in table 6, soils in group 1 were more favorable for growing crops than soils in group 2. In other words, plant care in group 1 is usually done well, and therefore, the VHI of this group was only affected by unusual weather conditions.Thus, it was easier to predict than in group 2.

Conclusions
In summary, this study aimed to select a model and evaluate drought-wetness factors contributing to the prediction of VHI in Dak Nong province of Vietnam.The MLR, xGBoost, and ANN models were evaluated; meanwhile, the SPEI, SM, and VHI in previous periods were considered predictors for building the prediction model.
By analyzing the correlation coefficient, the results showed that SM and SPEI in the previous period gained a pretty good relationship with VHI at the predicted time.Specifically, VHI had a high correlation coefficient with SM at a lag time of one to two months, with SPEI1 at a lag time of one to three months, and with SPEI3 and SPEI5 at a lag time of three to six months.In addition, the VHI of the predicted month had a good relationship with the VHI of the immediately preceding month.These characteristics contribute to the ability of the model to predict VHI.
Regarding model selection, ANN gave higher VHI prediction accuracy than MLR and xGBoost.The results also revealed that the relative contribution of SM and SPEI was high in the last 4 months of the dry season and accounted for 42% to 52%.The accuracy of prediction results depended on the year's season, and it was possible to predict VHI for the dry season one month in advance with high accuracy.Furthermore, the best VHI prediction ability was for the land use group susceptible to drought impacts, which included AnC, PeA, and PdR.
As mentioned, this study still has some limitations related to low forecast quality and short prediction time, the weight of VCI and TCI contributing to VHI, as well as the experiments with other forecasting approaches.These issues need to be taken into account in future studies.

Figure 1 .
Figure 1.Study area and location of meteorological stations.

oo
I was the heat index.It was calculated based on the temperature of 12 months as follows: m was a coefficient based on I as follows:

Figure 3 .
Figure 3. (a) Correlation between predictor variables and VHI (a) in the dry season and (b) in the rainy season.Note: Rcr is the critical value of Pearson's correlation coefficient at p = 0.01.

Figure 5 .
Figure 5. Values of R, d between observed and predicted results using MLR and standard deviation of VHI by month.

Figure 6 .
Figure 6.(a) Dispersion of actual and forecasted VHI values and (b) forecast error for April by using ANN.
)-(c) present the average values of the prediction quality assessment indices for the dry season.Figures9(d) and (e) show land use and soil types whose abbreviations are explained in tables 3 and 4.

Figure 7 .
Figure 7. (a) Difference in R between using xGBoost and ANN compared to MLR; and (b) VHI prediction accuracy using ANN.

Figure 8 .
Figure 8. Relative contribution in percentage of key predictors to VHI prediction results.

Figure 9 .
Figure 9. Assessment of VHI prediction quality for the dry season through statistical indices: (a) RMSE, (b) R, and (c) d; and the spatial distribution of (d) land use types and (e) soil types in the study area.

Table 1 .
Predictor variables in the models using MLR.

Table 2 .
Coefficients of the MLR model for March with standardized data.

Table 3 .
Prediction accuracy using the ANN by land use types.

Table 4 .
Prediction accuracy using the ANN by soil types.

Table 5 .
Percentage of area by land use and soil types.

Table 6 .
Characteristics of soil types.