Evaluation of Combined Agricultural Drought Index (CDI), Prediction Methods Based on Satellite Images via Deep Learning and Machine Learning Approaches

Drought is one of the greatest challenges of our time due to the dangers it poses to the world. In arid and semi-arid regions, it is necessary to continuously monitor agricultural systems that face water shortages and frequent droughts. Therefore, it is necessary to have large-scale information about agricultural systems and land use for managing and making decisions for the sustainability of food security. Continuous monitoring of drought requires a large amount of information to be processed with great speed and accuracy. Due to the complexity and impact of various factors on drought, in recent years, the methods of combining several factors to create a comprehensive drought index have received much attention. Machine learning and deep learning methods can provide a more accurate and efficient tool to predict droughts and be used in drought risk management. The review of sources shows that until now no studies have been conducted in the field of drought monitoring using deep learning approach and satellite images in the catchment area of Lake Urmia in Iran. A large part of its economic activities is dedicated to agriculture. The increase in temperature, the increase in evaporation-transpiration and the excessive use of water resources for agriculture have caused an upward trend in the frequency of droughts in this basin during consecutive years, one of the harmful effects of which is a significant decrease in the lake level. Therefore, for drought management in this basin, it is very important to identify drought behavior so It is very important to determine appropriate and reliable indicators to measure and predict the effects of droughts. According to the investigations, it was observed that most of the studies in the field of drought in this basin have been carried out from the meteorological point of view, or by individual plant indicators, so in this study, using the approach of principal component analysis, we tried to provide a composite drought index for drought modeling and forecasting.

In this research, satellite images and deep learning and machine learning methods have been used to predict the Combined Drought Index. For this purpose, satellite images were first obtained for the study area and pre-processing was done on the data. Then, all the data were converted to a scale with a spatial resolution of 500 meters, and the VCI index was calculated using NDVI data, the TCI index using the land surface temperature product, and the CWSI index using the Modis evapotranspiration product, and finally, CDI drought index was calculated using principal component analysis method. Then the correlation between CDI data and other meteorological variables including evapotranspiration, potential evapotranspiration, land surface temperature during the day, and land surface temperature at night was calculated. Finally, the CDI index is modeled using deep learning and machine learning methods.

This study modeled the Combined Drought Index based on a different combination of input variables and deep learning and machine learning methods. Examining the results showed that the variables of the normalized difference vegetation index, the land surface temperature during the day and at night, evapotranspiration, and potential evapotranspiration were the most influential parameters for modeling the CDI index, and all four methods with acceptable accuracy and error have been able to model the combined drought index. The CART model with a correlation coefficient of 0.96, RMSE equal to 0.029, and Nash Sutcliffe coefficient of 0.92 was chosen as the best model among the methods.

In this research, different combinations of input variables extracted from satellite image products were evaluated in the form of 6 independent scenarios to predict the Combined Drought Index. By examining the evaluation parameters including correlation coefficient, Nash Sutcliffe coefficient, and root mean square error, it was found that all four methods can estimate the combined drought index with acceptable accuracy and error. Among all the methods, the CART method performed better (R=0.96 and RMSE=0.029) than the other methods for predicting the time series of the Combined Drought Index. On the other hand, the SVM method has been able to model the combined drought index with acceptable accuracy (R=0.94 and RMSE=0.034). However, contrary to expectations, two deep learning methods were able to model the combined drought index with less accuracy than machine learning methods. In general, by examining the results, it was found that with the method presented in this research, it is possible to accurately predict the CDI combined drought index time series and predict drought in different periods of plant growth, and use its results for regional drought management and policies, especially in Basins without statistics.