Deep Learning Oriented Satellite Remote Sensing for Drought and Prediction in Agriculture

Drought is a challenging problem in agriculture due to its random and nonlinear nature. Moreover, in bad weather situations, satellites do not capture the precise data of the earth, which ultimately affects the training of machine learning models. This project proposes an approach for drought prediction and anti-drought strategies. The task is divided into two parts- (1) Predicting the short-term severity of drought, (2) Crop selection. Based on time duration, agricultural drought occurs after meteorological drought. Meteorological drought leads to the deficiency of precipitation, increased evaporation, and transpiration. Therefore, we employed the Standard Precipitation and Evapotranspiration Index (SPEI) to predict the first-stage by passing it to Long-Short Term Memory (LSTM) oriented neural network demonstrated an accuracy of 74%. Subsequently, for agricultural drought, which leads to reduced biomass yield and soil water deficiency, indices known as Normalized Difference Vegetation Index (NDVI) and Soil Moisture Index (SMI) are formulated. Stage 2 prediction is made using these indices as input to time distributed Convolutional Neural Network (CNN) network, which demonstrated an accuracy of 96%. Furthermore, based on the proposed methodology, anti-drought strategies are suggested to improve the agriculture outcome.