Application of deep learning in forecasting crop water stress in Kitenga Sub-County, Mubende District

Abstract

Water stress poses a serious threat to food security, leading to considerable financial losses for both farmers and the broader national economy. Accurate assessment of crop water stress will enhance agricultural productivity. Several direct and indirect methods for crop water stress detection exist, but they are tedious and require highly sophisticated equipment. This study aims to evaluate spatial and temporal patterns of crop water stress and forecast future stress levels using satellite-derived data and deep learning techniques, specifically the Bidirectional Long Short Term Memory (BiLSTM). The model was trained on CWSI time series data of 2000 to 2020 derived from satellite imagery. The results indicate significant temporal fluctuations in water stress, with notable peaks aligning with various months within the seasons. The BiLSTM model achieved a RMSE of 0.097, a MAE of 0.0691 and R2 score of 0.82, which indicates good performance of the model. These results suggest that this model can be used to effectively forecast crop water stress.