Agricultural drought prediction using machine learning.

Abstract

This project develops a short-term agricultural drought prediction system for Uganda using satellite-derived Vegetation Condition Index (VCI) data and Long Short-Term Memory (LSTM) deep learning models. Monthly VCI data from Google Earth Engine (2000-2022) is used in time series analysis, including decomposition, autocorrelation, and Mann-Kendall tests, to identify seasonal drought patterns and long-term trends. A Bidirectional LSTM model is trained on gridded VCI data from 107 locations in Uganda to forecast drought conditions. The model achieves approximately 75% prediction accuracy (1 - normalized RMSE), demonstrating its ability to capture Uganda's characteristic seasonal patterns, evidenced by strong autocorrelation at 12-months lags. Resulting forecasts are used to generate drought severity maps and planting advisories, supporting an early warning system that provides timely recommendations to farmers on optimal planting windows. The system aims to enhance agricultural planning and mitigate drought impacts on food production and food security in Uganda.