Predicting Crop Yields Using Satellite Data and ML

Abstract

Agribusiness planning, economic security, and world food security lie in the proper prediction of crop yields. The research paper introduces a sophisticated machine learning model that exploits the application of satellite imagery and climatic data in order to accurately predict crop yields. This system combines the multi-spectral satellite data (Sentinel-2 2, Landsat 8 ) measuring the most important vegetation indices (NDVI, EVI ) with the weather variables (precipitation, temperature, soil moisture ) to provide accurate yield predictions several weeks before the harvest. We fit XGBoost, Random Forest, and Long Short-Term Memory (LSTM) machines and perform a combination of machine learning techniques altogether based on the hybrid approach. The model, trained with five years of data in three large corn fields in the USA (corn, wheat, soybean), has an accuracy prediction of 92.4 percent (R2 score) with regards to predictions of corn yield, 27 percent better than conventional models. Because the system offers early yield predictions (8-12 weeks before harvest) at less than 10% average relative error, profound yield-limiting parameters, including drought tension and nutrient shortages, may also be detected. The cloud-based design of the framework allows scalable deployment and thus is available to large-scale agribusiness as well as to smallholder farmers. The usage advantages of field validations include precision farming, point-to-point product market forecasting, and climate response strategy. The study finds application in the sustainable intensification of food production in the sense that it would provide information that would be used in making agricultural decisions optimally.