A HYBRID LSTM–DNN DEEP LEARNING FRAMEWORK FOR CROP YIELD PREDICTION IN PRECISION AGRICULTURE

Abstract

Crop yield prediction sits at the core of precision agriculture, directly influencing food security planning, resource allocation, and economic decision-making. The problem is genuinely difficult: yield depends on dynamic factors that shift daily (temperature, precipitation) and static factors that barely change season to season (soil chemistry, elevation). Conventional models—whether simulation-based or standard machine learning—tend to handle one side well but not both. This paper presents a Hybrid Long Short-Term Memory–Deep Neural Network (LSTM–DNN) framework that treats these two data types separately before combining them. The LSTM branch handles sequential climatic records to extract temporal dependencies; the DNN branch processes soil and geographic constants to capture non-linear agronomic relationships. Their outputs are concatenated for final yield regression. Tested on a dataset comprising 15 years of weather, yield, and soil records, the model achieved a Mean Absolute Error (MAE) of 0.370 T/ha, a Root Mean Square Error (RMSE) of 0.674 T/ha, and an R² of 0.979— outperforming Random Forest (RF), Support Vector Machines (SVM), 1D-CNN, and standalone LSTM baselines by a wide margin.