Crop Yield Prediction Using a Hybrid Convolutional and Recurrent Neural Network Model

Abstract

Precise forecasting of crop yields signifies a critical authority for addressing global food security challenges, optimising supply chain logistics, and informing evidence-based agricultural policy formulation. Conventional yield prediction methodologies frequently demonstrate limitations in capturing the intricate, non-linear interdependencies among meteorological conditions, edaphic properties, and cultivation practices that collectively determine agricultural productivity. This research proposes CLDNet, a novel deep learning framework that architecturally integrates Convolutional Neural Networks with Long Short-Term Memory networks to predict soybean yield across the United States. The CNN module specialises in discerning spatially-distributed patterns from multidimensional environmental inputs, while the LSTM component effectively models temporal sequences and dependencies throughout the crop development cycle. Our framework was developed and validated using an extensive public dataset from Kaggle, incorporating historical yield records, meteorological measurements, and pedological characteristics. Empirical findings indicate that our integrated CLDNET paradigm surpasses the predictive performance of conventional approaches, including Random Forest, Support Vector Regression (SVR), and isolated LSTM configurations, manifesting superior performance through reduced Root Mean Square Error (RMSE) and enhanced coefficient of determination (R²) metrics. The model was optimised using adaptive moment estimation (Adam), culminating in a root mean square error of 3.78 bushels/acre and R² score of 0.897 on independent test data. This research substantiates the transformative potential of deep learning systems in decoding complex agro-climatic relationships and establishes a scalable, data-centric paradigm for agricultural yield projection.