DEEP TRANSFER LEARNING FOR PREDICTING CROP YIELDS UNDER CROP ROTATION AND MULTI-CROP CONDITIONS

Abstract

Crop production plays a fundamental role in sustaining global food systems and maintaining ecological balance, especially in regions where crop rotation and multi-crop cultivation are essential for soil health and yield optimization. However, the variability of environmental factors and limited availability of high-resolution data make it challenging to develop robust and scalable yield prediction models. This study introduces a novel deep transfer learning framework that integrates convolutional neural networks (CNNs) with spatio-temporal satellite data to predict crop yields under diverse crop rotation systems. The model leverages pretrained convolutional features from data-rich crop domains and fine-tunes them using region-specific datasets from Uzbekistan. By transferring learned representations across multiple crop types—such as wheat, maize, barley, and rice—the approach significantly enhances generalization and reduces the computational cost of training individual models. Experimental results demonstrated that the proposed method achieved an average coefficient of determination (R² = 0.90) and a root mean squared error (RMSE = 0.54 t ha⁻¹) across all crop types. Compared to traditional regression models and crop-specific CNNs, the transfer learning model achieved up to 18% faster convergence and superior robustness under data-scarce conditions. The findings indicate that transfer learning can effectively support crop rotation planning and agricultural forecasting by enabling high-efficiency, multi-crop yield prediction systems adaptable to regional variability..