Region-based Crop Yield Prediction using a Transformer-CNN Hybrid Model for Agricultural Forecasting

Efficient agricultural implementation and food security largely depend on precise prediction of crop yield: Traditional methods struggle to capture the diverse dependency factors, spatial and temporal, that influence yield productivity. In this paper, we tackle this issue systematically using a hybrid model where CNNs are used for spatial feature extraction and transformer networks for modeling data dependency in the sequential sense. The model works with historical yield records, weather data, soil information, and crop type which eliminates the need of image data. The CNN portion captures the spatial dependencies, while the transformer discriminates and models the long-term temporal dependencies’ seasons. A regional crop yield dataset was utilized for testing this model. The results of the tests yielded a 12% decrease in Root Mean Square Error (RMSE) along with 0.08 increases of the R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> value in comparison to the baseline models of the linear regression, CNN, and Transformer only models. This data infused spatial and temporal dynamics showcases the hybrid model’s strength in predictive accuracy enhancement. The proposed framework integrates various subsystems and serves as a base for a scalable, resource effective, and data-based solution for precision agriculture; contributing towards improved policy making, agriculture resource management, and crop production planning. This method guarantees greater applicability in various agricultural areas where satellite data could be restricted by removing the dependency on remote sensing imagery. In this study, we stress the use of hybrid deep learning algorithms in the agricultural forecasting systems to eliminate the adverse effects of climate change variability and food supply shift.