A Hybrid Transformer Framework for Crop Yield Forecasting Using Multispectral and Climatic Time-Series Data

In modern times, prediction of crop harvest is vital task in farming, management of food supply as well as early warning systems, by enabling farmers and policymakers to make knowledgeable decisions on resource allocation, risk mitigation and trade. However, the existing 3D Convolutional Neural Network with Convolution Long Short-Term Memory (3D-CNN-ConvLSTM) faced challenges such as limited spatial detail, weak temporal modeling and climatic context. To address these limitations, a hybrid Deep Learning (DL) model integrating Swin 3D Transformer with Temporal Fusion Transformer (YieldSwin-TFT) is proposed. Initially, the satellite imagery data and climate data are collected from the Moderate Resolution Imaging Spectroradiometer (MODIS) and CropHarvest respectively. Next, this collected data is preprocessed using various techniques such as spatial clipping, patch extraction, temporal alignment and normalization. Then, this preprocessed data is fed to the Swin transformer, which captures Spatio-spectral patterns. Further, these captured patterns are forwarded to the TFT module, which captures temporal dependencies and static features. Finally, Multi Head Self-Attention (MHSA) is employed which enhances these features and predict annual crop yield per region. Hence, the proposed YieldSwin-TFT achieved greater results in Mean Absolute Error (MAE) with 4.10, Root Mean Squared Error (RMSE) with 5.20, Mean Absolute Percentage Error (MAPE) with 9.85 % and Mean Squared Logarithmic Error (MSLE) with <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{0. 0 1 7 6}$</tex>, demonstrating better accuracy and generalization.