Vision Transformer-Based Temporal Attention Recalibrated Network for Efficient Crop Yield Prediction

To ensure food security, optimize resource allocation and support sustainable agricultural practices, accurate crop yield estimation plays a critical role. However, the existing approaches failed to capture global spatial and temporal growth patterns. Additionally, these approaches suffered from limited global representation and high computational complexity, which resulted in limited generalization and scalability. To address these limitations, a Vision Transformer with Temporal Attention Recalibrated Network (ViT-TARN) is proposed for accurate crop yield estimation for precise farming. Initially, a sequence of multi spectral field images is collected from soybean crop yield. In preprocessing stage, the collected crop images are normalized to scale the pixel values to a fixed range. Next, these preprocessed images are divided into overlapping patches and these patched are flattened. Followed by, spectral indices such as Normalized Difference Vegetation Index (NDVI) and Enhanced Vegetation Index (EVI) are utilized to enrich feature representation. These feature representations are transformed into feature embeddings and fed to the Vision Transformer (ViT) to capture global spatial dependencies. In addition, token wise feature recalibration is performed to focus on most relevant regions. Moreover, temporal self-attention with gradient scaling is applied to dynamically assign importance. Finally, the fused embeddings are passed to the regression head which predicts the continuous crop yield values. Experimental results demonstrate that the proposed ViT-TARN achieved Root Mean Squared Error (RMSE) of 8.1 and Mean Absolute Error (MAE) of 6.5, outperforming the existing approaches.