TKSF-KAN: Transformer-enhanced oat yield modeling and transferability across major oat-producing regions in China using UAV multisource data

Abstract

Accurate and efficient estimation of crop yield is crucial for enhancing crop variety testing, optimizing cultivation practices, and supporting effective crop management to ensure sustainable production. However, remote sensing-based yield models often face limitations due to geographical variability and diverse cultivation techniques, affecting both their model accuracy and transferability. This study utilized multiple features, including vegetation indices (VI), color indices (CI), texture features (TF), structural indices (SI), and canopy thermal information (TIR), obtained from RGB, multispectral, and thermal infrared sensors of unmanned aerial vehicles (UAV), to create six scenarios for oat yield estimation across major oat-producing regions in China. We developed a novel deep learning-based architecture, TKSF-KAN, which combines Transformer and Kolmogorov–Arnold Network (KAN) to fuse multimodal data across key growth stages, and benchmarked its performance against stacking ensemble learning (SEL) and standalone Transformer model. While SEL demonstrated the highest accuracy in single-modal scenarios, TKSF-KAN outperformed SEL in multimodal settings (R² = 0.76–0.81). Particularly, TKSF-KAN, with integrated VI, CI, TF, SI, and TIR inputs improved R² by 53.77 % compared with single-modal data sources. By combining Adaptive Batch Normalization (AdaBN) with fine-tuning strategies, transfer performance improved as the proportion of fine-tuned data increased, reaching a peak R² of 0.83 at one study site. In contrast, transferability was more influenced by cultivation practices at another site, with a maximum R² of 0.78. This study presents an innovative framework that seamlessly integrates agricultural practices with remote sensing and transfer learning methodologies, offering a more robust and scalable solution for yield prediction and advancing precision of agricultural management.