Comparison of Attention Mechanism-Based Deep Learning and Transfer Strategies for Wheat Yield Estimation Using Multisource Temporal Drone Imagery

Accurate agricultural yield estimates are vital for effective food resource management, and the efficacy of yield remote-sensing models is influenced by factors such as geographic location and temporal variations, posing challenges to both the accuracy and transferability. This study used multispectral (MS), thermal infrared (TIR), red-green-blue (RGB) sensors, and LiDAR, alongside soil and precipitation data, over three locations for two years, resulting in six datasets. Three models, stacking ensemble learning (SEL), long short-term memory (LSTM), and LSTM-multi-head self-attention (LSTM-MH-SA), were compared for wheat yield estimation. The flowering stage showed the highest accuracy with SEL (<inline-formula> <tex-math notation="LaTeX">$R^{2} = 0.83$ </tex-math></inline-formula>), but the LSTM-MH-SA model outperformed SEL when integrating multitemporal features. As more data types were introduced, precision improved. Specifically, the LSTM-MH-SA model using vegetation indices, texture features (TFs), geographical information, and climate data showed superior accuracy, enhancing <inline-formula> <tex-math notation="LaTeX">$R^{2}$ </tex-math></inline-formula> by 29.58% and reducing errors compared to a basic vegetation indices model. Incorporating the joint distribution adaptation (JDA) method enabled model transfer between varying datasets, maintaining a stable accuracy (<inline-formula> <tex-math notation="LaTeX">$R^{2} = 0.81$ </tex-math></inline-formula>) with just 4% target dataset supplementation. Combining deep learning (DL) with transfer learning provides an innovative method for more efficient agricultural yield prediction.