TAM-Net: A deep network combining tabular diffusion algorithm, attention mechanism, and multi-task learning for monitoring crop water status from UAV multi-source images

• Published in: European journal of agronomy Vol. 170; p. 127778
• Main Authors: Cheng, Zhikai, Gu, Xiaobo, Zhang, Zhengtao, Xu, Yang, Zhao, Tongtong, Li, Yupeng, Sun, Shikun, Du, Yadan, Cai, Huanjie
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.09.2025
• Subjects: Data synthesis; Data-driven model; Deep learning; Maize; Remote sensing; Deep learning; Maize; Data-driven model; Data synthesis; Remote sensing
• ISSN: 1161-0301
• DOI: 10.1016/j.eja.2025.127778

Rapid and accurate monitoring of crop water status is essential for ensuring sustainable agricultural development and food security. Crops exhibit a complex set of growth and physiological responses under water deficit. Existing studies primarily focused on the monitoring of phenotypic parameters, while the physiological indicators highly relevant to crop water status were ignored. In this context, we aimed to develop a novel model to comprehensively and accurately monitor maize water status using multi-source UAV data and multiple growth and physiological indicators. We first composed the original dataset, including feature variables based on multi-source UAV data (spectral indices, texture indices, thermal indices, and structural indices) and prediction variables based on field measurements (equivalent water thickness, stomatal conductance, transpiration rate, and actual photochemical efficiency) in 2023 and 2024. Next, the tabular denoising diffusion probabilistic model (TabDDPM) was employed for synthesizing new samples to adequately train the models. Then, a deep learning network named TAM-Net, with the hybrid attention mechanism and multi-task learning, was trained on the synthetic dataset. Finally, the fuzzy comprehensive water index (FCWI) considering uncertainty and variability was obtained in 2023–2024. The results indicated that multi-source data significantly improved the model performance, with the R2 of 0.52–0.63, and NRMSE of 27.89 %–29.86 %. TabDDPM was able to synthesize new datasets with high similarity and effectiveness. TAM-Net achieved the highest monitoring accuracy for the four indicators of crop water status (R2 of 0.76–0.90, NRMSE of 12.92 %–22.95 %). FCWI effectively assessed the water status across different treatments. Overall, TAM-Net was demonstrated with powerful performance for monitoring maize water status, which has potential in supporting precision irrigation practices. [Display omitted] •Multi-source data showed advantages in EWT, Gs, Tr and ΦPSⅡ monitoring.•Data synthesis algorithm (TabDDPM) produce high-quality samples.•Attention mechanisms and multi-task learning improved model performance.•FCWI effectively assessed the crop water status across different treatments.