Potato production is highly sensitive to water deficit, yet UAV thermal infrared monitoring still suffers from inaccurate canopy temperature extraction and limited soil moisture inversion accuracy. This study collected UAV thermal infrared imagery and ground measurements in Zhangbei County, Hebei Province, China, covering the seedling, tuber initiation, and tuber bulking stages of two potato cultivars. To mitigate soil-background interference, we propose HA-ResUNet++ (Hierarchical Attention ResNet34-UNet++) for canopy segmentation. Built on the densely connected skip pathways of UNet++ with a ResNet34 encoder, HA-ResUNet++ replaces the initial 7 × 7 convolution with DCNv4 deformable convolution, embeds CBAM in shallow layers to enhance locally salient canopy features, and introduces CoT Attention in deeper layers to capture long-range structural dependencies. A four-channel input combining pseudo-color thermal images and the temperature matrix is adopted to fuse spatial-structural and thermal information. Compared with UNet++, HA-ResUNet++ improves IoU, mIoU, PA, and F1-score by 9.58%, 7.31%, 3.98%, and 6.09%, respectively, and outperforms several classical segmentation models. Based on the segmented canopy, four canopy temperature determination strategies (Tpc, Tmc, Tmax, and Tcf50%) were evaluated for Crop Water Stress Index (CWSI) construction and phenology-specific soil moisture inversion. The optimal strategies were stage-dependent, with Tmc most robust in the seedling stage, Tmc/Tpc more suitable during tuber initiation, and Tcf50%/Tmax performing best during tuber bulking. Using a second-order polynomial model with growth-stage-specific optimal indicators, soil moisture inversion achieved R² = 0.52–0.66 across the three stages and two cultivars. These results indicate that robust canopy segmentation combined with phenology-adaptive temperature determination improves the stability and applicability of UAV thermal infrared–based soil moisture inversion, providing a rapid and non-destructive approach for precision irrigation in potato fields.
Liu et al. (Sun,) studied this question.