HPA-UNet-LSNet: An LSNet-based U-Net with hybrid pooling attention for accurate segmentation of Haloxylon ammodendron crowns from UAV RGB imagery

Abstract

Accurate segmentation of Haloxylon ammodendron crowns from UAV RGB imagery remains challenging in desert environments because of sparse crown distribution, weak crown–background contrast, and interference from sandy soil and co-occurring shrubs. To address this problem, this study developed HPA-UNet-LSNet , an enhanced U-Net framework that replaces the original encoder with LSNet and introduces hybrid pooling attention (HPA) for feature fusion. On the independent test set, HPA-UNet-LSNet achieved a Precision of 0.8890, a Recall of 0.9198, an F1-score of 0.9041, and an mIoU of 0.8456. Compared with the baseline U-Net, it reduced false positives from 454 ± 53 to 267 ± 18 and false negatives from 224 ± 11 to 185 ± 10. The improvement was especially evident for small crowns, where the F1-score increased from 0.7318 ± 0.0179 to 0.7611 ± 0.0102, and the mIoU increased from 0.6498 ± 0.0045 to 0.6929 ± 0.0089. Grad-CAM results further showed more concentrated responses over crown regions and relatively reduced activation in irrelevant background areas. Overall, HPA-UNet-LSNet provides an effective and practical RGB-based solution for Haloxylon ammodendron crown segmentation in desert environments.