HPA-UNet-LSNet: An LSNet-BasedU-Net with Hybrid Pooling Attention for Accurate Segmentation of Haloxylon ammodendron Crowns fromUAV RGB Imagery

HPA-UNet-LSNet dataset for Haloxylon ammodendron crown segmentation from UAV RGB imagery This dataset was developed for crown segmentation of Haloxylon ammodendron from high-resolution UAV RGB imagery in desert environments. The study area is located in the Junggar Basin, Xinjiang, China, where Haloxylon ammodendron is sparsely distributed and often mixed with sandy background and co-occurring shrubs. These characteristics make crown extraction challenging because of weak crown–background contrast, irregular crown shapes, and the presence of small targets. The original imagery was acquired in July 2023 using a DJI Matrice 300 RTK (M300) UAV equipped with a Zenmuse L1 sensor. The collected RGB images were processed in DJI Terra to generate an orthomosaic. The orthomosaic was then cropped into image patches of 640 × 640 pixels for manual annotation and model development. The dataset contains 1600 RGB image patches and their corresponding binary segmentation labels. All samples were divided into training, validation, and test subsets at a ratio of 70%, 15%, and 15%, respectively. To reduce spatial leakage, the subsets were assigned from different flight lines, ensuring that adjacent patches did not appear in different subsets. Each label mask represents a binary crown segmentation task: background: 0 Haloxylon ammodendron crown: 1 At the pixel level, the dataset is imbalanced, which reflects the real distribution of sparse desert vegetation: foreground crown pixels: 4.71% background pixels: 95.29% background-to-foreground ratio: approximately 20.25:1 This dataset was constructed to support research on UAV-based vegetation segmentation, desert shrub monitoring, and lightweight deep-learning models for complex arid environments. Dataset contents The dataset includes: RGB image patches binary segmentation masks train/validation/test split files A typical directory structure is as follows: VOCdevkit/└── VOC2007/├── JPEGImages/├── SegmentationClass/└── ImageSets/└── Segmentation/├── train.txt├── val.txt└── test.txt Image format Image type: RGB Patch size: 640 × 640 pixels Label type: single-channel binary mask Label values: 0 = background 1 = Haloxylon ammodendron crown Recommended use This dataset can be used for: semantic segmentation of desert shrub crowns benchmark comparison of segmentation models ablation studies on lightweight encoders and attention modules evaluation of small-target detection under complex desert backgrounds Notes The dataset was designed for research purposes. Users should cite the associated paper when using this dataset. If redistributed or reused, please retain the original attribution information.