TY  - JOUR
AU  - Cao, Qingjin
AU  - Xuan, Chuanzhong
AU  - Tang, Zhaohui
AU  - Zhang, Tao
AU  - Song, Yaobang
PY  - 2026
TI  - Desert steppe vegetation community classification using a deep learning method fusing attention and multi-scale features
JO  - Transactions of the Chinese Society of Agricultural Engineering
SN  - 1002-6819
SP  - 268
EP  - 276
VL  - 42
IS  - 13
AB  - Accurate classification of desert steppe vegetation communities is often required to prevent desertification of fragile grassland in sustainable agriculture. Conventional botanical field surveys rely on manual sampling, which is labor-intensive, time-consuming, and difficult to deploy over vast geographic landscapes. Meanwhile, satellite remote sensing techniques frequently struggle to delineate the fine-grained spatial patterns and complex morphologies of individual vegetation patches, due to restricted spatial resolutions. Unmanned Aerial Vehicle (UAV) can be expected to acquire ultra-high-resolution optical imagery, effectively capturing the phenotypic characteristics, resource distributions, and spectral signatures of sparse desert vegetation. Nevertheless, the complex environment of desert steppes—such as severe background spectral interference from exposed soils, various illumination conditions, and multi-scale phenotypes among different plant species—has posed great challenges for classification algorithms. This study aims to classify desert steppe vegetation communities using multi-scale features and fusing attention mechanisms in deep learning. An advanced deep learning framework was proposed (termed the UNetAC model) for the precise semantic classification of desert steppe vegetation communities using UAV-derived imagery. UNetAC architecture was constructed to optimize the classical U-Net fully convolutional baseline network. Convolutional Block Attention Module (CBAM) and Atrous Spatial Pyramid Pooling (ASPP) module were selected for UNetAC model. The dual-domain CBAM was embedded into the feature extraction encoder to recalibrate semantic representations over channel and spatial dimensions. Channel and spatial attention weights were applied to suppress visually distracting background noise, such as barren earth and shadows. While the features of the target vegetation species were amplified selectively. Furthermore, the ASPP module was introduced at the bottleneck layer to capture rich multi-scale contextual information without degrading the spatial resolution of intermediate feature maps. A significantly expanded receptive field was acquired to implement multiple parallel Atrous convolutional branches with progressively varying dilation rates. Multi-scale feature aggregation was utilized to identify vegetation patches with extreme morphological size variations, particularly for complex ecological community structures. Experiments were conducted on an annotated UAV imagery dataset from a representative desert steppe region. The UNetAC model was trained to focus exclusively on the grazing-forbidden zone. The results demonstrated that the exceptional classification performance of the optimal network was achieved during evaluation, with a mean Intersection over Union (mIoU) score of 83.28%, an Overall Accuracy (OA) of 91.32%, and a mean Pixel Accuracy (mPA) of 90.53%. These metrics outperformed the rest in semantic segmentation using deep learning. Specifically, the fusing attention mechanism significantly reduced misclassification rates in the fuzzy boundary transition zones separating vegetation categories from the background soil matrix. Concurrently, the multi-scale pooling strategy was achieved in the morphological structural integrity and boundary clarity of the vegetation patches after extraction. A zero-shot spatial transfer experiment was conducted to evaluate the structural robustness and spatial cross-regional generalization of the algorithm after architectural modifications. The optimal UNetAC model was directly applied—without any supplementary fine-tuning—to classify heterogeneous vegetation communities in a light-grazing area. This secondary testing environment was characterized by different spatial patterns of vegetation distribution, community densities, and background soil conditions. UNetAC model successfully maintained consistent efficacy of classification under various tests. An overall accuracy (OA) of 90.42% was obtained under an ecological environment. Cross-regional generalization confirmed that the hierarchical feature representations were robust, highly discriminative, and invariant to environmental shifts. In conclusion, the UNetAC model can be expected to deliver an effective, accurate, and scalable solution for the classification and assessment of desert steppe vegetation communities. An analytical benchmark has been established to integrate the spatial-channel attention mechanisms and multi-scale contextual feature aggregation using complex UAV optical imagery. Deep learning models can be deployed to dynamically monitor the large-scale grassland and vegetation evolution in sustainable pastoral strategies.
UR  - https://doi.org/10.11975/j.issn.1002-6819.202512051
DO  - 10.11975/j.issn.1002-6819.202512051