Rangeland ecosystems have been sources for pastoral communities. However, traditional seasonal mobility patterns are disrupted by climate change, requiring more dynamic, data-driven plant-based rangeland assessment. In this study, we propose amultiscale transformer-based network to address the challenge of automatically classifying rangeland plant species for livestock pasture scoring in Africa, given the complex environments and limited data. Accurately distinguishing similar plants with varying livestock utility is important for sustainable management. This study investigated Vision Transformers, known for multiscale features important for fine-grained visual differentiating. The initial comparative analysis of ViT, DEiT, and Swin Transformer models demonstrated the promise of Swin architecture. Building on this, we introduce a Multiscale Swin Transformer model incorporating multiscale feature fusion and weighting mechanism to enhance plant image classification. The model combines global and fine feature extraction, followed by fusion module. Early features capture local patterns (e.g., leaf), and later layers capture semantic information (e.g., general morphology). The proposed Multiscale approach utilizing a weighted decoder provides better performance improvements over the Swin base model, achieving 89.71 % accuracy compared to 88.0%, demonstrating that fusing features at different scales leads to better recognition. Moreover, analysis of the collected data shows class imbalance, including dominance of invasive species and useful herbs, sparse representation of rare unuseful (e.g., poisonous) and other sparse useful livestock forage species. This highlights an essential need for systematic data collection and optimization strategies, like synthetic image generation, to mitigate limitations and improve model generalization.
Alizadehsani et al. (Thu,) studied this question.