A robust detection method of apple leaf spectral image Starscream based on CRS-YOLO algorithm

Due to the significant variation in red spider mite infestations across different growth stages and environmental conditions, traditional methods struggle to achieve precise and efficient detection. To address this, a robust detection method based on the CRS-YOLO algorithm was designed for apple leaf spectral images. This approach enhances red spider mite leaf spot images by combining CS-LBP features with chi-square test-estimated thresholds. The Visual Transformer (ViT) was introduced as a feature extractor, preserving spatial information by segmenting small image patches and incorporating positional embeddings. Integrating Principal Component Analysis (PCA) transformation with spectral angle classification enabled precise extraction of spider mite-infested leaf areas from healthy regions. Building upon the YOLOv5s model, the original SIoU loss function was replaced with an improved Alpha-IoU loss function. Zoom loss is introduced to address the positive-negative sample imbalance in IACS regression. Experimental results demonstrate: map coverage reaches 84.63%, detection speed achieves 44.07 frames per second, high detection accuracy, and the ability to identify all spider mite lesions in apple leaf spectral images.