Real-time detection of plant leaf diseases based on improved YOLOv13-LM in complex field environments

Plant leaf diseases are a major factor leading to crop yield reduction, seriously threatening food security and sustainable agricultural development. Timely and accurate detection is crucial for scientific prevention and control. Traditional manual detection is time-consuming, labor-intensive, and highly subjective. Existing deep learning detection models still have significant shortcomings in complex field scenarios, struggling to balance detection accuracy, real-time performance, and stability. They are easily affected by background interference, differences in lesion scale, and sample imbalance, and some models are not lightweight enough to meet the needs of real-time field detection. Therefore, this paper uses YOLOv13 as the baseline model and constructs a YOLOv13-LM model through multi-module collaborative optimization. The optimization directions cover the backbone, neck, detection head, and loss function, strengthening lesion feature extraction and multi-scale fusion, reducing task interference, and improving localization accuracy. Model validation was completed in a complex farmland environment. The results show that the model's mAP@0.5 is improved by 5.4 percentage points to 87.9% compared to the original YOLOv13, the FPS is improved by 21.1% to 46 frames/second, and the number of parameters and computational cost are reduced by 27.3% and 25.7% respectively. The overall performance is better than the mainstream YOLO models of the same scale. However, the lightweight nature of the model is still not as good as that of the ultra-lightweight model, and its generalization and interpretability need to be improved. In the future, we will focus on ultra-lightweight design, expanding the generalization ability of multiple crops and diseases, and studying the interpretability of the model to further adapt to the actual needs of field applications.