LMMA-YOLO: Lightweight Multi-scale Multi Attention Architecture for Comprehensive Cashew Tree Stress Detection

Anacardium occidentale L. (cashew) plays a major economic role in Côte d'Ivoire, but its production is heavily affected by diseases, pest insects, and abiotic stress. This work presents LMMA-YOLO (Lightweight Multi-Scale Multi-Attention YOLO), an optimized object detection model for the automatic identification of these stresses from aerial images. The proposed architecture combines a lightweight MobileNetV4 backbone, a hierarchical multi-scale fusion (PAFPN), and three complementary attention mechanisms: Coordinate Attention (CA), Context Gating (CG), and Multi-Head Attention (MHA). The model was trained on the CADI AI dataset (4,736 images, three classes) with a composite loss function incorporating SIoU, Focal Loss, and Binary Cross-Entropy. Experimental results show that LMMA-YOLO achieves a mAP@50 of 97.6% with an inference latency of 5.1 ms per image on GPU and 28.4 ms on Jetson Orin Nano, surpassing YOLOv8 models by 2.5 percentage points while maintaining real-time processing capability at approximately 35 FPS on embedded platforms.