A lightweight multitask adaptive fusion network with channel attention for UAV-based crop detection and classification

Unmanned aerial vehicles (UAVs) have become essential tools in precision agriculture, facilitating large-scale monitoring of crop health, weed proliferation, and overall field conditions. However, conventional single-task recognition models encounter significant challenges when implemented on resource-constrained UAV platforms, particularly in detecting densely distributed small objects, managing scale variations, and addressing occlusions. To address these challenges, this paper proposes a channel attention-based adaptive fusion network (CNAF-Net) specifically designed for agricultural scenarios. CNAF-Net utilizes MobileNetV3 for hierarchical feature extraction and integrates a channel attention-guided adaptive fusion mechanism that dynamically combines multiscale features across detection, classification, and segmentation tasks. Experiments conducted on the WeedMap dataset demonstrate that CNAF Net consistently surpasses state-of-the-art lightweight models, including YOLOv5-s and Lite-FPN-SSD, in terms of precision, recall, F1-score, and mean average precision. In addition to quantitative analyses, visualization experiments further confirm the model’s applicability to agricultural scenarios. Specifically, CNAF-Net generates canopy coverage and leaf area index heatmaps for growth monitoring and produces crop–weed–soil segmentation masks that enable fine-grained analyses of vegetation density and interspecies competition. Ablation studies highlight the contribution of the adaptive fusion module to enhanced feature selectivity, convergence, and robustness. Overall, CNAF-Net offers an efficient and scalable solution for UAV-based agricultural monitoring. Its compact architecture and robust generalization capability enable real-time deployment on edge devices, providing practical support for weed detection, crop growth assessment, and sustainable precision agriculture.