An efficient detection method for tea leaf blight in UAV remote sensing images under intense lighting conditions based on MLDNet

•A fusion strategy based on RGB and SIR images is proposed to overcome the problem of image information loss caused by intense light.•A multimodal fusion (MF) module is designed based on a strategy of asymmetric feature extraction.•An SR branch is designed to reconstruct HR images at the training stage to teach the detector.•The proposed MLDNet model has a higher accuracy and a lower deployment cost compared with the baseline model. The detection of tea leaf blight (TLB) in UAV remote sensing images under intense lighting conditions is a challenging task. A large number of light spots are produced on the leaf surface by intense light, which results in information loss from the overexposed area; in addition, TLB disease spots are small and easy to overlook against a complex background. These factors lead to low accuracy of TLB detection by current methods. In this study, an efficient TLB detection method based on MLDNet is proposed for UAV remote sensing images under intense lighting conditions. Simulated infrared (SIR) images are obtained by reweighting and summing the RGB images channels after separation, and these generated SIR images can effectively supplement the information lost from the overexposed area. A multimodal fusion (MF) module is designed to fuse the RGB and SIR images. The design of the MF module is based on an asymmetric feature extraction strategy to ensure that the features of different modality images are effectively used. To solve the problem of small target detection, a super-resolution (SR) branch is designed, which uses the low- and high-level features extracted by the Backbone to reconstruct a high-resolution (HR) feature map to guide detector learning and achieve accurate detection of small TLB disease spots. Furthermore, a lightweight Backbone is designed to significantly reduce the computational cost without affecting the detection accuracy. Experimental results show that the proposed method achieves good performance. Its precision, recall, and mAP@0.5 are 78.4 %, 67.4 %, 73.4 % respectively, values that are 9.1 %, 2.7 % and 5.2 % higher than for the baseline network YOLOv8s. The parameters of the MLDNet model require only 2.6 MB, less than a third of the memory consumption of YOLOv8s.