Detection of Tea Leaf Blight From UAV Remote Sensing Images in Overcast and Low-Light Conditions

Detecting tea leaf blight (TLB) in unmanned aerial vehicle (UAV) remote sensing images is a challenging task. Under overcast and low-light conditions, existing detection methods face problems such as the small size of the TLB spots, their similarity to the backgrounds, and poor clarity, which results in low detection accuracy. To solve these problems, this study proposes a method based on a combination of an image enhancement network and an efficient object detection network. First, the image is feature-enhanced by the detection task-driven data enhancement network, called UAV detection enhancement network (UDENet), to reduce the influence of insufficient light under overcast weather condition and retain the discriminative information conducive to TLB detection. Second, TLB is detected in the enhanced UAV images by the TLB detection network UTDet. UTDet performs feature extraction on the feature-enhanced UAV remote sensing images through space-frequency multiaxis subnetwork (SFMASNet), which effectively improves the feature extraction capability for small objects in UAV remote sensing images. SFMASNet also avoids the influence of complex backgrounds on the detection results and improves the detection accuracy through the use of an aggregated contextual attention subnetwork (ACASNet). The experimental results show that compared with the existing detection methods, the proposed UDENet combined with UTDet gives the best accuracy for the detection of TLB in UAV remote sensing images under overcast and low-light conditions, with values for the precision (P), recall (R), and mean average precision (mAP) of 67.4%, 65.0%, and 68.6%, respectively. The proposed detection method achieves an mAP value that is 4.7% higher than the baseline network and more than 3.6% higher than the state-of-the-art detection methods.