Simple weakly supervised deep learning pipeline for detecting individual red-attacked trees in VHR remote sensing images
Rui Qiao, Ali Ghodsi, Honggan Wu, Yuanfei Chang, Chengbo Wang
Abstract
After an attack the by pine wood nematode, pine tree needles turn red. Using convolutional neural networks (CNNs) based object detection methods, machines can detect red-attacked trees. However, most deep learning object detection algorithms (such as Faster R-CNN and YOLO among others) often require a large number of labelled training datasets, where in each image every object must be given a bounding box label. To increase the cost-effectiveness of this process, we propose a simple yet efficient weakly supervised processing pipeline, based on class activation maps to locate the target. Unlike object detection methods that require bounding-box-labelled data for training, the proposed pipeline only needs image-level-labelled data. Using the proposed pipeline, we could achieve an average precision (AP) of 91.82% on test dataset. Comparing with sliding window-based method which achieves an average precision (AP) of 89.95%, our method not only gets a better AP but also runs faster than sliding window-based pipeline. This result not only indicates that the pipeline is a highly effective one but also demonstrates that image-level-labelled aerial images can be used for the detection of red-attacked tree. The proposed method should also find use in other object detection applications in the field of remote sensing.