Research on anti-shadow tree detection method based on generative adversarial network

Abstract: High-scoring remote sensing imaging has widely been applied in most management of agriculture and forestry, especially to monitor and evaluate crops and forest resources on a large scale. Nevertheless, there is a great challenge to the accuracy of single tree identification and detection during the image acquisition, due mainly to the fact that the shadow area is inevitably formed by the light. The shadow areas in the remote sensing images can be assumed as a kind of noise in the image sampling. As such, the degradation of high-resolution parameters can cause image distortion after post-processing. In this study, an anti-shadow tree detection method was proposed to detect the single tree with shadow interference using a generative adversarial network (GA-Faster RCNN). This framework consisted of a Faster RCNN network and a tree generator. The Faster RCNN network was mainly used for the tasks of feature extraction and detection. The tree generator was utilized to process the shadows in tree detection. The adversarial generation strategy was adopted by the tree generator to learn generating the minimum feature information characterizing trees. The generator was first trained separately and then put into the Faster RCNN network to finally lock its parameters. Two parts were then trained end-to-end to further improve the tree recognition ability of the network. The GA-Faster RCNN was also compared with 3 state-of-the-art methods, including region-growing, progressive cascaded convolutional neural network, and Faster RCNN on three test areas with shadows. Test area 1 presented a lot of shadows of trees, where the canopy density of trees was very high. Test area 2 showed fewer tree shadows and lower canopy closure, compared with test area 1. The shades of trees and the canopy density of trees in test area 3 were between those in test area 1 and 2. Results demonstrated that the GA-Faster RCNN achieved the highest harmonic average of precision and recall (F1) on the test area 1, 2, and 3, which were 78.4%, 91.6%, and 81.7%, respectively. The average F1 of three test areas was 84.7% for the GA-Faster RCNN, 6.2 percentage point higher than that of Faster RCNN. The user accuracy (UA) and producer accuracy (PA) of GA-Faster RCNN were also the highest among four methods, where UA was 79.8%, 95.0%, 85.3%, and PA was 77.0%, 88.5%, and 78.4% on test area 1, 2, 3, respectively. Moreover, a significance analysis, McNemar, was performed to eliminate the interference of experimental errors and other factors. It was found that there was a statistically significant difference between the three comparison methods and GA-Faster RCNN. The shadow misrecognition rate SR (proportion of the count of shadows misrecognized as trees to the count of total recognized trees) of GA-Faster RCNN was compared with that of Faster RCNN on test area 1, in order to clarify the effect of the mask on tree identification. Although the SR of GA-Faster RCNN was 13.8%, higher than that of Faster RCNN (8.6%), the UA and the number of missed tree identification were both better than those of Faster RCNN. Therefore, the GA-Faster RCNN behaved significant advantages over the other identification. In addition, the GA-Faster RCNN can still maintain the detection stability, when using different feature extraction networks, including ResNet101, ResNet50, and DenseNet. Consequently, the adversarial generative training strategy is highly suitable for learning the minimum feature information characterizing trees, while effectively reducing the interference of shadows, indicating the promising practical value for higher accuracy of tree detection.