Progressive Cascaded Convolutional Neural Networks for Single Tree Detection with Google Earth Imagery

Abstract

High-resolution remote sensing images can not only help forestry administrative departments achieve high-precision forest resource surveys, wood yield estimations and forest mapping but also provide decision-making support for urban greening projects. Many scholars have studied ways to detect single trees from remote sensing images and proposed many detection methods. However, the existing single tree detection methods have many errors of commission and omission in complex scenes, close values on the digital data of the image for background and trees, unclear canopy contour and abnormal shape caused by illumination shadows. To solve these problems, this paper presents progressive cascaded convolutional neural networks for single tree detection with Google Earth imagery and adopts three progressive classification branches to train and detect tree samples with different classification difficulties. In this method, the feature extraction modules of three CNN networks are progressively cascaded, and the network layer in the branches determined whether to filter the samples and feed back to the feature extraction module to improve the precision of single tree detection. In addition, the mechanism of two-phase training is used to improve the efficiency of model training. To verify the validity and practicability of our method, three forest plots located in Hangzhou City, China, Phang Nga Province, Thailand and Florida, USA were selected as test areas, and the tree detection results of different methods, including the region-growing, template-matching, convolutional neural network and our progressive cascaded convolutional neural network, are presented. The results indicate that our method has the best detection performance. Our method not only has higher precision and recall but also has good robustness to forest scenes with different complexity levels. The F1 measure analysis in the three plots was 81.0%, which is improved by 14.5%, 18.9% and 5.0%, respectively, compared with other existing methods.