A STUDY OF USING FULLY CONVOLUTIONAL NETWORK FOR TREETOP DETECTION ON REMOTE SENSING DATA
Individual tree detection and counting are critical for the forest inventory management. In almost all of these methods that based on remote sensing data, the treetop detection is the most important and essential part. However, due to the diversities of the tree attributes, such as crown size and branch distribution, it is hard to find a universal treetop detector and most of the current detectors need to be carefully designed based on the heuristic or prior knowledge. Hence, to find an efficient and versatile detector, we apply deep neural network to extract and learn the high-level semantic treetop features. In contrast to using manually labelled training data, we innovatively train the network with the pseudo ones that come from the result of the conventional non-supervised treetop detectors which may be not robust in different scenarios. In this study, we use multi-view high-resolution satellite imagery derived DSM (Digital Surface Model) and multispectral orthophoto as data and apply the top-hat by reconstruction (THR) operation to find treetops as the pseudo labels. The FCN (fully convolutional network) is adopted as a pixel-level classification network to segment the input image into treetops and non-treetops pixels. Our experiments show that the FCN based treetop detector is able to achieve a detection accuracy of 99.7 % at the prairie area and 66.3 % at the complicated town area which shows better performance than THR in the various scenarios. This study demonstrates that without manual labels, the FCN treetop detector can be trained by the pseudo labels that generated using the non-supervised detector and achieve better and robust results in different scenarios.