Forest Mapping Using a VGG16-UNet++& Stacking Model Based on Google Earth Engine in the Urban Area

Abstract

Accurate, detailed urban forest mapping contributes to ecological status monitoring and formulating sustainable development policies in cities worldwide. However, accurate urban forest identification in southern Chinese cities is challenging when samples are insufficient because of high fragmentation and the influence of mountain shadows and cloudy weather. Therefore, this study combined the advantages of transfer, deep, and ensemble learning to propose a VGG16-UNet++&Stacking algorithm for urban forest mapping in heavily urbanized areas based on the Sentinel dataset. Initially, the algorithm mined deep features of an image by pre-training the convolutional layer. Then, the deep feature set was fed into ensemble learning for classification to improve accuracy and robustness. The classification results showed that the VGG16-UNet++&Stacking had an overall accuracy (OA) of 95.87% and a Kappa coefficient of 0.9481. Furthermore, the user accuracy of the forest was 97.94%. The OA of the method in this study was improved by 2.2% and 3.83% compared with that of UNet++ and random forest (RF), respectively. Compared to that of UNet++, the results showed a modest improvement in OA for the VGG16-UNet++&Stacking method; VGG16-UNet++&Stacking is more effective in eliminating cloud influence, a feature that UNet++ lacks.