Topographic constrained land cover classification in mountain areas using fully convolutional network.

Mountains are an important kind of landform on the earth's surface. Due to harsh mountainous environment, such as steep slopes and cliffs, remote sensing has become an indispensable tool for surveying mountain areas instead of traditional ground surveys. However, the accuracy of current land cover products derived from remote sensing in mountain areas is still low. In this paper, we propose a three-level architecture for land cover classification in mountain areas. Topographic partitioning is first performed in order to partition a large area into several smaller zones, and then, multiresolution segmentation is implemented in each individual zone. Thus, we can obtain initial geo-semantic objects with terrain, spectrum and texture homogeneities. A fully convolutional network (FCN)-based classifier (U-Net) is further introduced for supervised classification of land cover. From the perspectives of both visual interpretation and quantitative evaluation, the proposed method achieved robust and high-precision results for all land cover types. We also investigate the contributions of multimodal features for classification accuracy improvement. First, the results showed that additional features resulted in higher classification accuracies than 3-features only; 6-features achieved the best performance on farmland, impervious surfaces and coniferous forests, while 5-features performed well on water and broad-leaved forests. The elevation feature did not have a positive effect on water and broad-leaved forests, which can be explained by their physical distribution in the landscape. Second, the most significant improvement was achieved on water (Kappa coefficient increased from 0.741 to 0.924), followed by coniferous forests (Kappa coefficient increased from 0.629 to 0.805), whereas only a minor improvement was observed for the other three types. Furthermore, the accuracies of farmland and impervious surfaces remained relatively high even without the assistance of additional features, and texture feature plays a key role. The final land cover map was generated by combining the optimal results of each type via a hierarchical integrating strategy. The overall accuracy of classification achieved 90.6%. [ABSTRACT FROM AUTHOR]