Semantic Segmentation on Remote Sensing Images with Multi-Scale Feature Fusion

Semantic segmentation of remote sensing images has drawn extensive attention both from academy and industry for its wide range of applications,such as urban planning,urban change detection and geographic information system.Nevertheless,many complicated factors,such as complex background,shadows,objects with various scales,topological shapes and appearances in different regions,make this task quite challenging.Accordingly,this paper proposes a deep convolutional neural network model with multi-scale information fusion for semantic segmentation of remote sensing images.The structure of our model is composed of two parts:encoder and decoder.In the encoder part,a strategy is proposed to fuse multi-scale features based on DenseNet network.Specifically,global average pooling is first used to extract regional semantic information of different sub-regions to make network understand complex background in remote sensing images;sub-region global average pooling and multiscale convolution are then used to deal with complex background areas.In the decoder part,we design a shorter decoder which can fuse features from different levels of convolution to accurately restore the image details.For the overall model construction,we design a hierarchical monitoring mechanism with multiple outputs.This trick allows our model to obtain supervised information from different levels,which can help guide the training of the network.Extensive experiments on ISPRS benchmark datasets and Beijing remote sensing dataset demonstrate the effectiveness of our approach.