Multi-Scale Feature Aggregation Network for Semantic Segmentation of Land Cover

Land cover semantic segmentation is an important technique in land. It very practical land resource protection planning, geographical classification, surveying and mapping analysis. Deep learning shows excellent performance picture recent years, but there are few algorithms for cover. When dealing with tasks, traditional networks often have disadvantages such as low precision weak generalization due to the loss of image detail information limitation weight distribution. In order achieve high-precision segmentation, this article develops a multi-scale feature aggregation network. Traditional convolutional neural network downsampling procedure has problems resolution degradation; fix these problems, extraction spatial pyramid module made assemble regional context data from different areas. address issue incomplete at multiple sizes, fusion developed fuse attributes various layers several sizes boost accuracy. Finally, attention presented enhance segmentation’s target that classic capacity building waters segmentation. Through contrast experiment experiment, it can be clearly demonstrated algorithm proposed paper realizes high