摘要
Abstract
Although convolutional neural networks have achieved great success in the field of change detection,the detection effect of change regions with different shapes and scales is obviously different,the feature information contained in the small target region tends to decrease significantly with the increase of network depth.To address this issue,we proposed a network to enhance the feature information of small areas(ESANet).First,it uses two-dimensional Gaussian fitting to generate small-area feature information.Second,the enhancement of the feature in-formation is fused with the decoding layer features,the mixed loss function is used to calculate the loss of each lay-er and the loss is weighted.At last,the binary classification of predicted values is got through threshold settings.The method is evaluated on datasets such as LEVIR-CD,CDD,and SYSU.The experimental results show that the network enhancing the feature information of small regions can significantly enhance the detection effect of small ar-eas without changing the original detection performance.The accuracy rates on Levi-CD,CDD and SYSU data sets reached 93.24％,97.17％and 84.89％,respectively.关键词
变化检测/遥感影像/小区域特征增强/卷积神经网络/多尺度融合Key words
change detection/remote sensing imagery/enhancement of small area features/convolutional neural networks/multi-scale fusion分类
天文与地球科学
