基于改进U-Net网络与联合损失函数的海南自然保护区高分辨率遥感变化检测模型

海南自然保护区遥感监测对森林资源监测、生态保护及热带亚热带地表研究具有重要意义。基于深度学习方法,针对海南省自然保护区大范围变化检测问题,对U-Net网络结构进行了改进,在每一个卷积层后加入标准化层,以跳线连接的形式将原有卷积模块改进为优化模块,同时在编码器底端添加金字塔池化模块以更好提取全局信息,形成了改进U-Net网络模型。模型训练采用基于交叉熵损失函数和广义骰子损失函数构建的联合损失函数,配合多种优化策略实现端到端的地物变化信息提取。该模型应用于公开数据集和研究构建的海南自然保护区数据集的变化检测任务,总体精度分别为97.21%(Kappa系数0.88)和95.12%(Kappa系数0.90),相比原始U-Net效果提升显著。

Remote Sensing Change Detection Model of Hainan Nature Reserves Based on Improved U-Net and Joint Loss Function

Remote sensing monitoring of Hainan nature reserves is of great significance to forest resource monitoring, ecological protection and tropical and subtropical surface researches.This study focuses on the large-scale change detection in Hainan nature reserves with deep learning methods.To improve the U-Net structure, a batch normalization layer was added after each convolution layer, the original convolution modules in encoder and decoder were transformed into different improved modules with short connections and a pyramid pooling module was attached to the end of the encoder, which was helpful to extract global information.The weighted addition of cross-entropy loss and generalized dice loss was used to form a joint loss function.A good end-to-end change detection results were achieved by adjusting parameters such as learning rate and learning rate decay.The model was applied to the change detection task of the public dataset and the Hainan nature reserves dataset.The overall accuracy was 97.21% (Kappa coefficient 0.88) and 95.12% (Kappa coefficient 0.90), which was significantly improved compared to the original U-Net.