基于孪生残差神经网络的GF-2影像林地变化检测——以浙江省建德林场为例

Forest Change Detection based on Siamese Neural Network with GF-2 Image:A Case of Jiande Forest Farm,Zhejiang

森林是世界上生产力最高的可再生自然资源之一,但由于火灾、洪水、砍伐等多种自然或人为因素干扰,森林的生态环境受到严重威胁。准确掌握林地资源变化的情况,可以为森林资源的管理与保护提供有效信息。由于林地类别及树种差异较大,在林地变化检测任务中传统的机器学习变化检测方法难以捕捉深层次语义信息,存在提取特征适应性差、识别能力弱以及因季相导致的伪变化等问题。提出以孪生残差神经网络构建深度学习模型,进行林地变化的检测实验。分别采用残差神经网络ResNet50、添加不同轻量级注意力机制如卷积注意力机制模块CBAM和压缩和激励模块SE 3种不同特征提取方法作为主干特征提取模块。3种主干特征提取网络都基于预训练权重进行训练,通过将提取的多尺度的特征图进行融合,使得不同特征图中信息粗略细节和精细细节互补,从而改善变化检测效果,同时具有相同数量的参数,共享权值的优点。以浙江省建德林场为实验区,获取2015年和2020年两期高分二号卫星影像,构建一套分辨率为1 m的林地变化检测数据集。对孪生残差神经网络变化检测的结果和真实变化标签进行比较,其中主干特征提取网络SE-ResNet50综合结果最好,精确率为0.91,召回率为0.78,F1分数为0.83,要优于主流的变化检测模型FC-Siam-conc、FC-Siam-diff。证明孪生残差神经网络在高分辨率遥感影像林地变化检测任务中能够准确捕捉林地变化,为森林资源管理部门提供了新的林地变化检测方法。

Forest is a valuable non-renewable resource,but the ecological environment of forest is seriously threatened by many natural or man-made factors such as fire,flood,and deforestation interference.Accurate grasp of forest resource changes can provide effective information for forest resource management and protec-tion.In the task of forest change detection,traditional machine learning change detection methods have difficul-ty in capturing deep semantic information due to large differences in forest categories and tree species,and suffer from poor adaptability of extracted features,weak recognition ability,and pseudo-change due to seasonal phas-es.We propose to build a deep learning model with Siamese neural networks for forest change detection experi-ments.Three different feature extraction methods,ResNet50(Residual neural network),CBAM(Convolu-tional Block Attention Module)and SE(Squeeze and Excitation)with different lightweight attention mecha-nisms are used as backbone feature extraction modules,respectively.All three backbone feature extraction net-works are trained based on pre-trained weights,which improve change detection by fusing the extracted multiscale feature maps so that the coarse and fine details of information in different feature maps complement each other.It also has the advantage of sharing weights with the same number of parameters.Taking Jiande Forest Farm in Zhejiang province as the experimental area,two phases of GF-2 images in 2015 and 2020 are acquired to construct a forest change detection dataset with a resolution of 1m.The results of Siamese neural network change detection are compared with the true change labels(Ground truth),where the backbone feature extrac-tion network SE-ResNet50 has the best combined results with Precision(0.91),Recall(0.78)and F1-score(0.83),which is better than mainstream change detection models FC-Siam-conc,FC-Siam-diff.It is proved that Siamese neural networks can accurately capture forest changes in the task of forest lad change detection from high-resolution remote sensing images,and provide a new forest change detection method for forest re-source management departments.