| Landslides,as a natural disaster that occurs frequently and poses a serious threat to people’s lives and property,it is important to accurately obtain their location and distribution range for disaster mitigation and relief work.However,most of the current landslide monitoring studies focus on the same resolution remote sensing images,and there are fewer studies on landslide extraction on different resolution images,making it difficult to objectively evaluate the relocatability of the model.To address this issue,this study used landslide data from the Iburi area after the 6 September 2018 Hokkaido earthquake in Japan(Planet imagery,spatial resolution 3m),landslide data after the 8August 2017 Jiuzhaigou earthquake(GF1 imagery,spatial resolution 2m)and open landslide data from Bijie(Triple Sat imagery,spatial resolution 0.8m)as data sources A cross-regional study of landslide extraction using different machine learning methods was carried out.Details are as follows:(1)To address the problem that the landslide extraction model is easily disturbed by impurities such as light,shading and clouds,the image enhancement strategy based on a combination of bilateral filtering and histogram equalization and pre-processing methods such as threshold segmentation are used to reduce the difficulty of landslide extraction under complex background feature conditions.The experimental results show that the combination of bilateral filtering and histogram equalization can effectively smooth out the noise and improve the spectral difference between the landslide and the background features in the image,while retaining the edge contour detail information.(2)To address the problem of difficult landslide feature extraction under complex background features,spatial and frequency domain features are designed based on expert knowledge.The spatial domain features are constructed using image spectral,texture and edge information.The FT saliency detection algorithm is used to construct the frequency domain salient features.Based on the constructed spatial and frequency domain features,random forest and support vector machine models are constructed for landslide extraction respectively.Through training and validation on different locations and different resolution images,the results show that the constructed spatial and frequency domain features have significant advantages in constructing landslide extraction models.(3)As the manually constructed features are limited by expert knowledge,the robustness of application in different study areas is difficult to guarantee.In this paper,a machine learning classification model based on automatic feature extraction with full convolutional self-encoders is designed.Firstly,the full convolutional self-encoder is used to automatically learn the deep features of the data,and a random forest and support vector machine model is constructed for landslide extraction based on the learned features.By comparing with deep learning algorithms on different study areas and different spatial resolution images,it is well demonstrated that the machine learning models constructed based on the deep features learned by the full convolutional self-encoder have higher portability and accuracy.In summary,the results of this paper show that the combination of image enhancement and thresholding pre-processing methods can remove cloud interference and can improve the accuracy of landslide extraction models.A machine learning model using a combination of spatial and frequency domain features can significantly improve the accuracy of landslide detection,especially the frequency domain features are more sensitive to the edges of feature information.The application of full-convolution autoencoder is able to learn deeper features of landslides,avoiding the limitations of manual feature design.In addition,the results of validation experiments in different study areas show that the random forest model has better transferability compared to the support vector machine model,with an accuracy of up to 92.1%,while the support vector machine model has an accuracy of up to 87.3%.The results of the feature importance assessment showed that spectral features are very important for landslide extraction.These findings provide an important reference for landslide detection and cataloguing construction. |
