A Modified Mask-RCNN Algorithm for Intelligent Identification of Landslide Based on High-resolution Remote Sensing data

<p>Deep learning is a data-driven approach that requires high-quality labeled data to construct training and evaluation datasets. However, there are few open landslide data sets at present, and the degree of standardization of data sets is low. Now, the most advanced instance segmentation algorithms require strongly supervised learning. The cost of acquiring new categories of images is prohibitive. A question is raised: Is it possible to train high-quality instance segmentation models for early landslide disaster identification on the premise that not all categories are marked with complete instance segmentation annotations?</p><p>This article mainly deals with the intelligent identification of the small and medium-scale loess-bedrock historical landslides in the east Gansu Province. We proposed a modified instance segmentation algorithm based on transfer learning. Specifically, (1) A self-made landslide dataset was constructed. Google Earth images were used as the data source, and Arc GIS was selected as the landslide interpretation software. Based on DEM and 1:50,000 detailed regional geological hazard survey data, landslide boundaries were manually circled using the dataset annotation software&#65288;label me&#65289;according to the landslides' features of color, spectrum, vein, and surface roughness in optical images. The method of regional separation of datasets was used, with Anding district of Dingxi city as the validation set (15%), and Tianshui city, Longnan city, and Qingyang city as the sampling areas of the training set (70%) and testing set (15%) in the dataset. (2) A novel segmentation algorithm for landslide instances was proposed. The algorithm combined partially supervised training with weight transfer function to achieve high precision landslide classification and boundary recognition on data set constructed by mixed label annotation method. (3) A new method of Mask scoring was adopted to solve the problem that the accuracy of instance segmentation was affected by the lack of Mask scoring in Mask-RCNN.</p><p>The results show that the proposed method is superior to other algorithms in precision, accuracy, and recall rate. In addition, the Mask-IOU threshold value of 0.5 was used to estimate the average accuracy higher than the Mask-IOU threshold value of 0.75. The improved algorithm is in the segmentation of small and medium-sized landslides better than for large landslides, which will help solve the problem that it is difficult to comprehensively monitor the small and medium-sized landslides in the geological field survey. And our algorithm is not sensitive to the diffident backbone network and can achieve stable improvement on different Backbones. The average accuracy is about 3.1. The result of the experiment verified with the landslide field survey data in the validation area demonstrates this algorithm is stable and adaptable.</p>