Landslide semantic segmentation using satellite imagery

This work aims to improve the accuracy of landslide detection and detecting landslide scar with single satellite imagery by using slope factor extraction and normalized difference vegetation index (NDVI) extraction and by combining three semantic segmentation models based on a convolutional neural network with a classification/decision tree. The proposed method used three semantic segmentation models with a decision tree. The first model is trained by a training set of color images.The second model is trained by a training set of slope factor images. The third model is trained by a training set of NDVI images. The slope factor and NDVI are extracted from color images that contain red, green, and blue bands. The results from threemodels are used to generate features for training the classification tree. Evaluation metrics (precision, recall, and F2 score) can be improved by using slope factor and NDVI. In combining three models, the F1 score and F2 score are increased more thanusing a single of color images 16.71% and 24.15%, respectively in Resnet18, and increased16.80% and 26.20% respectively in Resnet50. Moreover, the slope factor detection model and NDVI detection model can support some areas that the color imagedetection model cannot detect.

本研究旨在通过提取坡度因子与归一化差分植被指数（Normalized Difference Vegetation Index, NDVI），并将三种基于卷积神经网络（Convolutional Neural Network, CNN）的语义分割模型（Semantic Segmentation Model）与分类/决策树相结合，提升单卫星影像滑坡及其滑坡痕迹的检测精度。
所提方案采用三类语义分割模型配合决策树：第一类模型以彩色图像训练集完成训练；第二类模型以坡度因子图像训练集完成训练；第三类模型以NDVI图像训练集完成训练。其中，坡度因子与NDVI均从包含红、绿、蓝波段的彩色影像中提取得到。
随后将三个模型的输出作为特征，用于训练分类树。借助坡度因子与NDVI，可提升精准率（precision）、召回率（recall）与F2分数等评估指标。在融合三类模型的场景下，基于Resnet18架构的模型中，F1分数与F2分数分别较单一彩色影像模型提升16.71%与24.15%；而在Resnet50架构下，二者分别提升16.80%与26.20%。此外，坡度因子检测模型与NDVI检测模型可覆盖部分彩色影像检测模型无法识别的区域。