Landslide mapping is the primary step for landslide investigation and prevention. At present, both the accuracy and the degree of automation of landslide mapping with remote sensing (LMRS) are still lower than those of general remote sensing classification. In order to improve the performance, previous attempts have been made to develop new features, classifiers, and rules, whereas few studies have investigated the in-depth causes and the corresponding solutions. In this paper, after reviewing the related literature, some of the fundamental difficulties hindering the improvement of LMRS are disclosed and discussed. Firstly, landslides do not have distinguishable spectral, spatial, or temporal characteristics, as they may actually be covered by other land covers. Secondly, the surface features of a landslide can vary greatly, affected by the different geological factors, geomorphological factors, hydrological factors, weather conditions, and other factors. Thirdly, the differences in the surface features are often remarkable and nonnegligible, and thus it is difficult to identify a landslide with only a few simple criteria. Finally, some solutions to the above difficulties are suggested. It is expected that the accuracy and applicability of LMRS could be greatly improved, by exploiting big data, utilizing the deep learning technique, and modelling the surface spatial structure of the landslide.

In recent years, hundreds of Earth observation satellites have been launched to collect massive amounts of remote sensing images. However, the considerable cost and time to process the significant amount of data have become the greatest obstacle between data and knowledge. In order to accelerate the transformation from remote sensing images to urban thematic maps, a strategy to map the bare land automatically from Landsat imagery was developed and assessed in this study. First, a normalized difference bare land index (NBLI) was presented to maximally differentiate bare land from other land types in Wuhan City, China. Then, an unsupervised classifier was employed to extract the bare land from the NBLI image without training samples or self-assigned thresholds. Experimental results showed good performance on overall accuracy (92%), kappa coefficient (0.84), area ratio (1.1321), and match rate (83.96%), respectively. Results in multiple years disclosed that bare lands in the study site gradually moved from inner loops to the outer loops since 2007, in two main directions. This study demonstrated that the proposed method was an accurate and reliable option to extract the bare land, which led to a promising approach to mapping urban land use/land cover (LULC) automatically with simple indices.

Landslides have been frequently occurring in the high mountainous areas in China and poses serious threats to peoples’ lives and property, economic development, and national security. Detecting and monitoring quiescent or active landslides is important for predicting risks and mitigating losses. However, traditional ground survey methods, such as field investigation, GNSS, and total stations, are only suitable for field investigation at a specific site rather than identifying landslides over a large area, as they are expensive, time-consuming, and laborious. In this study, the feasibility of using SBAS-InSAR to detect landslides in the high mountainous areas along the Yunnan Myanmar border was tested first, with fifty-four IW mode Sentinel-1A ascending scenes from 12 January 2019 to 8 December 2020. Next, the Yolo deep-learning model with Gaofen-2 images captured on 5 December 2020 was tested. Finally, the two techniques were combined to achieve better performance, given each of them has intrinsic limitations on landslide detection. The experiment indicated that the combination could improve the match rate between detection results and references, which implied that the performance of landslide detection can be improved with the fusion of time series SAR images and optical images.

Post-seismic vegetation recovery is critical to local ecosystem recovery and slope stability, especially in the Wenchuan earthquake area where tens of thousands of landslides were triggered. This study executed a decadal monitoring of post-seismic landslide activities all over the region by investigating landslide vegetation recovery rate (VRR) with Landsat images and a (nearly) complete landslide inventory. Thirty thousand landslides that were larger than nine pixels were chosen for VRR analysis, to reduce the influence of mixed pixels and support detailed investigation within landslides. The study indicates that about 60% of landslide vegetation gets close to the pre-earthquake level in ten years and is expected to recover to the pre-earthquake level within 20 years. The vegetation recovery is significantly influenced by topographic factors, especially elevation and slope, while it is barely related to the distance to epicenter, fault ruptures, and rivers. This study checked and improved the knowledge of vegetation recovery and landslide stability in the area, based on a detailed investigation.

Land use/land cover (LULC) change is one of the most important indicators in understanding the interactions between humans and the environment. Traditionally, when LULC maps are produced yearly, most existing remote-sensing methods have to collect ground reference data annually, as the classifiers have to be trained individually in each corresponding year. This study presented a novel strategy to map LULC classes without training samples or assigning parameters. First of all, several novel indices were carefully selected from the index pool, which were able to highlight certain LULC very well. Following this, a common unsupervised classifier was employed to extract the LULC from the associated index image without assigning thresholds. Finally, a supervised classification was implemented with samples automatically collected from the unsupervised classification outputs. Results illustrated that the proposed method could achieve satisfactory performance, reaching similar accuracies to traditional approaches. Findings of this study demonstrate that the proposed strategy is a simple and effective alternative to mapping urban LULC. With the proposed strategy, the budget and time required for remote-sensing data processing could be reduced dramatically.

Landslide inventories could provide fundamental data for analyzing the causative factors and deformation mechanisms of landslide events. Considering that it is still hard to detect landslides automatically from remote sensing images, endeavors have been carried out to explore the potential of DCNNs on landslide detection, and obtained better performance than shallow machine learning methods. However, there is often confusion as to which structure, layer number, and sample size are better for a project. To fill this gap, this study conducted a comparative test on typical models for landside detection in the Wenchuan earthquake area, where about 200,000 secondary landslides were available. Multiple structures and layer numbers, including VGG16, VGG19, ResNet50, ResNet101, DenseNet120, DenseNet201, UNet−, UNet+, and ResUNet were investigated with different sample numbers (100, 1000, and 10,000). Results indicate that VGG models have the highest precision (about 0.9) but the lowest recall (below 0.76); ResNet models display the lowest precision (below 0.86) and a high recall (about 0.85); DenseNet models obtain moderate precision (below 0.88) and recall (about 0.8); while UNet+ also achieves moderate precision (0.8) and recall (0.84). Generally, a larger sample set can lead to better performance for VGG, ResNet, and DenseNet, and deeper layers could improve the detection results for ResNet and DenseNet. This study provides valuable clues for designing models’ type, layers, and sample set, based on tests with a large number of samples.