Identifying recurrent and persistent landslides using satellite imagery and deep learning: A 30-year analysis of the Himalaya.

This paper presents a remote sensing-based method to efficiently generate multi-temporal landslide inventories and identify recurrent and persistent landslides. We used free data from Landsat, nighttime lights, digital elevation models, and a convolutional neural network model to develop the first multi-decadal inventory of landslides across the Himalaya, spanning from 1992 to 2021. The model successfully delineated >265,000 landslides, accurately identifying 83 % of manually mapped landslide areas and 94 % of reported landslide events in the region. Surprisingly, only 14 % of landslide areas each year were first occurrences, 55-83 % of landslide areas were persistent and 3-24 % had reactivated. On average, a landslide-affected pixel persisted for 4.7 years before recovery, a duration shorter than findings from small-scale studies following a major earthquake event. Among the recovered areas, 50 % of them experienced recurrent landslides after an average of five years. In fact, 22 % of landslide areas in the Himalaya experienced at least three episodes of landslides within 30 years. Disparities in landslide persistence across the Himalaya were pronounced, with an average recovery time of 6 years for Western India and Nepal, compared to 3 years for Bhutan and Eastern India. Slope and elevation emerged as significant controls of persistent and recurrent landslides. Road construction, afforestation policies, and seismic and monsoon activities were related to changes in landslide patterns in the Himalaya.

Higher depression risks in medium- than in high-density urban form across Denmark.

Urban areas are associated with higher depression risks than rural areas. However, less is known about how different types of urban environments relate to depression risk. Here, we use satellite imagery and machine learning to quantify three-dimensional (3D) urban form (i.e., building density and height) over time. Combining satellite-derived urban form data and individual-level residential addresses, health, and socioeconomic registers, we conduct a case-control study (n = 75,650 cases and 756,500 controls) to examine the association between 3D urban form and depression in the Danish population. We find that living in dense inner-city areas did not carry the highest depression risks. Rather, after adjusting for socioeconomic factors, the highest risk was among sprawling suburbs, and the lowest was among multistory buildings with open space in the vicinity. The finding suggests that spatial land-use planning should prioritize securing access to open space in densely built areas to mitigate depression risks.

Multi-hazard susceptibility and exposure assessment of the Hindu Kush Himalaya.

Mountainous regions are highly hazardous, and these hazards often lead to loss of human life. The Hindu Kush Himalaya (HKH), like many mountainous regions, is the site of multiple and overlapping natural hazards, but the distribution of multi-hazard risk and the populations exposed to it are poorly understood. Here, we present high-resolution transboundary models describing susceptibility to floods, landslides, and wildfires to understand population exposure to multi-hazard risk across the HKH. These models are created from historical remotely sensed data and hazard catalogs by the maximum entropy (Maxent) machine learning technique. Our results show that human settlements in the HKH are disproportionately concentrated in areas of high multi-hazard risk. In contrast, low-hazard areas are disproportionately unpopulated. Nearly half of the population in the region lives in areas that are highly susceptible to more than one hazard. Warm low-altitude foothill areas with perennially moist soils were identified as highly susceptible to multiple hazards. This area comprises only 31% of the study region, but is home to 49% of its population. The results also show that areas susceptible to multiple hazards are also major corridors of current migration and urban expansion, suggesting that current rates and patterns of urbanization will continue to put more people at risk. This study establishes that the population in the HKH is concentrated in areas susceptible to multiple hazards and suggests that current patterns of human movement will continue to increase exposure to multi-hazards in the HKH.

A framework for large-scale mapping of human settlement extent from Sentinel-2 images via fully convolutional neural networks.

Human settlement extent (HSE) information is a valuable indicator of world-wide urbanization as well as the resulting human pressure on the natural environment. Therefore, mapping HSE is critical for various environmental issues at local, regional, and even global scales. This paper presents a deep-learning-based framework to automatically map HSE from multi-spectral Sentinel-2 data using regionally available geo-products as training labels. A straightforward, simple, yet effective fully convolutional network-based architecture, Sen2HSE, is implemented as an example for semantic segmentation within the framework. The framework is validated against both manually labelled checking points distributed evenly over the test areas, and the OpenStreetMap building layer. The HSE mapping results were extensively compared to several baseline products in order to thoroughly evaluate the effectiveness of the proposed HSE mapping framework. The HSE mapping power is consistently demonstrated over 10 representative areas across the world. We also present one regional-scale and one country-wide HSE mapping example from our framework to show the potential for upscaling. The results of this study contribute to the generalization of the applicability of CNN-based approaches for large-scale urban mapping to cases where no up-to-date and accurate ground truth is available, as well as the subsequent monitor of global urbanization.