Advanced Workflow for Landslide Detection and Scar Analysis
This study outlines a comprehensive workflow for detecting landslides and analyzing their scars, primarily using Airborne Laser Scanning (ALS) data.
The process involved several key stages, from initial data preparation to detailed geomorphological analysis.
Data Collection and Pre-processing
The initial phase involved collecting and preparing various datasets for landslide analysis within the Sellrain valley. Crucial inputs included ALS acquisitions from 2013 and 2017, which provided classified point-clouds and 0.5-meter resolution Digital Terrain Models (DTMs).
A Difference of DTMs (DoD) was subsequently created using these DTMs. Furthermore, the 2013 ALS data was utilized to generate several derived products: a binary forest canopy cover mask, single tree positions, a tree distance map, and a tree neighbor count map. Ground point density and elevation roughness were also derived from the ALS point-clouds for both years.
Supplementary data included a 0.1-meter resolution orthophoto from June 2015 for validation, historic orthophotos (1988-2015) for landcover change analysis, and stream and road network data with 15-meter buffers for masking purposes.
Landslide Inventory and Non-Landslide Area Construction
A robust, polygon-based landslide inventory was developed from an existing point-based inventory using a seeded region growing algorithm in SAGA GIS. This intricate process utilized a downscaled DoD and the 2015 orthophoto. The delineated scars underwent rigorous validation against the orthophoto and various DoD derivatives, including slope, hillshade, and topographic openness.
To enhance the training accuracy for subsequent models, areas without any signs of landslides, referred to as "non-landslide areas," were also meticulously visually delineated using the orthophoto and DoD data.
Landslide Detection Using a Random Forest Model
An additional landslide detection component specifically focused on identifying landslides within forested areas, where existing inventories typically had limited samples. A Random Forest (RF) model, implemented with the scikit-learn Python library, was trained using the newly created polygon-based inventory and the non-landslide areas.
The model's input features were comprehensive, including the DoD and four critical uncertainty proxy layers: point density and elevation roughness derived from both the 2013 and 2017 ALS data.
The RF model underwent meticulous hyperparameter tuning using the Optuna library, with optimization focused on maximizing the Area Under the Curve (AUC) and Jaccard index through a three-fold cross-validation approach.
Following training, the RF model's probability output was further refined through a sophisticated segmentation workflow. This workflow comprised several stages: initial filtering, seed extraction, seeded region growing, and segment clumping, all performed using GRASS GIS. The parameters for this detailed segmentation process were also optimized using Optuna, prioritizing the Jaccard index to effectively filter false positives and improve accuracy.
Landslide Scar Mapping and Analysis
The final step involved precisely mapping landslide scars (depletion zones) within the areas identified by the RF model. This was primarily achieved by applying the convergence index from SAGA GIS to the DoD dataset. A specific negative threshold of -30% on the convergence index proved effective in delineating scar areas distinctively from other erosion features.
Resulting segments that overlapped with the landslide detection output were then subjected to further validation against DoD and orthophoto data. They were subsequently filtered based on a minimum scar size, which was derived from the characteristics of the initial inventory.
The meticulously mapped scars were then analyzed for their morphological and topographic characteristics, providing crucial insights.
This analysis considered their location relative to a forest canopy cover mask (classified as "within forest" for areas with ≥90% canopy cover or "outside forest" for 0% canopy cover). It also included metrics like the closest distance to a tree and the average number of trees within a 10-meter radius of scar cells. To ensure historical context and robust analysis, only scars located in areas without landcover changes in the 20 years preceding the 2015 storm event were included in this final analysis, a condition determined by visual inspection of historic aerial imagery.