Object‐based Segmentation and Machine Learning Classification for Landslide Detection from Multi‐temporal WorldView‐2 Imagery

Abstract

Landslides are pervasive hazards that pose significant risk to human populations. Routine quantification of landslide occurrence is necessary for hazard mitigation, traditionally compiled from manual interpretation of aerial imagery. To increase precision, reduce costs, and expedite analysis, much effort is focused on landslide identification from satellite imagery, with object-based methods rapidly emerging as a viable approach. Recent work has also utilized machine learning classifiers to increase automation and transferability. This study built on previous work to apply object-based image analysis (OBIA) and machine learning classification to sub-meter and multi-temporal WorldView-2 imagery. The primary objective was to explore scenarios resulting in optimal classification, considering: (1) random forest (RF) versus support vector machine (SVM) classifiers, (2) multispectral versus fused image resolutions, (3) binary versus multi--‐class structures, and (4) variations in sample size. A study area was selected involving challenging image composition and an extended capture window to test the robustness of the method to non-ideal conditions. The eCognition software allowed for image segmentation. Following selection of training samples, the R software was then utilized for machine learning classification with both RFs and SVMs. Classification was performed for each parameter combination over 100 replications, with accuracy assessed against a manual reference inventory. Optimal results were observed for RF at the largest sample size using a binary class structure and fused resolution, with an average F-score of 60.2 ± 1.3%. RF classifications consistently reached ∼3-5% higher accuracy versus SVM when compared between specific parameter combinations. RFs demonstrated higher run-to-run stability, both in terms of spatial results and lower variance by area, as well as lower processing cost by an order of magnitude. These findings aid future work in determining optimal classification frameworks. The need for future research is also highlighted, including automation of sample selection and further refinement of the image segmentation task.