Boosting Ensemble Learning Technique for Landslide Activity Classification using Vegetation Anomalies Indicators

Abstract

Having detailed information and inventories about landslides is important for studying landslides. These inventories have been created for various purposes. However, detecting and mapping landslides in highly dense vegetated areas and evaluating their activity state is a major challenge due to various factors, such as the dense tree canopy, undulating terrain, and fast-growing vegetation. Therefore, this paper presents a new technique for categorising landslide activity using vegetation anomalies indicators (VAIs) extracted from high-resolution remote sensing data. The data were utilised to support manual landslide inventory and VAI production. The landslide inventory map was divided randomly into two groups of datasets, one for training (70%) and the other for validation (30%). The classification process used a boosting ensemble learning approach, specifically Decision Tree (DT) and Stochastic Gradient Boosting (SGB), with seven primary VAIs as inputs. The study compared the classification models’ performance against various parameters, including spatial resolution and landslide depth. To evaluate the accuracy of the classification methods, metrics such as overall accuracy, kappa, producer’s accuracy, and user’s accuracy were measured from the validation dataset. The results demonstrated that both methods performed best under high spatial resolution. Among the two approaches, DT performed better, with an overall accuracy value of 89.6% for deep-seated translational, 64.0% for shallow translational, 67.0% for deep-seated rotational, and 80.0% for shallow rotational. This reliable accuracy that has been attained in landslide activity classification from VAI allows (i) to map and classify the landslide activity in the forested area, (ii) characterise the different types of vegetation characteristics to specific landslide activity, and (iii) permits for the continuous landslide activity monitoring in the area where monitoring activity is not practically feasible to be conducted.