Geographic object-based image analysis for landslide identification using machine learning on google earth engine

Landslides significantly threaten human life and infrastructure, requiring accurate and timely identification for effective hazard assessment and management. This study proposes a new approach combining Geographic Object-Based Image Analysis (GEOBIA) and machine learning on the Google Earth Engine (GEE) platform, utilizing high-resolution Sentinel-2 imagery and NASADEM data. Our methodology begins with Simple Non-iterative Clustering (SNIC) segmentation, which divides the images into homogeneous super-pixels. This step is crucial for reducing 'salt and pepper' noise and enhances the differentiation of spectrally similar objects through advanced texture, shape, and contextual analysis. Following segmentation, Gray Level Co-occurrence Matrix (GLCM) feature extraction is employed to gather critical textural information, which is pivotal in discerning surface roughness, heterogeneity, and composition-key factors in identifying landslide-prone areas. To manage the high dimensionality of the data, Principal Component Analysis (PCA) is utilized for dimensionality reduction, transforming original variables into a set of uncorrelated principal components that facilitate more efficient subsequent analysis. Various machine learning algorithms are utilized, including Support Vector Machine (SVM), Random Forest (RF), and Classification and Regression Trees (CART). We use the GEE platform to leverage extensive geospatial data and computational power. The performance of SVM, RF, and CART algorithms is evaluated for landslide detection. RF demonstrates superior accuracy in detecting landslides, achieving an overall accuracy of 87.41%, surpassing SVM (85.47%) and CART (68.45%). Integrating SNIC segmentation, GLCM feature extraction, PCA analysis, and RF algorithm within the GEOBIA framework using the GEE platform shows promising results for improving landslide identification, monitoring, and risk assessment.