A Comprehensive Review of Deep Learning Approaches for Landslide Detection Using Remote Sensing and Object-Based Image Analysis

Landslides are among the most destructive natural hazards worldwide, causing loss of life, damage to infrastructure, and severe environmental degradation. Rapid and accurate detection of landslides is essential for effective risk mitigation and early warning. Recent developments in remote sensing (RS), deep learning (DL), and object-based image analysis (OBIA) have significantly improved the precision and automation of landslide detection. This paper presents a comprehensive review of the integration of DL and OBIA in landslide detection using multisource RS data. It traces the evolution from traditional pixel-based and machine learning (ML) approaches to advanced deep neural architectures such as CNN, U-Net, ResU-Net, and Transformer models. The study highlights the advantages of combining DL with rule-based OBIA for improved boundary delineation and reduced noise. Comparative analyzes of hybrid models show that the integration of DL-OBIA improves the F1score by 8 to 22 % compared to the DL or OBIA techniques alone. The review identifies key limitations-limited annotated datasets, weak generalization across regions, and challenges in multi-source fusion-and discusses future research directions for scalable, interpretable, and efficient landslide detection frameworks.