Every year, landslides kill thousands of people and wipe out infrastructure worth billions of dollars, yet predicting where and when they will occur remains stubbornly difficult. Over the last decade or so, the combination of freely available satellite imagery and fast-maturing machine learning (ML) and deep learning (DL) techniques has begun to change that picture in meaningful ways. This paper reviews the trajectory of that change examining how researchers have moved from simple statistical classifiers to sophisticated transformer-based segmentation models, and what that progression has actually delivered in terms of practical detection capability. We surveyed around 45 peer-reviewed studies published between 2015 and 2024, covering landslide susceptibility mapping, pixel-level inventory mapping, change detection from multi-temporal imagery, and real-time early warning integration. Our analysis draws on work using optical sensors (Sentinel-2, Landsat), SAR platforms (Sentinel-1, ALOS-2), and a growing variety of public benchmark datasets such as Bijie, COOLR, and HR-GLDD. We find that while ensemble classifiers like Random Forest still hold their own for susceptibility mapping, encoder-decoder architectures particularly U-Net variants have become the workhorse for segmentation tasks, with more recent transformer hybrids pushing IoU scores above 0.85 on standard benchmarks. That said, the field carries some persistent and underappreciated weaknesses: almost all top-performing models have been trained and tested within narrow geographic windows; labeled data remains scarce outside China, Italy, and Central Europe; and the jump from research prototype to operational warning system has proven far harder than benchmark numbers suggest. We close the review by pointing to foundation models, physics-informed learning, and federated training as directions that may genuinely move the needle on these limitations.
D.B.Mirajkar et al. (2026) studied this question.