Satellite Imagery Analysis For Landslides Prediction Via Artificial Intelligence

Landslides in mountainous and hilly terrains arise due to both natural phenomena and human activities. Natural triggers such as excessive rainfall, seismic shifts, and soil moisture fluctuations play a major role, while human-induced factors like uncontrolled infrastructure development further amplify the risks. These occurrences lead to significant destruction of both human life and property, emphasizing the necessity of early detection to mitigate their impact. In recent years, machine learning methods have gained prominence in automating landslide identification. Various techniques, including feature extraction and classification, have been applied to satellite imagery to create semi-automated systems for predicting and detecting landslides. However, achieving fully automated detection with superior precision remains a critical challenge. One major limitation in landslide classification and prediction using satellite images is the need for a systematically structured dataset to ensure accurate and reliable testing outcomes. This study examines a range of approaches used for landslide detection and classification through satellite images, aiming to identify gaps in existing research. Moreover, it proposes an advanced prototype designed to enhance classification accuracy. A comprehensive review of 50 academic papers from well-regarded journals has been conducted, focusing on machine learning and deep learning algorithms. The study evaluates and contrasts various classification models based on their accuracy and effectiveness. To address the identified limitations, an optimized deep learning framework built upon ResNet101 is introduced, attaining an accuracy of 96.88% on an augmented Beijing dataset containing 770 satellite images. This research contributes valuable knowledge on emerging advancements, ongoing challenges, and prospective developments in machine learning and deep learning for landslide detection. It serves as a significant resource for researchers aiming to explore innovative methodologies in this domain