Leveraging CNNs-RNNs and Adam Optimization: A Revolutionary Deep Learning approach to landslide prediction in district of Than Uyen, Lao Cai province, Vietnam

Abstract

This paper presents an innovative deep learning framework for predicting landslides in the District of Than Uyen, Lao Cai Province, Vietnam, harnessing the power of Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), and the Adam Optimization algorithm. Given the multifaceted nature of landslide phenomena, influenced by a myriad of geographical and meteorological factors, there is a growing need for advanced computational strategies that can decipher complex patterns and temporal correlations.

Our cutting-edge model employs CNNs to analyze and extract significant spatial attributes from topographical and geological data. Concurrently, RNNs -specifically Long Short-Term Memory (LSTM) networks - are deployed to manage time-series data, such as weather conditions and other temporal elements. The Adam Optimization algorithm, renowned for its superior efficiency and effective performance, is used to optimize the model parameters.

The model was trained and validated using a comprehensive dataset from the Than Uyen district, comprising 114 landslide and 114 non-landslide locations, along with ten key influential factors: elevation, slope, curvature, aspect, relief amplitude, soil type, geology, and proximity to faults, roads, and rivers. The results demonstrate a noteworthy predictive accuracy, sensitivity, and specificity, with the model surpassing benchmarks in prediction power (PPV=93.3%, NPV=83.2%, Sen=82.3%, Spe=94.1%, Acc=88.2%, F-score=0.875, Kappa=0.765, and AUC=0.968).

The study advances deep learning in landslide prediction, aiding proactive disaster mitigation and showing potential for application in global regions prone to geographical hazards.