Attention Based CNN-BiLSTM Framework for Explainable Flood and Landslide Prediction Using Multimodal Environmental Data

Abstract

Natural disasters such as floods and landslides impose severe socioeconomic losses across vulnerable geographies. Accurate early prediction demands model capable of capturing both Spatial and a Temporal patterns embedded in environmental data. This paper presents a novel Attention-Based Convolutional -Neural Network(CNN) and Bidirectional Long short-term memory (CNN-BiLSTM-Attention) architecture for simultaneous, multi-output prediction of flood and landslide events. The proposed framework integrates seven geophysical and meteorological parameters—rainfall, temperature, soil moisture, river discharge, slope gradient, Normalized Difference Vegetation Index (NDVI), and historical event frequency—processed through a deep learning pipeline comprising one-dimensional convolution, bidirectional recurrence, and a self-attention mechanism. A GIS-interactive interface built with Folium and Streamlit enables real-time location selection and parameter input. Experiments on 7,000 records yield flood prediction accuracy of 95.7% and landslide prediction accuracy of 94.3%, outperforming baseline LSTM, BiLSTM, and CNN-LSTM models. An integrated rule-based explainable AI (XAI) module provides human-interpretable risk justifications alongside quantitative predictions.