arefbakali/drought-risk-prediction-tunisia

Machine Learning, Deep Learning and spatial visualization for drought risk prediction in Tunisia. # Drought Risk Prediction in Tunisia ## Project Date **May 2026** ## Overview This project focuses on drought risk prediction in Tunisia using Machine Learning, Deep Learning and spatial visualization. The objective is to model drought-related risk from climatic, temporal and spatial variables, compare several predictive approaches, and generate a territorial interpretation of drought exposure across Tunisian governorates. The project follows a complete modeling workflow: tabular Machine Learning models, clustering enrichment, neural networks, temporal LSTM forecasting, and interactive cartographic visualization. ## Project Context Drought is a major environmental and agricultural challenge, especially in regions exposed to climate variability. Predicting drought risk can support decision-making by identifying areas that may require more attention in terms of water resource management, agriculture and environmental monitoring. This project investigates the following question: > Can Machine Learning and Deep Learning models predict drought risk in Tunisia using climatic, temporal and spatial indicators? ## Dataset The project uses the `ex16_secheresse_spi.csv` dataset. * Number of observations: **2520** * Number of columns: **32** * Period covered: **15/01/2017 to 15/09/2025** * Missing values: **No missing values detected** * Geographic scope: **Tunisia** Main types of variables: * Climatic indicators * Temporal lag variables * Monthly variables * Latitude and longitude * UTM coordinates * Drought-related target variables The dataset is stored in the `data/` folder: ```text data/ex16_secheresse_spi.csv ``` ## Methodology The project is divided into four modeling stages and one spatial visualization stage. ### Part A — Baseline Tabular Models Random Forest and XGBoost were trained as baseline models to predict the tabular drought risk score. ### Part B — KMeans Enrichment A `cluster_kmeans` feature was added after selecting the number of cluster …