Paradise-theking/Eswatini-Water-Stress-ML
Domaine:
environment and energyType de record:
softwaremodel
Créateur:
Par
Hôte:
Machine learning-based one-month-ahead forecasting of hydroclimatic water stress in Eswatini using remote-sensing and climate data.
# Machine Learning-Based Forecasting of Hydroclimatic Water Stress in Eswatini
A leakage-aware machine learning framework and interactive forecasting application for **one-month-ahead prediction of hydroclimatic water stress in Eswatini** using remotely sensed and reanalysis-derived environmental variables.
The project investigates whether hydroclimatic conditions observed at month **t** can provide useful predictive information about water stress at month **t+1**, with particular emphasis on chronological validation, leakage prevention, regularization, independent testing, and comparison against simple forecasting baselines.
The trained forecasting pipeline is also integrated into a **FastAPI backend** and **TypeScript/Vite dashboard** for interactive model inference and historical Water Stress Index visualization.
---
## Research Overview
Eswatini is vulnerable to recurrent drought, rainfall variability, increasing temperatures, and associated pressures on agriculture and water resources.
Reliable short-term hydroclimatic forecasting could complement conventional drought monitoring by providing advance information to support climate-resilient water-resource management.
This project develops a monthly water-stress forecasting framework using environmental variables including:
- precipitation;
- potential evapotranspiration (PET);
- temperature;
- soil moisture;
- runoff;
- surface runoff;
- solar radiation;
- dew point;
- wind speed; and
- lagged and accumulated hydroclimatic variables.
The final modelling experiment uses **Ridge Regression** to forecast the Water Stress Index one month ahead.
---
## Forecasting Objective
The forecasting problem is formulated as:
> Use hydroclimatic information available at month **t** to predict the Water Stress Index at month **t+1**.
This distinction is important because the project is designed as a genuine forecasting experiment rather than a same-month estimation exercise.
The forecasting horizon is therefore: …