Using machine learning algorithms to develop a remotely-sensed framework for drought monitoring in different climate regions in South Africa

Loading...
Thumbnail Image

Date

2025

Journal Title

Journal ISSN

Volume Title

Publisher

University of the Western Cape

Abstract

Droughts pose a significant threat to rainfed smallholder farming systems, particularly in regions with varying climatic conditions. This study aimed to develop a spatial modelling framework for assessing the occurrence and frequency of droughts across different climatic zones, with a focus on rainfed smallholder farms. Specifically, this study sought to develop a spatial modelling framework for assessing the occurrence and frequency of droughts on rainfed smallholder farms across different climatic zones in South Africa i.e. the Limpopo, North West, Mpumalanga and Gauteng Provinces. By integrating satellite-derived vegetation indices, specifically the Normalised Difference Vegetation Index (NDVI) and the Modified Soil-Adjusted Vegetation Index 2 (MSAVI2), with state-of-the-art machine learning algorithms, the accuracy of drought mapping in rainfed smallholder farms was significantly enhanced. The study employed the Gradient Boosting Decision Tree (GBDT), Support Vector Machine (SVM), Convolutional Neural Network (CNN), Random Forest (RF) and Transformer models to capture complex spatial and temporal patterns of drought dynamics. The results showed that the Transformer model was effective in detecting rainfed smallholder farms (with an Overall Accuracy of 0.85 and a mean Intersection over Union (IoU) of 0.86). The study further evaluated the agricultural and meteorological drought conditions from 2004 to 2023. The Vegetation Condition Index (VCI) from SPOT VEGETATION 1 and PROBA-V data and the Standardised Precipitation Index (SPI) from CHIRPS data were computed.

Description

Keywords

Advanced Very High-Resolution, Radiometer, Convolutional Neural Network, Digital Elevation Model, Evapotranspiration

Citation

N/A