Mapping Irrigated Agriculture in Fragmented Landscapes of Sub-Saharan Africa: An Examination of Algorithm and Composite Length Effectiveness

This study evaluates the effectiveness of different machine learning algorithms and temporal composite lengths for mapping irrigated agriculture in heterogeneous landscapes of sub-Saharan Africa. We tested four classifiers (Random Forest, Support Vector Machine, Artificial Neural Networks, and K-Nearest Neighbors) and four composite lengths (1×12-monthly, 2×6-monthly, 4×3-monthly, and 6×2-monthly) across study areas in Mozambique. Results show that no single algorithm consistently outperforms others, but Artificial Neural Networks, Support Vector Machines, and Random Forest all performed effectively. For complex and heterogeneous landscapes, shorter composites were more suitable, while longer composites sufficed for more uniform areas. The study introduces ‘agreement maps’ that highlight areas of classification consensus, improving irrigated area identification accuracy.

Journal Article

International Journal of Applied Earth Observation and Geoinformation