Mapping irrigated agriculture in fragmented landscapes of sub-Saharan Africa: An examination of algorithm and composite length effectiveness

Abstract

Accurately identifying irrigated areas is crucial for sustainable development, food security, and effective land and water resource management. However, incomplete or outdated national estimates of irrigated areas underestimate the extent of it, particularly among smallholders. This study aimed to address this issue by investigating the impact of different algorithms and composite lengths on predicting irrigated agriculture in four study areas in Mozambique. The study found that the choice of algorithm and composite length notably impacted the accuracy of identifying irrigation. Shorter composite lengths, such as 2-monthly or 3-monthly composites, were more effective in identifying irrigation in fragmented and dynamic landscapes, while longer composite lengths were better suited to stable classes and homogeneous landscapes. Artificial neural networks, support vector machines, and random forests were all effective algorithms for classifying irrigation. However, the study emphasised the importance of considering hotspots and agreement maps when identifying irrigation. Agreement maps combine the classification results of multiple models, providing better insights into the core areas of irrigated agriculture and allowing for a better understanding of irrigation dynamics and policy decision-making, particularly among smallholder systems. This research provides valuable insights for those working on remote sensing-based irrigation mapping and monitoring in sub-Saharan Africa, focusing on identifying smallholder irrigation with greater certainty.

Keywords

• Random forest
• Support vector machine
• Neural network
• k-nearest neighbor
• Hotspot map
• Remote sensing