Tropentag, September 16 - 18, 2026, Göttingen "Towards multi-functional agro-ecosystems promoting climate-resilient futures"

Leveraging deep learning with very high resolution satellite imagery to map Faidherbia albida (DEL.) trees, and support agroforestry decision-making in northern Cameroon

¹The University of Maroua, Cifor-Icraf Cameroon, Cirad, Geography, Cameroon
²CIRAD, Forests and Societies Research Unit - Environment and Society, France

Abstract

Faidherbia albida is a key agroforestry species widely distributed across the Sudano-Sahelian zone of Cameroon, where it plays a crucial role in enhancing soil fertility, providing fuelwood, and strengthening climate resilience. Since the 1990s, several initiatives have promoted the expansion of agroforestry systems through farmer-managed natural regeneration (FMNR), supported by financial incentives. However, rapid population growth, climatic constraints, and restrictive forestry regulations have often limited the effectiveness of these interventions. In addition, the lack of accurate, long-term spatial data has constrained the monitoring of the species’ distribution and dynamics.
This study aims to improve the mapping of individual tree crowns (ITCs) of F. albida using very high-resolution (VHR) satellite imagery. A deep learning-based approach (CNN) was applied to a sub-metric spatial resolution imagery (0,5 m) Wordview-2 to detect and accurately delineate ITCs (crown area > 3 m²) across a 15,000 ha (150 km²) over the study area. Using a georeferenced field inventory dataset, five classifiers (CART, kNN, Random Forest, SVM-L, and SVM-R) were trained to predict species map at the individual tree level. Predictor variables included statistical metrics derived from spectral bands and vegetation indices. Model performance was assessed using an independent ground-truth dataset collected during a dedicated field campaign.
The results show strong agreement between field inventory and remote sensing estimates of tree density (31 vs. 33 individuals ha^-1). Among the tested models, the linear SVM achieved the best performance (r = 0.83), followed by Random Forest and radial SVM. Species mapping accuracy was also satisfactory (r = 0.79), confirming the robustness of the proposed approach for individual tree detection and classification.
By combining remote sensing with field-based data, this study provides a starting point for investigating the spatiotemporal dynamics and keys drivers of F. albida regeneration or decline. The approach offers a scalable framework to support evidence-based decision-making for sustainable agroforestry management and climate change mitigation in dryland regions.

Keywords: Cameroon, CNN, deep learning, machine learning, tree crown segmentation, tree species classification

Contact Address: Faustin Ambomo Tsanga, The University of Maroua, Cifor-Icraf Cameroon, Cirad, Geography, Maroua, Cameroon, e-mail: ambomotsangafyahoo.com