• Citer

The importance of the sampling design in mapping woody cover in arid ecosystems

par Ramalason, Felana Nantenaina;Rakotondrasoa, Olivia Lovanirina;Vander Linden, Arthur;Renard, Guillaume;Randriamalala, Josoa Ramarolanonana;Vereecken, Nicolas ;Razakamiaramanana, Aina ;Ranaivoharivelo, Mbolatiana;Raholinarivo, Sismondi;Rakoto Ratsimba, Harifidy;Bogaert, Jan;Bastin, Jean-François
Référence GIScience and remote sensing, 63, 1
Publication Publié, 2026-04-13

Article révisé par les pairs

• Accès en ligne
• Détails
• Contenu
• Statistiques

Résumé :

Accurate calibration data remains a major constraint in ecological remote sensing, particularly in arid ecosystems where sparse and heterogeneous woody vegetation is challenging to detect. Despite advances in sensor technology and modelling algorithms, optimal calibration sampling design has received limited attention, with current approaches varying unsystematically (35–1,000 plots/1,000 km²) without empirical benchmarks. Here, we develop and evaluate a quantitative framework for optimizing sampling strategies in remote-sensing models of woody cover through systematic comparison of (i) calibration data source (field surveys versus photointerpretation), (ii) spatial configuration (clustered versus dispersed), and (iii) sampling density. This integrated approach enables isolating the relative contributions of each design component to mapping accuracy—a critical gap in current remote sensing methodology. We applied this framework to Sentinel-1, Sentinel-2, combined Sentinel-1 + 2, and AlphaEarth Foundations across Madagascar's arid southwest, validating predictions through spatial cross-validation using independent field plots. Photointerpretation substantially outperformed field-based calibration under typical arid-zone constraints (R² = 0.88, RMSE = 0.11 versus R² = 0.46–0.66, RMSE = 0.17–0.21). Performance saturated at 20.7–41.4 dispersed calibration plots per 1,000 km² across all predictors, beyond which gains became marginal. Dispersed strategies required half as many samples as clustered designs to achieve comparable accuracy, demonstrating that spatial distribution outweighs sample size. Once adequate sampling density and distribution were achieved, Sentinel-1 + 2 and AlphaEarth performed similarly (R² ≈ 0.85–0.86), indicating that sampling design outweighs predictor complexity. Our framework provides empirically derived operational thresholds (minimum: 15.5 plots/1,000 km²; optimal: 20.7–41.4 plots/1,000 km²) and a transferable methodology for determining optimal calibration densities in heterogeneous ecosystems where logistical constraints limit field sampling.