Comparative machine learning for forest landscape mapping in Mount Argapura, East Java, Indonesia

Le résumé fourni par la source

Abstract. Wulandari PW, Hernawan E, Sumarga E, Sulistyawati E. 2026. Comparative machine learning for forest landscape mapping in Mount Argapura, East Java, Indonesia. Asian J For 10 (1): r100122. https://doi.org/10.13057/asianjfor/r100122. Monitoring land cover is essential for understanding forest landscape structure and supporting sustainable management in mountainous regions. This study mapped land cover in the Mount Argapura landscape, East Java, Indonesia, using 2024 Landsat 8/9 imagery combined with topographic variables and topographic and texture metrics. Seven land-cover classes were identified: Natural Forest, Plantation Forest, Plantation Estate, Dryland Agriculture, Open Land, Settlements, and Water Bodies. Three machine-learning classifiers Classification and Regression Tree (CART), Random Forest (RF), and Support Vector Machine (SVM) were evaluated using Overall Accuracy (OA), Kappa (κ), 95% Confidence Intervals (CI), and McNemar’s test. Results indicate that RF achieved the highest performance (OA=0.98; κ=0.97), significantly outperforming CART (OA=0.97; κ=0.96) and SVM (OA=0.87; κ=0.82). At the class level, RF showed high stability (User’s Accuracy 0.98 for forest classes), while SVM struggled with spectral confusion between plantation and Dryland Agriculture. Spatial analysis reveals a landscape dominated by managed forest systems, with Plantation Forest covering 36.3% (35,544.33 ha) and Natural Forest covering 32.7% (32,043.51 ha). Together with Plantation Estates (22.1%), these vegetated classes account for over 91% of the 97,923.33 ha study area. Comparison with the official Indonesian Ministry of Environment and Forestry (MoEF) map showed moderate agreement (OA=0.54; κ=0.30), with this study identifying a larger extent of managed plantation landscapes (35,544.33 ha vs. 12,053 ha in MoEF). This discrepancy highlights the importance of local-scale mapping for capturing site-specific details. The findings provide a robust baseline for BKSDA and Perum Perhutani in biodiversity conservation, carbon stock assessment, and agroforestry planning. Future research should incorporate multi-temporal data to better detect fine-scale forest dynamics in this complex tropical terrain.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.