Fiche détaillée
| Id | 547 |
| Auteur | Verge Salomé |
| Titre | Mapping Spectrally Similar Vegetation Types in Cloud-Persistent Tropical Landscapes: |
| Sous-titre | A Multi-Sensor Approach to Oil Palm-Forest Discrimination in Sarawak, Borneo |
| Résumé | |
| Summary | Tropical deforestation for agricultural monocultures such as oil palm is a major driver of biodiversity loss, yet accurately mapping this land-cover change remains difficult in cloud-persistent tropical regions where optical satellite imagery is frequently unusable. This study evaluated how multi-sensor remote sensing and derived spectral and structural indices could address the classification challenges posed by persistent cloud cover and spectral similarity between forest and oil palm, using a case study of six oil palm plantations in Sarawak, Malaysian Borneo. Sentinel-1 SAR and Sentinel-2 optical imagery from the 2024 dry season were processed in Google Earth Engine and classified into four land-cover classes (forest, oil palm, infrastructure, water). Six scenarios were compared to isolate the individual and combined contribution of each data source and spectral and structural index. Combining Sentinel-1 data with a GLCM metric and raw Sentinel-2 data achieved the highest mean overall accuracy (90.53%) and the most consistent performance across estates, exceeding both a widely recognised threshold for strong classification accuracy. However, this improvement over Sentinel-2 alone was modest and site-dependent rather than universal, with Sentinel-2 with a GLCM metric alone outperforming the combined stack in half of the six estates tested. Contrary to expectations, neither vegetation indices nor SAR texture metrics reliably improved on their respective raw sensor data, and classifier choice had only a minor influence on accuracy relative to the choice of predictor variables. Misclassification concentrated predictably along forest-oil palm boundaries and around narrow linear features, consistent with a structural limitation of pixel-based classification at 20 m resolution rather than a shortfall specific to any sensor or predictor configuration. This study does not establish a single universally optimal methodology for mapping oil palm and forest in cloud-persistent tropical landscapes; rather, it provides empirical evidence that multi-sensor fusion offers a real but modest, site-dependent benefit, that derived indices should not be assumed to improve classification without empirical testing, and that predictor selection matters more than classifier choice. |
| Volee | 2025 |
| Soutenance | 2026-08-31 |
| Sujet | |
| Url 1 | Mémoire disponible sur demande (admincg) |
| Url 2 | |
| Url 3 |