Characterizing tropical forest dynamics by remote-sensing using very high resolution and Sentinel-2 images

Tropical forest is a major Earth ecosystem, in terms of extend, biodiversity and carbon stock. Anthropogenic degradations put this ecosystem under increasing pressure. Some practices such as selective logging or shifting cultivation however, not always lead to net forest cover loss, but rather in a change in forest structure. Vegetation regenerate through different stages, which results in what is called forest dynamics. Observing those dynamics allow a more accurate monitoring. Remote-sensing offers faster and cheaper ways to monitor forests, from local to global scale, than in situ observations. ESA new Sentinel-2 satellites designed for vegetation observation shows promising perspectives for forest monitoring. It is therefore important to investigate its potential in the case of tropical forest dynamics. This study is based on a collaboration with the Environmental Science for Social Change (ESSC) organization in Mindanao, Philippines, and has been the subject of a field campaign in February-March 2018. Forest dynamics at stake in the study area are characterized, and a typology suitable for monitoring is established. Two methods to detect and map the proposed forest classes are developed, one based on a Sentinel-2 images per-pixel classification, the other one based on a very high resolution image (Pléiades) per-object classification. The Sentinel-2 method turns out to be able to discriminate all proposed forest classes, with a 95% classification accuracy. The most discriminating features are the two short-wave infrared (SWIR) image bands followed by the blue band. The very high resolution method shows a confusion between two close classes and an overall model prediction accuracy of 70%. Combining the confused classes into one makes prediction reach 85% accuracy. By comparing both developed methods, the Sentinel-2 per-pixel classification prove to be more accurate, but also to better represent transition areas between classes than the Pleiades method. An Upper Pulangi watershed full landcover is realized as application of the developed method.