Improving near real time tropical forest change monitoring with multiple data sources

Forest cover loss in the tropics is mainly driven by agriculture and other activities such as mining and timber logging. Tropical countries need reliable and timely measurements of the extent of forest disturbances to prevent and reduce unsustainable and illegal activities. Time series-based forest monitoring at near real time (NRT) has the capacity of detect forest changes once a new satellite image is available. NRT forest multi-sensor monitoring approaches have proven to increase accuracy in tropical forest change detection; although, current methods are still not capable of detecting changes with high spatial accuracy after a few observations. The inclusion of ancillary datasets, e.g. road networks, in combination with satellite time series via machine learning approaches has the potential to provide information about the drivers of forest change and therefore to increase the change detection accuracy. The main objective of this study is to develop such a multi-source approach. A NRT scenario will be simulated in the province of Madre de Dios, Peru. Sentinel 1, Sentinel 2 and Peru’s road network datasets will be combined through the approach developed by Reiche et al. (2018) to detect forest changes. This approach calculates the conditional probability of forest cover change once a new image of the input time series is available. The conditional probability of forest change is computed using Bayesian updating, and forest change events are detected. New satellite observations are used to update the conditional probability of forest change along the time axis, and to confirm or reject forest change events detected previously. Very high resolution images, available through Planet Archive, will be used as guided reference data to collect training and validation data. The TimeSync tool will be used to estimate the temporal accuracy of the proposed method. Both spatial and temporal accuracy will be evaluated for the forest cover loss maps; therefore enabling us to discuss the utility of the data combination pipeline, as well as the importance of the single sources.