ASSESSING FOREST CHANGE DETECTION IN TROPICAL SEASONAL BIOMES THROUGH LANDTRENDR ALGORITHM

There is no optimal disturbance detection method without limitations that can be applied to all ecosystems. Different vegetation structures present different responses to seasonal variations inducing disturbance errors along geographical regions. This study applied the LandTrendr algorithm, developed in GEE platform, in order to examine its accuracy in different tropical seasonal biomes: Savanna and Atlantic Forest. LandTrendr was run by a default parameter configuration and compared to a large-scale reference dataset. In general, the Atlantic Forest presented higher accuracies than Savanna in the disturbance. Both biomes presented low producer’s accuracy for some years but high user’s accuracy for almost all years. These results are explained due to change detection in high seasonal areas is affected by seasonal modifications in spectral signature due to phenology, leading to misclassification of spectral changes as having human-induced changes. Future researches can also follow-up this approach by exploring different disturbance detection algorithms and parameters simulations as well as the implementation of stratification by vegetation class in order to reach higher accuracies, reducing the effects of land cover type on climate fluctuations.