Item

Abstract

Several forest change detection algorithms are available for tracking and quantifying deforestation based on dense Landsat and Sentinel time series satellite data. Only few also capture regrowth after clearing in an accurate and continuous way across a diversity of forest types (including dry and seasonal forests) and are thus suitable to address the need for better information on secondary forest succession and for assessing forest restoration activities. We present a new change detection algorithm that makes use of the flexibility of kernel density estimations to create a forest reference phenology, taking into account all historical phenological variations of the forest rather than smoothing these out by curve fitting. The AVOCADO (Anomaly Vegetation Change Detection) algorithm allows detection of anomalies with a spatially explicit likelihood measure. We demonstrate the flexibility of the algorithm for three contrasting sites using all available Landsat time series data; ranging from tropical rainforest to dry miombo forest ecosystems, with different time series data densities, and characterized by different forest change types (e.g. selective logging, shifting cultivation). We found that the approach produced in general high overall accuracies (> 90%) across these varying conditions, but had lower accuracies in the dry forest site with a slight overestimation of disturbances and regrowth. The latter was due to the similarity of crops in the time series NDMI signal, causing false regrowth detections. In the moist forest site the low producer accuracies in the intact forest and regrowth class was due to its very small area class (most forest disappeared by the nineties). We showed that the algorithm is capable of capturing small-scale (gradual) changes (e.g. selective logging, forest edge logging) and the multiple changes associated to shifting cultivation. The performance of the algorithm has been shown at regional scale, but if larger scale studies are required a representative selection of reference forest types need to be selected carefully. The outputs of the change maps allow the estimation of the spatio-temporal trends in the proportions of intact forest, secondary forest and non-forest - information that is useful for assessing the areas and potential of secondary forests to accumulate carbon and forest restoration targets. The algorithm can be used for disturbance and regrowth monitoring in different ecozones, is user friendly, and open source.