Using spatial context to improve early detection of deforestation from Landsat time series

Abstract

Mapping deforestation using medium spatial resolution satellite data (e.g. Landsat) is increasingly shifting from decadal and annual scales to sub-annual scales in recent years, but this shift has brought new challenges on how to account for seasonality in the satellite data when detecting deforestation. A seasonal model is typically used to account for seasonality, but fitting a seasonal model is difficult when there are not enough data in the time series. Here, we propose a new approach that reduces seasonality in satellite image time series using spatial context. With this spatial context approach, each pixel value in the image is spatially normalised using the median value calculated from neighbouring pixels whose pixel values are above the 90th percentile. Using Landsat data, we compared our spatial context approach to a seasonal model approach at a humid tropical forest in Brazil and a dry tropical forest with strong seasonality in Bolivia. After reducing seasonal variations in Landsat data, we detected deforestation from the same data using the Breaks For Additive Season and Trend (BFAST) method. We show that, in dry tropical forest, deforestation events are detected much earlier when the spatial context approach is used to reduce seasonal variations in Landsat data than when a seasonal model is used. In the dry tropical forest, the median temporal detection delay for deforestation from the spatial context approach was two observations, seven times shorter than the median temporal detection delay from the seasonal model approach (15 observations). In the humid tropical forest, the difference in the temporal detection delay between the spatial context and seasonal model approach was not significant. The differences in overall spatial accuracy between the spatial context and seasonal model were also not significant in both dry and humid tropical forests. The main benefit for using spatial context is early detection of deforestation events in forests with strong seasonality. Therefore, the spatial context approach we propose here provides opportunity to monitor deforestation events in dry tropical forests at sub-annual scales using Landsat data.