Mapping Deforestation and Forest Degradation in Mosaic Landscapes of Southeast Asia using Dense Landsat Time Series
Pflugmacher, Dirk1; Grogan, Kenneth2; Hostert, Patrick1
1Humboldt University Berlin, GERMANY; 2University of Copenhagen, DENMARK

In mainland Southeast Asia, much of the forest is found in complex mosaic landscapes dominated by forest succession of different ages and composition, a result of the shifting cultivation practices of the indigenous people. Quantifying extent and quality of these forests is important for a variety of ecological and socio-economic reasons, but mapping with remote sensing has been challenging as these landscapes are characterized by high temporal dynamics and fine spatial patterns. The challenge is exacerbated by the persistence of cloud cover and atmospheric aerosols in tropical regions. Recently, there has been a promising development towards methods that can utilize uneven and (compared to hyper-temporal sensors) relatively sparse time series of high-resolution multi-spectral imagery from Landsat to detect forest vegetation changes. The complexity of these algorithms ranges from simple threshold models to more complex functional or segmentation based time series models. In this study, we tested and compared several of these time series approaches to map extent and timing of vegetation changes associated with shifting cultivation practises using Landsat time series. As study area we selected a Landsat scene in Northern Laos. The dominant land cover and land-use ranged from intense agriculture, short rotation (< 2 years), medium (3-10), and long rotation (>10 years) fallows to protected primary and secondary natural forest. We constructed a time series using all orthorectified Landsat images with less than 80% cloud cover available from the USGS archive between 1985 and 2012 (148 images). Then, we performed atmospheric correction using the Landsat Ecosystem Disturbance Adaptive Processing Systems, and cloud and cloud-shadow masking using the automated Fmask algorithm. The resulting time series were then spectrally enhanced by using the tasselled cap (TC) transformation and the normalized burn ratio (NBR), and temporally aggregated to annual time steps using different functions such as median and maximum value compositing and compositing based on acquisition day of year. We validated our results using a combination of recent high resolution RapidEye imagery and visual interpretation of Landsat image chips. This study demonstrates the utility of dense time series of high-resolution optical data for tropical forest monitoring. While data availability will likely improve with Landsat LDCM and the upcoming Sentinel-2, historic analyses of forest cover and land-uses will remain relevant.