Mapping Annual Land Use and Land Cover Changes Using MODIS Time Series

Mapping land use and land cover change (LULCC) over large areas at regular time intervals is a key requisite to improve our understanding of dynamic land systems. In this study, we developed and tested an automated approach for mapping LULCCs at annual time intervals using data from the Moderate Resolution Imaging Spectroradiometer (MODIS). Our approach characterizes changes between land cover types based on annual time series of per-pixel land cover probabilities. We used the temporal segmentation algorithm MODTrendr to identify trends and changes in the probability time series that were associated with land cover/use conversions. Accuracy assessment revealed good performance of our approach (overall accuracy of 92.0%). The method detected conversions from forest to grassland with a user's accuracy of 94.0 +/- 2.0% and a producer's accuracy of 95.6 +/- 1.6%. Conversions between cropland and grassland were detected with a user's and a producer's accuracy of 65.8 +/- 4.8% and 72.2 +/- 9.2%, respectively. We here present for the first time an approach that combines probabilities derived from machine learning (random forest classification) with time-series-based analysis (MODTrendr) for land cover/use change analysis at MODIS scale.