Characterizing land cover/land use from multiple years of Landsat and MODIS time series: A novel approach using land surface phenology modeling and random forest classifier

Over the last 20 years, substantial amounts of grassland have been converted to other land uses in the Northern Great Plains. Most of land cover/land use (LCLU) assessments in this region have been based on the U.S. Department of Agriculture - Cropland Data Layer (USDA - CDL), which may be inconsistent. Here, we demonstrate an approach to map land cover utilizing multi-temporal Earth Observation data from Landsat and MODIS. We first built an annual time series of accumulated growing degree-days (AGDD) from MODIS 8-day composites of land surface temperatures. Using the Enhanced Vegetation Index (EVI) derived from Landsat Collection 1's surface reflectance, we then fit at each pixel a downward convex quadratic model to each year's progression of AGDD (i.e., EVI = alpha + beta x AGDD - gamma x AGDD(2)). Phenological metrics derived from fitted model and the goodness of fit then are submitted to a random forest classifier (RFC) to characterize LCLU for four sample counties in South Dakota in three years (2006, 2012, 2014) when reference point datasets are available for training and validation. To examine the sensitivity of the RFC to sample size and design, we performed classifications under different sample selection scenarios. The results indicate that our proposed method accurately mapped major crops in the study area but showed limited accuracy for non-vegetated land covers. Although all RFC models exhibit high accuracy, estimated land cover areas from alternative models could vary widely, suggesting the need for a careful examination of model stability in any future land cover supervised classification study. Among all sampling designs, the "same distribution" models (proportional distribution of the sample is like proportional distribution of the population) tend to yield best land cover prediction. RFC used only the most eight important variables (e.g., three fitted parameter coefficients [alpha, beta, and gamma]; maximum modeled EVI; AGDD at maximum modeled EVI; the number of observations used to fit CxQ model; and the number of valid observations) have slightly higher accuracy compared to those using all variables. By summarizing annual image time series through land surface phenology modeling, LCLU classification can embrace both seasonality and interannual variability, thereby increasing the accuracy of LCLU change detection.