Using the 500 m MODIS land cover product to derive a consistent continental scale 30 m Landsat land cover classification

Classification is a fundamental process in remote sensing used to relate pixel values to land cover classes present on the surface. Over large areas land cover classification is challenging particularly due to the cost and difficulty of collecting representative training data that enable classifiers to be consistent and locally reliable. A novel methodology to classify large volume Landsat data using high quality training data derived from the 500 m MODIS land cover product is demonstrated and used to generate a 30 m land cover classification for all of North America between 20°N and 50°N. Publically available 30 m global monthly Web-enabled Landsat Data (GWELD) products generated from every available Landsat 7 ETM + and Landsat 5 TM image for a three year period, that are defined aligned to the MODIS land products and are consistently pre-processed data (cloud-screened, saturation flagged, atmospherically corrected and normalized to nadir BRDF adjusted reflectance), were classified. The MODIS 500 m land cover product was filtered judiciously, using only good quality pixels that did not change land cover class in 2009, 2010 or 2011, followed by automated selection of spatially corresponding 30 m GWELD temporal metric values, to define a large training data set sampled across North America. The training data were sampled so that the class proportions were the same as the North America MODIS land cover product class proportions and corresponded to 1% of the 500 m and < 0.005% of the 30 m pixels. Thirty nine GWELD temporal metrics for every 30 m North America pixel location were classified using (a) a single random forest, and (b) a locally adaptive method with a random forest classifier derived and applied locally and the classification results spatially mosaicked together. The land cover classification results appeared geographically plausible and at synoptic scale were similar to the MODIS land cover product. Detailed visual inspection revealed that the locally adaptive random forest classifications and associated classification confidences were generally more coherent than the single random forest classification results. The level of agreement between the 30 m classifications and the MODIS land cover product derived training data was assessed by bootstrapping the random forest implementation. The locally adaptive random forest classification had higher overall agreement (95.44%, 0.9443 kappa) than the single random forest classification (93.13%, 0.9195 kappa). The paper concludes with a discussion of future research including the potential for automated global land cover classification.