This research examined sub-pixel land-cover classification performance for tree canopy, impervious surface, and cropland in the Laurentian Great Lakes Basin (GLB) using both time-series MODIS (Moderate Resolution Imaging Spectro radiometer) NDVI (Normalized Difference Vegetation Index) and surface reflectance data. Classification training strategies included both an entire-region approach and an ecoregion-stratified approach, using multi-layer perceptron neural network classifiers. Although large variations in classification performances were observed for different ecoregions, the ecoregion-stratified approach did not significantly improve classification accuracies. Sub-pixel classification performances were largely dependent on different types of MODIS input datasets. Overall, the combination of MODIS surface reflectance bands 1-7 generated the best sub-pixel estimations of tree canopy (R2 = 0.57), impervious surface (R2 = 0.63) and cropland (R2 = 0.30), which are considerable higher than those derived using only MODIS-NDVI data (tree canopy R2 = 0.50, impervious surface R2 = 0.51, and cropland R2 = 0.24). Also, sub-pixel classification accuracies were much improved when the results were aggregated from 250 m to 500 m spatial resolution. The use of individual date MODIS images were also examined with the best results being achieved for Julian days 185 (early July), 217 (early August), and 113 (late April) for tree canopy, impervious surface, and cropland, respectively. The results suggested the relative importance of the image data input selection, spatial resolution, and acquisition dates for the sub-pixel mapping of major cover types in the GLB.