Comparing the Effects of Temporal Features Derived from Synthetic Time-Series NDVI on Fine Land Cover Classification

Landsat data are an ideal data source for deriving fine-resolution land cover maps, and integrating temporal features extracted from time-series normalized difference vegetation index (NDVI) data achieves better performance. This paper compares the different roles of NDVI statistic features and phenology features in land cover classification at a finer scale. Time-series NDVI with fine resolution is first obtained by fusing Landsat-8 Operational Land Imager and moderate resolution imaging spectrometer (MODIS) NDVI via spatiotemporal fusion algorithm. Statistic and phenology features are then extracted from the fused data and added into random forest (RF) classifier. Performance under different classifiers and importance of phenology features are further discussed. Results show that both NDVI statistic features and phenology features have great effects on improving the classification accuracy after adding them to Landsat spectral bands. The overall accuracy is improved approximately 3% and 5%. Phenology features contain majority information of statistic features, and better reflect the seasonal variations of time-series NDVI, especially for vegetation types. Additionally, neural network classifier achieved similar trends of results with RF but lower accuracy, while support vector machine classifier seems to be poor in dealing with high-dimension temporal features, especially in regions with abundant vegetation. Among phenology features, maximum value, large integrated value, and base value have the highest importance scores, while start, end, and middle times of season provide extra information for identifying grass and nongrass.