The Method Of LULC Classification Base On Random Forest And MODIS Phenology Features

The land use/land cover map obtained by remote sensing technique is indispensable for monitoring, understanding and predicting the complex conditions of local regions, as well as the global human-nature interactions. However, it is still difficult to monitor the conditions of larger regions rapidly and accurately. At present, spatial resolution of many produces only500m and the produces are influenced by many factors, such as classification system, the representation of the features and the property of classification. Therefore, we put forward a modified time-series land cover mapping method to handle these problems. MODIS product data were further processed. The high time-resolution advantage of MODIS was realized, while spatial resolution was not taken as an influence condition. Thus the final classification accuracy was significantly improved. The investigation is based on Hebei Province, and we collected a series of predicted variables from the2010VI product of MODIS250m C5(MOD13Q1). The variables included red and near-red wave band, NDVI, phenological parameters, slope and slope aspect from NVDI time-series data. As for the reference data, we proposed a sampling method at pixel scale on the Google map, which contributes to a more accurate determination and selection of sample. We classified the land cover with a random forest combination classifier, and optimized all the parameters. Meanwhile, the paper optimizes the model, such as: predictive variable、sample sizes and mixed classification. Finally, the overall accuracy of land cover mapping exceeded84.3%with the Kappa coefficient higher than0.79. The user accuracy of rice, forest land and grass reached94.29%,92.23%and85.71%respectively. The overall accuracy was raised by2.6%with the involvement of phenological characteristics, and the accuracy for dry land and building land was increased by6.7%and11.9%respectively. Obviously the mapping method we proposed is accurate, and it can also be applied in the mapping of other regions, as well as in annual national land use/land cover mapping.