Land Cover/use Classification Optimization Model (LC-COM): new fusion model by considering spatial heterogeneity

The Land use/Cover Classification Optimization Model (LC-COM) is designed to integrate the strengths of the classification results from multiple classifiers and existing products. In LC-COM, the reconciliation index was developed to align the existing LULC products with the composite approach of Landsat images to be classified. Training samples were then auto-generated from these LC products and refined by the spectral indices to further match the selected Landsat images. Six classifiers provided by the Google Earth Engine platform were applied to make their classification to fully explore the detailed and specific information from the Landsat images. The results of these classifiers with the five LULC products were then integrated into an accuracy-weighted hybrid map by using producer accuracy, user accuracy and the especially designed index of matching accuracy reflecting spatial heterogeneity. The results show that the optimized land-cover classification after fusion effectively improved the overall accuracy by integrating all the strengths from each individual result, and the classification performance could be significantly improved when spatial heterogeneity considered.