Improved Land Cover Mapping using Random Forests Combined with Landsat Thematic Mapper Imagery and Ancillary Geographic Data

Large area land-cover mapping involving large volumes of data is becoming more common in remote sensing applications. Thus, there is a pressing need for increased automation in the land-cover mapping process. The main objective of this research was to compare the performance of three machine learning algorithms (MLAS) for mapping wetlands in the Sanjiang Plain combined Landsat TM imagery with ancillary geographical data. Three MLAS included random forest (RF), classification and regression tree (CART), and maximum likelihood classification (MLC). Comparisons were based on several criteria: overall accuracy, sensitivity to data set size, and noise. Our results indicated that first, the random forest and CART approach can achieve substantial improvements in accuracy over the traditional MLC method. Random forest produced the highest overall accuracy (91.3 percent) the kappa coefficient 0.8943, with marsh class accuracies ranging from 77.4 percent to 90.0 percent. Secondly, the random forest method was least sensitive to reduction in training sample size, and it was most resistant to the presence of noise compared to CART and MLC. The comparison between three MLAS revealed that the random forest approach was most resistant to training data deficiencies while improved land-cover map accuracy in marsh area.