An Assembly Method to Enhance the Classification of Satellite Imagery In Google Earth Engine
Abstract: Remote sensing has been an enormous aid in monitoring the changes of land cover and land use over time. This study was aimed to compare the performance of different classification models for Landsat imagery and to develop a new ensemble method to improve classification performance at the regional scale. Using the Corpus Christi region in south Texas as the study area, the comparative study involved four widely-used classification models: CART, minimum distance, random forest (RF), and naive Bayes. Using 1-m NAIP imagery, we generated a total of 2,400 sample points over the period from 2006 to 2016 at two-year steps, representing four major land cover categories: water, urban, vegetation, and bare land. The samples were randomly partitioned into two training and validation sets. Results show that individual classification models performed differently in certain regions and for certain land features. For instance, bare land was often misclassified as developed land in suburban areas while some vegetated pixels were confused with water. This misclassification could be masked by overall high training accuracies, such as 100% and 90.2% for CART and RF, respectively. However, naive Bayes, with a training accuracy of 79.4%, was able to outperform other models in some of those areas. We developed a post-analysis ensemble approach: first stacking the initial classification results of individual classifiers as an image collection and then performing a pixel-wise analysis as to determine if there is an agreement between the classifiers regarding the class for that particular pixel. Results shown that the ensemble approach was able to effectively reduce commission and omission errors, and maximize the classification results. All efforts of ground truth collection and image analysis were conducted in Google Earth Engine, a state-of-the-art geospatial analysis platform. In doing so, the proposed ensemble method can be readily applied to other regions around the globe.
Judging Forms – Official judges only