A Bayesian Inference Approach for Reducing Inter-Investigator Variability in Sampling-Based Land Cover Classification (2403.15720v2)

Abstract: Land cover classification plays a critical role in describing Earth's surface characteristics. However, these classifications can be affected by uncertainties introduced by variability in investigator interpretations. While land cover classification mapping is becoming easier due to the emergence of cloud geospatial platforms, such uncertainty is often overlooked. This study aimed to create a robust land cover classification map by reducing inter-investigator variability from independent investigators' maps using a Bayesian inference framework. In Saitama City, Japan, as a case study, 44 investigators used a point-based visual interpretation method via Google Earth Engine to collect stratified reference samples across four different land cover classes: Forest, Agriculture, Urban, and Water. These samples were then used to train a random forest classifier, resulting in the creation of individual classification maps. We quantified pixel-level discrepancies in these maps arising from inherent inter-investigator variability. To address them, we developed a Bayesian inference framework to produce a unified classification map. This framework updates the classification probability based on a Dirichlet distribution and yielded an overall accuracy of 0.851 for independent validation samples, an improvement over the average accuracy of 0.728 for the individual maps. We further improved the results by introducing K-Medoids to group the most reliable maps as the input for Bayesian inference, achieving an overall accuracy of 0.858, the highest among all approaches tested. Our approach also effectively reduced salt-and-pepper noise, which is often found in individual classification maps. This research underscores the intrinsic uncertainties present in land cover classification maps attributable to investigator variations and introduces a potential solution to mitigate these variations.