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

Abstract: Land cover classification faces persistent challenges with inter-investigator variability and salt-and-pepper noise. Although cloud platforms such as Google Earth Engine have made land cover classification more accessible, these issues persist, particularly when multiple investigators contribute to the process. This study developed a robust classification approach that integrates unsupervised clustering of investigator maps with a Bayesian inference framework using Dirichlet distributions. In this study, 44 investigators collected stratified reference samples across four land cover classes using point-based visual interpretation in Saitama City, Japan. We trained three different classifiers, Random Forests (RF), Support Vector Machines (SVM), and Single hidden layer Feed-forward Neural Networks (SFNN), and enhanced the system by implementing unsupervised clustering (k-Means or k-Medoids) to group reliable maps based on entropy characteristics. The Bayesian framework, employing Dirichlet distributions for both likelihood and prior distributions, enables sequential probability updates while preserving probabilistic class assignments. The Bayesian inference from the SVM classification maps achieved the highest mean overall accuracy of 0.857 for Monte Carlo sampling from the referenced JAXA land use land cover map, improving upon the non-Bayesian SVM map (0.855, p < 0.001). Analysis revealed a strong correlation (r=0.710) between investigators' labeling quality and classification accuracy, suggesting that selecting high-quality investigator maps improves the robustness of fusion. The Interspersion and Juxtaposition Index (IJI) showed that fused maps from SVM-based maps selected by k-Means reduced salt-and-pepper noise (IJI: 56.652) compared to baseline maps (IJI: 69.867). Our approach demonstrates an effective approach for combining multiple land cover classifications.