Earth Observation Semantic Data Mining: Latent Dirichlet Allocation-Based Approach

Recent advances in remote sensing technology have provided (very) high spatial resolution Earth Observation data with abundant latent semantic information. Conventional processing algorithms are not capable of extracting the information form these and harness their full potential. As a result, discovery methods, based on mining techniques, such as Dirichlet allocation bag visual words models, can discover Despite crucial rule, there only few studies field for applications. This article is focused this shortage. Three different scenarios used to evaluate various applications, including both optical synthetic aperture radar (SAR) resolutions. In first scenario, method correlated perception user machine correct enhance defined Ground Truth map very high-resolution RGB data. The potential evaluated wildfire affected area detection Sentinel-2 second scenario. Finally, third utilized detect misclassifications well patches ambiguous or multiple labels Sentinel-1 SAR patch-based benchmark dataset robustness accuracy annotation dataset. Our results three demonstrated capability data-mining-based methods sensing.