An Advanced Dirichlet Prior Network for Out-of-Distribution Detection in Remote Sensing

Remote sensing deals with a plethora of sensors, large number classes/categories, and huge variation in geography. Due to the difficulty collecting labeled data uniformly representing all scenarios, data-hungry deep learning models are often trained source domain that is limited above-mentioned aspects. However, during test/inference phase, such subjected distributional shift, also called out-of-distribution (OOD) samples, form unseen classes, geographic differences, multisensor differences. Deep can behave an unexpected manner when uncertainties. Vulnerability OOD severely reduces reliability trusting on predictions absence any indicator may lead wrong policy decisions or mishaps time-bound remote applications. Motivated by this, this work, we propose Dirichlet prior network-based model quantify uncertainty learning-based models. The approach seeks maximize representation gap between in-domain examples for better segregation samples at test time. Extensive experiments several image classification datasets demonstrate proposed uncertainty. To best our knowledge, first work elaborately study context sensing. codes publicly available https://gitlab.lrz.de/ai4eo/Uncertainty/-/tree/main/DPN-RS .