CAS: Confidence Assessments of classification algorithms for Semantic segmentation of EO data
GitHub repo of paper
Abstract of paper: Confidence assessments of semantic segmentation algorithms are important. It is a desirable property of models to a priori know if they produce an incorrect output. Evaluations of the confidence assigned to the estimates of models for the task of classification in Earth Observation (EO) are crucial as they can be used to achieve improved semantic segmentation performance and prevent high error rates during inference. The model we develop, Confidence Assessments of classification algorithms for Semantic segmentation (CAS), performs confidence evaluation at the segment and pixel levels, and outputs both labels and confidence. Our model CAS detects the segments with incorrect predicted labels using the proposed combined confidence metric, refines the model, and improves its performance. The outcome of this work has important applications. The main application is the evaluation of EO Foundation Models on semantic segmentation downstream tasks, in particular on land cover classification using satellite Sentinel-2 data. The evaluation shows that the proposed model CAS is effective and outperforms other baseline models.
Authors of paper: Nikolaos Dionelis and Nicolas Longepe
European Space Agency (ESA)
Φ-lab, ESA, ESRIN, Italy