Semantic World – A Novel Benchmark Dataset for Semi-Supervised Semantic Segmentation

Challenge:

Despite the increasing interest in deep learning models for remote sensing applications, large-scale data sets foundational to these models are still scarce. Data sets covering a range of biomes worldwide are of particular interest to train models generalisable across space. In this regard, the Dynamic World data set [1] represents an important milestone giving rise to the eponymous model [2] as well as the ESRI land cover product [3]. With these models global, continuously updatable land cover classifications have been produced for the first time. However, one factor that limits the accuracy and generalisability of the models is the availability of training data, based on a labour-intensive manual labelling process. Upscaling the training of such models requires an extension of the narrow view of supervised learning to semi-
supervised learning, which enables the leveraging the much larger archive of unlabelled satellite scenes.