Spatially Precise Contextual Features Based on Superpixel Neighborhoods for Land Cover Mapping with High Resolution Satellite Image Time Series

High resolution image time series as those provided by Sentinel-2 allow to target semantically rich nomenclatures for land cover mapping. However, at 10 m resolution, pixel based classification fails to correctly identify some classes for which pixel context is discriminative. Recent advances in deep convolutional neural networks show promising results to tackle this problem, but the lack of complete annotation over large areas, the computational cost and the dimensionality of the feature space (much larger than those used in computer vision) does not allow to use these approaches in operational mapping applications yet. Contextual information can be calculated by applying a fixed-size neighborhood filter, but this can cause the loss of linear objects and the rounding of sharp corners. In Object Based Image Analysis, segmentation is used to extract objects for calculating contextual features while maintaining the high-frequency elements in the image. However, these do not necessarily include spectrally diverse pixels in a neighborhood, which can be relevant for characterizing the context. Superpixels place themselves in between the fixed-neighborhood and the object-based methods, in that they include spectrally diverse pixels in the same segment by imposing size and compacity constraints, while remaining adaptive to the natural boundaries in the image. This study assesses and compares the ability of these three types of neighborhood to improve classification performance on context-dependent classes, in a high-resolution Sentinel-2 time series land cover mapping problem.