In land cover mapping at a high spatial resolution, pixel values alone are not always sufficient to recognize the more complex classes. Contextual features (computed with a sliding kernel or other kind of spatial support) can be discriminating for certain land cover classes, for example, different levels of urban density, or classes containing heterogeneous pixels, such as orchards and vineyards. However, the reference data used for training the supervised classifier are almost always sparsely labeled, in other words, not every pixel of the training area is labeled. This makes the selection of an appropriate contextual classification method for land cover mapping problematic. Indeed, the current state-of-the art contextual classification model, the Deep Convolutional Neural Network (D-CNN), encounters issues when the geometry of the desired output is absent from the training set. Data-driven methods like D-CNN rely heavily on the availability of extensive training labels to learn both the feature extraction and classification steps. With a sparse training set, sharp corners are rounded, and thin elongated elements may be either thickened, or entirely lost. Alternatively, there are several methods based on the manual selection of contextual features in a chosen neighborhood, guided by the knowledge of the data and past experience from similar problems. Such approaches should not be as sensitive to sparsely labeled data, as they do not rely on any training data for feature extraction. This paper presents a new process for including contextual information in an image classification scheme: the Histogram Of Auto Context Classes in Superpixels (HACCS), which involves classifying an image using the local class histograms as contextual features. These histograms are calculated within superpixels of different sizes in order to provide a multi-scale characterization of the neighborhood, while preserving the geometry of the image objects. This method is evaluated on two data sets presenting different spatial, temporal, and spectral resolutions, and each case is compared with a D-CNN in terms of class accuracy, but also of the quality of the geometry in the produced map. Experiments on the Sentinel-2 time series show that HACCS provides equivalent thematic accuracy compared to the D-CNN, while exhibiting a higher degree of geometric accuracy. On very high spatial resolution imagery (SPOT-7), the D-CNN provides significantly stronger thematic accuracy, but this comes at the cost of a lower level of geometric accuracy.
This is an open access article distributed under the Creative Commons Attribution License
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited