Segmentation networks reinforced with attribute profiles for large scale land-cover map production

Abstract

Segmentation networks have proven to be popular tools for large scale pixel-wise remote sensing image classification as they can deal with wide spatial areas efficiently, as opposed to convolutional neural networks trained with pixel centered patches. However, they are often criticized in terms of spatial consistency. As such, they have received various extensions through the last few years, in the form of dilated convolutions and skip connections and more. In this paper, we address the same issue by feeding attribute filtered images, that contain inherently a multiscale hierarchical representation of the underlying image, as input to a segmentation network, in an effort to both accelerate convergence and render easier the feature learning task of the bottom layers. We validate our approach through the production of land-use and land-cover maps for a large area of Turkey using Sentinel 2 multispectral images and ground truth from the Copernicus Land Monitoring Service.