Abstract
Abstract. This paper describes a methodology to produce a 7-classes land cover map of urban areas from very high resolution images and limited noisy labeled data. The objective is to make a segmentation map of a large area (a french department) with the following classes: asphalt, bare soil, building, grassland, mineral material (permeable artificialized areas), forest and water from 20cm aerial images and Digital Height Model.We created a training dataset on a few areas of interest aggregating databases, semi-automatic classification, and manual annotation to get a complete ground truth in each class.A comparative study of different encoder-decoder architectures (U-Net, U-Net with Resnet encoders, Deeplab v3+) is presented with different loss functions.The final product is a highly valuable land cover map computed from model predictions stitched together, binarized, and refined before vectorization.
Highlights
Land cover maps are cartographic products widely demanded by land managers
We propose to reuse top-class semantic segmentation architectures (U-Net with ResNet encoders, DeeplabV3+) on very high resolution aerial images
A Digital Surface Model - derived from the aerial images by photogrammetric techniques - is used in combination with a Digital Terrain Model to produce a Digital Height Model that we concatenate as a 5th channel to input images
Summary
Land cover maps are cartographic products widely demanded by land managers. As they provide a quantification of land resources into thematic categories (e.g. forest, water or asphalt surface), land use/land cover maps are used to measure current conditions and how they are changing. The different classes composing the final land cover map are derived from OCSGE product (Description of OCSGE). They have been selected to best fit the description of urban areas: asphalt, bare soil, building, grassland, mineral material, forest, and water. We propose a full methodology to produce a very high resolution 7-classes land cover map at a large scale with aerial images. Semantic segmentation of aerial images with convolutional neural networks is performed from this training dataset. We propose to reuse top-class semantic segmentation architectures (U-Net with ResNet encoders, DeeplabV3+) on very high resolution aerial images. This article describes some post-processing techniques to convert model predictions to a relevant land cover map
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
