Overview

This section introduces what we call patch based classification for land cover mapping. This task consists in classifying image patches of fixed size, and attributing the estimated class to the central pixel of the patch, to produce a land cover map. This kind of approach is suited when the available data is sparsely annotated: the only requirement is the location of patches of which the class is known. Typically, examples of terrain truth allowing this approach can be GPS coordinates associated to land cover class, or manually annotated polygons in GIS, etc. This approach involves deep networks that input a patch of image, and produce a single value that represents the class of this patch: hence training this kind of network requires only input image patches, and associated class values. Once the training of a network is done, producing land cover maps consist in applying the model on each patches of a remote sensing image, and storing the value of the estimated class into the output image. The advantage of this kind of approach is its suitability to sparsely annotated datasets. Its implementation subtleties and limitations will be discussed along this section.

First we will see how to collect patches in geospatial images. Then, we will build and train a simple convolutional neural network with otbtf and keras. Once this model is trained, we will apply it quickly on a remote sensing image to generate a landcover map and open the result in QGIS. After that, we will learn how to customize the training process, e.g. how to select the best model, how to add metrics and use tensorboard to monitor them. Then, we will consider an important subset of convolutional neural networks: the fully convolutional neural networks.