Deep Learning for Remote Sensing Images with Open Source Software

9

 

Discussion

 

In this section, we have introduced a simple approach to produce land cover maps from remote sensing images using deep nets. The patch-based approach, consisting of estimating the class value of one patch of image(s), is well suited to a sparsely annotated dataset. Architectures involved in patch-based classification generally input one or multiple patches of images, possibly from different modalities (as we saw in section 8.2), and derive a single-valued output corresponding to the estimated class of the input content. This kind of architecture can be trained with single-valued class samples. For instance, terrain truth data, like a set of GPS coordinates associated with land cover classes, suffices to train this kind of ...