Deep feature extraction in an object-based image analysis context

Since only a few attempts exist to extract deep features in OBIA, this work aims at developing novel deep learning encoder for satellite image objects.

​

It is composed of two main objectives: (i) comparing existing techniques, both traditional and deep extraction, on a downstream task, and (ii) proposing a new feature encoder taking account spectral, textural and geometric information of the objects.

Using ISPRS high resolution aerial dataset, the deep learning pipeline is as follows: extracting objects from satellite images thanks to a deterministic segmentation algorithm (like SLIC or Felzenszwalb), implementing different

feature extraction methods, and plug a simple existing neural network for the downstream task.

​

Second, we will develop a new method to extract spectral, textural and geometric information. The use of graph neural networks  is the most appropriate method thanks to their ability to handle non-square data (the objects are a grid of pixels of arbitrary shape). With an adapted architecture, we should be able to exploit the most of the object information and so, surpass handcrafted features performances without expert knowledge.