Sample database for DL-based dwelling extraction

Suggested / supervised by: Stefan Lang, Dirk Tiede, Yunya Gao

Background:
The Christian Doppler Laboratory on Earth Observation for Humanitarian Action – GEOHUM – seeks to advance technologies at the interface of EO*GI and AI (artificial intelligence). The applications range from mission planning and operations in crisis intervention to population estimation for food distribution or vaccination campaigns. Our vision is to enhance technical and organisational capacities matching specific needs from humanitarian organisations, in particular our partner MSF (Doctors Without Borders). Envisaged outcomes are a fundamental scientific substantiation and the development and innovative use of relevant information products to optimize aid delivery in conflict and humanitarian disaster situations.

With promising results obtained by deep learning (DL), the notion arises that DL is not agnostic to input errors or biases introduced, in particular in sample-scarce situations. Within the CDL, we seek to understand the influence of different sample quality aspects for creating a sample database: (1) inherited properties (quality parameters of the underlying image such as cloud cover, seasonality, etc.); (2) individual (i.e., per-sample) properties, i.e. data lineage and provenance, (3) reliability of the labelling (classification).

This master thesis explores the influence of samples collected from different camp settings were hand-selected and annotated with computed features in an initial stage. The supervised annotation routine shall be automated in a way that thousands of existing samples can be labelled with this extended feature set. This should help better condition the subsequent DL tasks in a hybrid AI approach.

Suggested reading:

• Lang, S., L. Wendt, D. Tiede, Y. Gao, V. Streifeneder, H. Zafar, A. Adebayo, G. Schwendemann and P. Jeremias (2021). "Multi-feature sample database for enhancing deep learning tasks in operational humanitarian applications." GI Forum - Journal for Geographic Information Science 9(1): 209-219. 

Start:
1 Oct 2021

Prerequisites/qualification:
remote sensing
image classification
machine / deep learning