Inductive transfer-learning for the classification of aerial and satellite images using Bayesian methods

• Applicant: Professor Dr.-Ing. Jörn Ostermann
• Subject Area: Geodesy, Photogrammetry, Remote Sensing, Geoinformatics, Cartography
• Term: from 2013 to 2016
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 246374192

High-resolution aerial- and satellite imagery are the most important sources of spatial information in remote sensing. Classification of these images using supervised learning algorithms is an essential procedure for image processing and interpretation. This requires a training-sample which must be carefully prepared in a laborious process. One paradigm that may help in reducing manual work during this process is called transfer-learning. A robust transfer-model allows us to reuse training-samples from similar data-sets that are related to our target image in either space (in close vicinity) or time (multi-temporal images). The critical issue is to automatically determine when this transfer will actually improve classification results.The aim of this project is the development and evaluation of such a transfer-model. The parameters of this model will be estimated from all available training-samples using a Bayesian estimator, which also requires a carefully devised prior-probability. Bayesian methods have already proved themselves on a wide range of tasks. Mainly, because they allow robust estimation even when only a minimal amount of data is available and secondly, because they allow methodical modelling of uncertainty and unknown parameters. The estimator will be implemented using modern simulation-based methods, such as Marko-Chain-Monte-Carlo (MCMC).Among the increased robustness compared to existing methods, the main innovations of this project can be summarized as follows: (1) Assessing the usefulness for knowledge transfer of a training-sample is a critical issue. Existing methods are often too optimistic (thus causing negative transfer) or require a large training sample from the target image. Our transfer-model will extend current notions of transferability by a new criteria. This will make our model more robust and more conservative when the available data allows no conclusive decision. (2) Our system combines the instance-transfer-strategy and the feature-representation-transfer-strategy in a unified transfer-model. (3) Existing transfer-methods are usually deeply integrated into a specific classification-method. On the other hand our transfer-model can be utilized with basically any classification-method through a thin abstract interface.A research project of Professor Heipke (Institut für Photogrammetrie und GeoInformation, Leibniz Universität Hannover) will cover a different aspect on the topic of transfer-learning. As part of a research cooperation we plan to share data and results.

• DFG Programme: Research Grants
• Participating Person: Professor Dr.-Ing. Christian Heipke