Institute of Photogrammetry and GeoInformation Research Research Groups Photogrammetric Image Analysis Completed projects
Transfer learning for hierarchical Conditional Random Fields for the classification of urban aerial and satellite images (2020)

Transfer learning for hierarchical Conditional Random Fields for the classification of urban aerial and satellite images (2020)

Team:  A. Paul, F. Rottensteiner
Year:  2014
Funding:  DFG
Duration:  since 2014
Is Finished:  yes

Motivation

Aerial and satellite images are an important source of information for the geo-information systems. Semantic information from the remotely sensed data can be derived by the classification. One problem related to the machine learning techniques used in this context is the necessity to provide a sufficient amount of training data for each new classification task or new data set. The acquisition of the data can be tedious and time consuming manual task. Therefore it is desirable to transfer classifier to the new data without any or with just few new training samples.

Goal

The goal of the project is to develop a new methodology based on transfer learning technique called domain adaptation. Domain adaptation techniques try to reduce the amount of training data required for a successful classification by adapting the classifier trained on samples from so called source domain to a new data set called target domain, where the features may have different distributions but assumed to be related.

Our approach for domain adaptation is based on logistic regression considering multiclass problems. Logistic regression is a simple and fast discriminative probabilistic classifier, which should require fewer training samples than a generative approach. The adaptation performs iteratively by removing labelled source samples from the current training data set and simultaneous addition of semi-labelled target samples to the current training data set. The semi-labels are determined automatically applying the current state of the classifier. After each iteration the classifier is retrained using the updated training data set. In this way the classifier and the decision boundaries can be slowly adapted to the distribution of the target domain data.

Currently, individual weights in the domain adaptation process are used, to achieve more stability and overcome problems of abrupt changes in the decision boundaries. These weights modulate the impact on the basis of a sample’s distance from the decision boundary. Additional stability is achieved by using the current state of the classifier for regularisation. An example for classification results applying the domain adaptation is shown in Figure 1(c).

(a)(b)(c)

Fig. 1: Results of classification of the target data: (a) classifier was trained and applied to the target data without domain adaptation (best possible performance using logistic regression); (b) classifier was trained on source domain and applied to target domain without domain adaptation; (c) results after domain adaptation procedure.