Selection of Unlabeled Source Domains for Domain Adaptation in Remote Sensing

Abstract

In the context of supervised learning techniques, it can be desirable to utilize existing prior knowledge from a source domain to estimate a target variable in a target domain by exploiting the concept of domain adaptation. This is done to alleviate the costly compilation of prior knowledge, i.e., training data. Here, our goal is to select a single source domain for domain adaptation from multiple potentially helpful but unlabeled source domains. The training data is solely obtained for a source domain if it was identified as being relevant for estimating the target variable in the corresponding target domain by a selection mechanism. From a methodological point of view, we propose unsupervised source selection by voting from (an ensemble of) similarity metrics that follow aligned marginal distributions regarding image features of source and target domains. Thereby, we also propose an unsupervised pruning heuristic to solely include robust similarity metrics in an ensemble voting scheme. We provide an evaluation of the methods by learning models from training data sets created with Level-of-Detail-1 building models and regress built-up density and height on Sentinel-2 satellite imagery. To evaluate the domain adaptation capability, we learn and apply models interchangeably for the four largest cities in Germany. Experimental results underline the capability of the methods to obtain more frequently higher accuracy levels with an improvement of up to almost 10 percentage points regarding the most robust selection mechanisms compared to random source-target domain selections.