Enriching satellite image annotations of forests with keyphrases from a specialized corpus

The automatic annotation of changes in satellite images requires examples of appropriate annotations. Alternatively, keyphrases extracted from a specialized corpus can serve as candidates for image annotation models. In the case of detecting deforestation in satellite images, there is a rich scientific literature available on the topic that may serve as a corpus for finding candidate annotations. We propose a method that utilizes a deep learning technique for change detection and visual semantic embedding. This method is combined with an information retrieval framework to find annotations for pairs of satellite images showing forest changes. Our evaluation is based on a dataset of image pairs from the Amazon rainforest and shows that keyphrases provide richer semantic information without any negative impact on the annotation compared to annotating with single words.