Progressive Unsupervised Deep Transfer Learning for Forest Mapping in Satellite Image

Abstract

Automated forest mapping is important to understand our forests that play a key role in ecological system. However, efforts towards forest mapping is impeded by difficulty to collect labeled forest images that show large intraclass variation. Recently unsupervised learning has shown promising capability when exploiting limited labeled data. Motivated by this, we propose a progressive unsupervised deep transfer learning method for forest mapping. The proposed method exploits a pre-trained model that is subsequently fine-tuned over the target forest domain. We propose two different fine-tuning echanism, one works in a totally unsupervised setting by jointly learning the parameters of CNN and the k-means based cluster assignments of the resulting features and the other one works in a semi-supervised setting by exploiting the extracted k-nearest neighbor based pseudo labels. The proposed progressive scheme is evaluated on publicly available EuroSAT dataset using the relevant base model trained on BigEarth-Net labels. The results show that the proposed method greatly improves the forest regions classification accuracy as compared to the unsupervised baseline, nearly approaching the supervised classification approach.