INTER-REGION TRANSFER LEARNING FOR LAND USE LAND COVER CLASSIFICATION

Land use land cover (LULC) classification is an essential task in Earth Observation (EO) as it helps in monitoring long-term developments, detecting changes and analysing their environmental impacts. Due to advancements in remote sensing, there is an abundance of open data available but annotating this data is expensive. As a result, many research works in EO create a labelled dataset for one selected region and perform a corresponding regional analysis. By employing transfer learning, we can reuse these labelled datasets for different regions and thereby minimize the manual annotation costs. However, there are some open questions: to what extent can the features learned in one region be transferred to another? Does a larger pre-training dataset mean better transfer learning performance? How can we estimate the transfer learning performance? To answer these questions, we divide a large EO dataset called BigEarthNet into sub-datasets by region and perform region to region transfer learning. We find that the models trained on one region do not perform well on another region. We applied transfer learning techniques and showed that the class imbalance can hinder learning. If the source region has additional classes which are dominant in the source region or has fewer images for the classes dominant in the target region, transfer learning can have negative impacts on the model performance in the target region. We also demonstrate the use of chi-squared distance in selecting an appropriate source region for transfer learning.