CaDS Seminar 2024 - May 7
Ankit Patnala (Research Group Earth System Data Exploration)
Exploring Self-Supervised Learning Methods for Landcover Applications using Remote Sensing Data
Abstract:
Land cover is defined as the outer surface of the Earth which is exposed to the atmosphere and mostly comprises grasses, trees, lands, water, etc. The Earth's surface has been subjected to rapid changes in recent years due to anthropogenic and natural factors. Such changes vividly impact humans and livestock in many ways. The term land cover is often associated with land use and land cover(LULC) nowadays. Land use refers to the activities undertaken by people on a certain land cover type, for example, grasslands often provide support to plant, animal, and bird species. The other example is the use of land for industrial purpose that leads to urbanization and migration. There are several other similar reasons why is it important to study and understand the land surface. The land surface plays an important role in many applications that influence on a wide range of ecological, social, and economic process.
The development of self-supervised learning methods has been much explored in the machine learning community recently. It has several advantages, one being there is no need for labeled data thus getting an edge over supervised learning, and the other being the ability to use deep learning models and are efficient on high dimensional such as images making it a better alternative than the traditional unsupervised algorithms. Self-supervised though starts with a huge amount of unlabeled data and using techniques such as transfer learning, fine-tuning as well and zero-shot learning allow it to be used for solving new tasks with fewer labels. The self-supervised learning has greater potential for applications in the field of Earth observation(EO). Satellite missions such as Sentinel2 and Landsat have been active for many years now and are tracking a whole globe in the order of a few days, so there is a huge availability of data, on the other hand as mentioned above, there is a huge cost and effort in annotating. So, self-supervised learning is believed to be beneficial for these applications. Despite being suitable, it does pose different challenges. One of the biggest challenges is that most of the algorithms are developed for natural images and text but those are still quite different to the type of data in this domain. This motivated us to explore different strategies for two different EO applications.