Last week EarthPulse participated in two events: The Space & AI Conference and the AI & Big Data Congress. In both events the current state of DL4EO was presented, the most pressing limitations were discussed and future developments were proposed in order to overcome the current barriers. We are currently living through a Deep Learning revolution thanks to the increase in data access, computing and open source libraries. Earth Observation is no different. When we talk about Deep Learning for Earth Observation (DL4EO) we are mainly talking about Computer Vision with Convolutional Neural Networks. Other applications of DL4EO exist, but these are the most common and successful applications nowadays.

The main limitation that we face in the field of DL4EO is the lack of big, quality, labeled datasets (necessary to train neural networks effectively). This is mainly due to the fact that data access can be expensive (depending on the data source) and manually labeling images can be very time consuming and require human expertise. Nevertheless, a lot of problems can be solved today with this technology.

For the past months we have been working, along SpaceKnow and ESA, in the AI-Pathfinder project with the objective of reviewing the current state of DL4EO, find its limitations and provide development lines to advance the field. Our main findings are summarized in the following table:

We are working hard on all these fronts with the objective of building a true Digital Twin Earth, where relevant information can be extracted from EO data without the associated complexity. You can learn more exploring our technology.

Learn more about the conferences in the following links.