Why are we not building one platform to rule them all?

At the current stage of development of the Earth Observation field we believe we can contribute more by building a service-based solution and open-sourcing machine learning libraries rather than providing yet another platform. It will serve better to our users and ensure faster advancement, dissemination and adaptation of remote sensing know-how.

Earth observation field is going through interesting times; after existing for decades it seems it started to change practically over night. We’ve seen such events in science already. For example biology, being around for centuries, giving us an understanding of evolution, medicine and many other useful things, was for a long time limited to qualitative research, confined to observing the world, finding extremely intriguing things, but mostly searching for clear “Yes or No” answers. But at some point this was no longer enough as quantitative analysis became essential to get an answer. This is where physicists got interested in it, launching biophysics, and things started to change fast.

I feel that Earth Observation is going through a similar phase now. For decades remote sensing scientists were analyzing satellite imagery. I imagine it was quite a challenge to understand multi-spectral aspects of EO, atmospheric distortions in the data and similar issues. However, results were mostly maps, pictures and stories, a chart here and there. Very few things were deterministic. For example, how did we come with the “fact” that vegetation has NDVI from 0.3–0.8, while soil has 0.1–0.2? There were various indices created, NDWI, EVI, NBR and others, but they are all plagued with similar issues as NDVI — interpretation of index values, impact of atmospheric conditions to these, etc.. Strangely enough it is sometimes difficult to even get a clear formula for a specific index as even these are often described, in scientific papers, in words alone. This is why we created the “Custom Scripts repository” — to have one place where this knowledge is described in a deterministic manner. Still, we cannot solve the “legend” issue as we get asked many times “what does a moisture index of 0.6 stand for”. These limitations resulted in most of the scientific results being overly fine-tuned for a specific case, making it very difficult to produce regional or global solutions.

All of this has been changing in the recent years. It seems to coincide with the Sentinel-2 mission and Copernicus starting to freely distribute good-quality data globally every few days, which made it interesting to a wider group of people. These are bringing missing “tools” to the remote sensing field — statistics and related neural networks, combined with accessible powerful cloud infrastructure and loads of data.

Automatic classification based on machine learning is a typical example where quantitative analysis is essential

I believe we are at this cross-road now. It’s fun being part of the infancy of the “new remote sensing” — the possibility to extract quantitative information from satellite imagery, almost like querying some virtual database. However, we all have to be honest to ourselves: we are still learning and the existing results are pretty constrained. I am confident that we can do much more than count cars on the parking spots, planes and ships and oil tanks. In this phase we will (all) get best results if we cooperate as much as possible. This is why we have designed our approach to EO machine learning, eo-learn, as an open-source library — we hope that other users will contribute to it or at least help us test it thoroughly. It will make progress faster.

Both eo-learn and Sentinel Hub are designed as a set of services, not as platforms. There is a reason for that as well. Creating a platform, which everyone is using, for sure sounds sexy. Unfortunately, it is also pretty limiting. If someone requires a function that is not available in a platform, they will simply go to whichever platform that has that functionality. There is also a matter of an IPR. People who have a really good idea on what to do with EO data are not too fond of putting their know-how into someone else’s platform, especially if this someone could end up becoming their competitor. This is why we believe that a service-based approach is the right thing. There is a bit more work to accomplish it (we try to minimize this by preparing various integration packages), but whoever plans to generate revenue with their solution is willing to invest this effort. And they have all the freedom to design their solution however they like, using our services for Sentinel data, some 3rd party service for weather data, eo-learn libraries, their own tools and another plethora of small pieces to make the best possible solution.

It was not always like this. When we started with Sentinel Hub, we were focusing on client-based tools to process the data. But we realized that there are not many people who are willing to pay for such a service as most of them were able to do it themselves, albeit taking quite a bit of their time. It was difficult to build a business model on top of it. Therefore, when creating EO Browser we ditched business approach and provided it free of charge. We might have lost some potential revenues, but we gained much more. EO Browser is not just a marketing tool, it is actually raising awareness about Sentinel data throughout the world, generating demand for it. Some of this demand comes back to us in terms of Sentinel Hub services subscriptions. Due to EO Browser’s open-source policy it also represents a good integration example and there are quite a few of our users, who decided to build commercial applications on top of it.

By releasing our products under an open source license we aim to gain in innovation, transparency, scalability, flexibility, and security through collaboration with the community, including existing EO groups dealing with similar problems as well as newcomers to the EO that will bring their fresh and unique views from their domains of expertise. We can see this in real-life, for example Pierre Markuse’s wild-fire script, which is now used by many people in the community and seen on several news platforms.

Or Harel Dan’s contribution related to forest monitoring:

We believe we will learn and benefit from the received feedback and ideas for further development, ideally also from code contribution, making it possible for us to provide a better service for everyone. We have decided to use the MIT license for eo-learn libraries to make the uptake as easy as possible, even for a commercial use. We hope that our example will lead to similar steps from other teams. The activity on sentinelhub-py package, which exposes Sentinel Hub services to Python users and was open-sourced six months ago, proves the case. Developers and community in general can advance on top of our new technology for bringing benefits to mankind. We are perfectly fine with it.

One way or another, we all benefit.