Adding more dates
Hi there,
I'm working in remote sensing analysis for ecological and biodiversity-related applications. I very much appreciate your efforts to make the Sentinel 2 data easily accessible to broad community in standartized format, and excited about new developments at the interface of deep learning and remote sensing.
I was wondering, is there a similar dataset or potential expansion of this dataset with multiple images through years for the same location?
The reason I'm asking is that for many ecological and agricultural applications vegetation seasonality is a very important characteristic and having the same image in different seasons would be very useful. To my understanding, the current implementation of Major-TOM takes a random image at each location, which is fine for analyzing general spatial structure and urban areas, but creates challenges for vegetation analysis when e.g. 2 images next to each other are taken in summer and winter season.
From practical perspective, I understand that this increases the size of stored images. But even adding one image in each season would allow to track vegetation cycles and enrich foster applications.
Kind regards,
Elena
Okay, I actually found one Sentinel 2 dataset with seasonality, but it is far from being uniformly distributed:
https://github.com/zhu-xlab/SSL4EO-S12?tab=readme-ov-file
Hey! We absolutely want time series as part of Major TOM eventually. The issue is that we had to prioritise spatial vs. temporal coverage and for the first release we opted for spatial. Seasonal time series would be ideal in the future, but we would probably need to limit locations somewhat to allow for so much data. Would you be interested in the future discussing what locations might be most impactful to focus on?
And yes, SSL4EO-S12 looks great for an application like yours, as they've prioritised temporal coverage a bit more over spatial.
Thanks for the answer!
For ecological applications it would be brilliant to keep the uniform spatial distribution or focus more in mountain regions, as they have more eco-zones.
Unfortunately, most of the other datsets I have seen (incl. SSL4EO-S12) focus too much on cities or agricultural areas. Btw, JFYI, here is the overview of the datasets that exist already for different applications:
https://earthnets.retool.com/embedded/public/676aa812-0dca-4e3b-a596-b043d852571d
Very curious for the updates of Major-TOM :)
Indeed time series is what keeps getting mentioned whenever we speak about the future of Major TOM.
We're currently expanding in several directions and can only do so much at a time, while time-series is one of the most demanding (in terms of compute and storage) expansions we could consider.
I think a time series with 1 image per season could be a good start, but that still requires a lot of storage. Perhaps we could balance that a bit by decreasing density of coverage and covering e.g. only 1/10 of the samples across the globe. Ideally, we would need a few stakeholders interested in time-series to start a discussion on the sampling strategy and then see how we can build a Major TOM expansion that follows that strategy.
It's a bit quiet there right now, but we have also set up a discord for conversations like that https://discord.gg/SsUJCDcrcu
@eplekh - would you be interested in engaging in a discussion about how we could design a time-series Major TOM? We should probably also look for other potential users who could offer a diversity of opinions.
Yep, I totally understand the storage concern. Most of the current multiseasonal datasets have no more than 200K images, so reducing the size to 1/10 seems reasonable. Plus, for the applications such as SeCo, in principle only 2 images are needed - from 2 contrasting seasons (i.e. "summer" and "winter"). The fact that the network can associate the same image with leaves-on and leaves-off seasons already means a lot.
The time-series is a bit different application, which allows to track the changes (e.g. in land use) over time. For that, a bit different data and architectures are used (e.g. Preso).
@mikonvergence
, yes, I would be interested in such discussion. In fact, at the moment I'm thinking of creating my own dataset with 2 or 4 seasons distributed globally and consisting of around 100-200K images, because I really want to build a foundational model for a variety of ecological applications based on it. If we can combine the efforts there, it would be great!
What is your timeline for potentially extending Major-TOM with seasons?
My project timeline is around 1 month or so.
What would be the best way for me to contribute?
If I would do it independently, I would use a script that just access GEE and downloads images for random locations that have cloud cover <20%. But then it would be maybe 250GB and would need to be tranfered to your database. And maybe you have a more efficient code for downloading.
@mikonvergence , thanks for the invite! I have joined the Discord and happy to respond there as well.
Hi @eplekh - that all sounds great, indeed it would be interesting to discuss in more detail different approaches to temporal expansions along the lines you're describing (e.g. multi-season pairs, sparse multi-season, or denser time series) - and the associated techniques each of those would enable.
It would be absolutely phenomenal if we could build this expansion within the next month. We currently do not have the capacity to lead the development of a time-series / season expansion of Major TOM, but would be very keen to collaborate on this and see how we can join the efforts to build something that will be easy to combine with other ongoing expansions of Major TOM. We should also have some helpful scripts laying around to make the process easier.
I think for the sake of clarity, it would be good to communicate over discord, which has channels designed specifically for discussing the future growth projects of Major TOM. If that's easier, I'm also happy to schedule a call to coordinate this development.
Let's continue this exchange on discord in any case.