1. Towards Generalizing NDVI image Synthesis
We promised to discuss how well our NDVI model works for other non-Estonian areas in this issue, and it is no trivial issue (pun intended). The images below illustrate the structural differences in agricultural areas for disparate regions considered. (Grayscale image intensities represent NDVI values, with dark to light corresponding with low to high values respectively).
For each sub-tile the difficulty in modeling (NDVI and structural changes in fields) can be conceptualized by the number of contiguous fields and layout variations per a 512px image tile. It also includes other key indicators such as differences in NDVI crop signatures, and time difference between historical inputs and target outputs. In a recent demo for Northeastern Germany, we used an ensemble model trained over Baltic areas and Germany (due to field similarities) to synthesize NDVI images illustrated in Figure 2. Evaluations over a test area (consisting of 360 512px images each representing 5242.88 km2 land patches) leave our German model at a 68% averaged structural similarity to target images.
We still have some way to go, and some improvement techniques involve active learning, training with smaller image tiles (256px for e.g.) amongst other data and model architectural changes. Looking forward to sharing progress in subsequent releases including how the results look in a demo web map!
Hudson Taylor Lekunze, Data analyst
2. 2021 was a good development and growing year for KappaZeta
Perhaps the most visible and ready to use product coming from us this year is KappaMask Sentinel-2 cloud mask. It is useful for wide variety of users from EO and non-EO sector. Everyone, who needs to use Sentinel-2 data in large volumes and wants to separate cloud-free areas from cloudy ones, will benefit from KappaMask.
Though still in development, in our first focus area – Northern Europe, it already seems to be convincingly the most accurate free and open Sentinel-2 cloud mask. Read more about it from our recently published research paper: https://www.mdpi.com/2072-4292/13/20/4100
Try it out: https://github.com/kappazeta/cm_predict
Or re-use our reference dataset: https://zenodo.org/record/5095024#.YP8HpY4zaUk
First trials by users confirm that it even works outside the focus area – in Southern Europe. Though, with some land cover types, especially with water bodies it sometimes gives strange results. It is most likely due to the limitations of the reference set, where our model is fitted. We will address it in the 2nd phase of the project.
The 2nd phase of the project has already started. The goal is very simple – to make KappaMask the most accurate free and open Sentinel-2 cloud mask globally in all seasons, not just for Northern European terrestrial summer season conditions. We are extending the reference set (manual labelling with clever active learning approach plus re-using free and open data sets by other teams) and improving the model. Next version of KappaMask will be released in summer 2022 – stay tuned.
In 2021 we also extended our Sentinel-1 data pre-processing services portfolio. Besides the parcel level aggregated feature set time series, you can also have raster output for our coherence and backscatter data layers. The biggest changes and major new product launches of our Sentinel-1 analysis ready data (ARD) layers for business to business (B2B) and business to government (B2G) are still ahead and coming in 2022 and 2023. To make it happen we were lucky to receive the investment from ESA InCubed program (https://incubed.phi.esa.int/), but even more proud we are about the pre-purchase deals we signed for close to 100 kEUR. It proves that users see the value and need our easy to use, well calibrated and noise removed Sentinel-1 ARD layers.
2021 has also been a good year in terms of growing interest from investors. So far, we haven’t taken any investor money on board, but who knows maybe in the coming years we join someone with who we have a great match, shared vision and ideals. The increased interest of investors reflects the fact that we are developing, growing and doing relevant work for the world. Here we must pay credit to the mentorship programs we took part in 2020-2021. The Design Masterclass of Enterprise Estonia gave us just an invaluable toolbox for fast prototyping, product-market-fit searching, user interest mapping and validation. Let’s now put it to daily use in KappaZeta, whatever new product or service we are about to create. Kristjan Raude’s mentoring hours within Põhjanael program were equally valuable for marketing and sales work. According to my opinion – pure gold, we just need to apply it in practice. Last, but not least this autumn marks the ending of our incubation and mentorship support in Superangel program. It is difficult to overestimate the value of the very practical advice and contacts we received from Kalev Kaarna & co. Superangel team – you really do a work with a mission, which is beyond making just money.
This year marks also the end of two important ESA projects – grazing detection and HaTRY. In grazing detection project, we were working towards grazing detection methodology development using satellite monitoring. In HaTRY the goal was to predict the optimal time for winter wheat, winter rapeseed and spring barley harvesting considering the ripeness and moisture content. Both projects served as a good preparatory work for future developments to reach tried and tested operational services one day.
Regarding subsidy checks services for Estonian Government (PRIA https://www.pria.ee/) 2021 was a challenging year. Crop classification model results in 2021 new time series were much worse than in the RITA project test sets. It underlines a fundamental problem – our crop classification model does not generalize well and is too dependent on the specific behavior of the Sentinel-1 and -2 feature set time series from the training set seasons. The drawback does not render all the work we did in RITA methodology development project (2019-2020) useless. The input features of S1 and S2 parameter time series still carry the information of which crop corresponds to a certain input data. The input data as prepared by us is very rich and fine, we only need to work on our model and methodology to make it more robust and better generalizable when moving from one season to another.
With mowing detection, it seems that we underestimated the varieties what our weather can bring once again. With experiences and training data from 2017, 2018, 2019 and 2020 in our pocket it seemed that we had largely solved the mowing detection task at least for Estonian conditions. But the tropical weather and drought conditions in June-July this year brough a surprise. The vegetation dried out so fast that the pattern in NDVI and coherence time series looked very similar to mowing events. Without knowing that there was a drought we would have marked them as mowing events. Should we bring weather data back to let the model learn the patterns of a drought and distinguish them from the mowing events? We will address this question for the seasons to come.
No matter the economic situation, inflation, deflation – the greatest value of KappaZeta is always our people. In that sense 2021 was a good year. I’m very glad to welcome young and talented new members – Olga, Hudson, Anton, Liza and Tanya to our team. Early 2022 will bring some more positive news regarding great people joining KappaZeta. Follow us in LI and FB to get the news.
Have peaceful Christmas time with your friends and families. Productive and positive new year!
