Fighting Yield Loss and Food Insecurity With a Platform for AI-ready Data: VISTA’s Plan for Ensuring Early Crop Growth Predictions
Yield loss is becoming a dour and often-seen occurrence in the world, especially as temperatures continue to rise. While we should strive to adopt practices which will prevent further rising of temperatures, we should also work towards mitigating food insecurity. One way this can be done is by early crop growth prediction through utilizing the agrifood data space.
That is why we have asked VISTA about their involvement in the STELAR project. As STELAR’s partners, they are at the forefront of Pilot B: Early crop growth predictions. Heike Bach, Silke Migdall, Marion Buddeberg, Christoph Jörges and Florian Appel told us the key elements leading to yield loss, and how we can predict them. Additionally, they highlight the advantages of STELAR’s planned Knowledge Lake Management System (KLMS), which transforms raw data into an AI-ready data lake, benefiting the agricultural industry.
Food security is a topic that is becoming ever more important. The planet keeps breaking heat temperature records. And temperatures, of course, have an impact on yield. What are the primary factors contributing to yield loss in agricultural production? Additionally, what are the key parameters that play a vital role in predicting food supply?
Every crop has “ideal” conditions under which it grows best. These “ideal” conditions are usually ranges consisting of temperatures that are neither too high, nor too low. Also, enough water (but not too much) and nutrients are needed. The weather conditions are the driving factor, which determines whether the plants get the right amount of sunshine and warmth as well as rain. But, of course, management decisions also play a part. With Earth Observation (EO) data, we can monitor how the crops are currently doing, e.g. how much green leaf area they have developed. If we assimilate this information into a crop growth model that is driven by the weather conditions, we can also model what this means for items like biomass growth and yield formation. This way, we can for example model what effects there are on the yield when it is very hot or very dry – for example, when necessary irrigation is not carried out. Using weather scenarios for the next few weeks, we can also predict how much yield there will be at the end of the season.
What are the key challenges regarding the prediction of crop growth and yields?
Ideally, crop predictions should come early enough in the year so that changes to outcomes can still be made, or infrastructure utilization can be optimized. On the level of the individual farm, this means that a crop prediction should happen before the last fertilization, when the crops are still in their vegetative phase and not already ripening. On the big scale, when predicting yields for whole continents, the earlier the predictions are available, the earlier warnings for food shortages due to yield loss can be given and political measures to support the regions affected by high yield losses can be taken.
There are several key challenges to this. Of course, the predictions need to be accurate to be helpful, so very good input data is needed. This means utilizing well-calibrated EO data and sophisticated methods of analyzing them. It also means using well-validated crop growth and yield models that we can have a high degree of trust in. Last but not least, we cannot look into the future, so when calculating yields several weeks – or ideally even months – in advance, we need expert knowledge not only in our crop growth model, but also in our weather scenarios for how the next months will most likely develop.
What benefits does STELAR’s Knowledge Lake Management System (KLMS) bring to the challenges for yield prediction? Which specific data sources will be utilized?
We combine data from different sources, mainly operational multispectral satellite data, as well as available hyperspectral data, and use the technologies available in the KLMS for data fusion and data imputation, as well as crop classification, to derive the best possible input data for our crop growth modeling.
STELAR is using data from the European Copernicus fleet like Sentinel-2 data, but also very current research data like acquisitions from the German hyperspectral satellite EnMAP, which was launched in April 2022. The data from Sentinel-2, which really is the “work horse” among the Earth Observation satellites for agricultural applications, is available in a very operational way through European cloud platforms (such as the Food Security TEP) and acquisitions happen globally in regular 3 to 5 day intervals. But the next-generation hyperspectral data from EnMAP is only available for selected test-sites that also need to be specifically targeted and at a low temporal frequency. Data fusion helps us here to integrate this data into larger service chains. This is highly needed, as, for the first time, the higher spectral resolution of hyperspectral data allows the remote sensing quantification of many biochemical plant parameters via absorptions and analysis of their depths.
Apart from the satellite data, we also use meteorological data and weather forecasts. For instance, seasonal forecasts are available for the next 6 months from the Copernicus Climate Data Store. Historical futures (i.escenarios on how the year will develop meteorologically from any given point in time on, which are based on historical weather records) can be built from different meteorological data sets depending on region and availability.
Last but not least, crop growth modeling is used to provide synthetic data about plant and soil characteristics. Using measured variables (e.g., meteorological data, Leaf Area Index (LAI)) as input, crop growth modeling can produce variables that cannot be seen from above or measured on the ground continuously and area-wide. This way, various important indicators such as the phenology, i.e. the growth stage of the plant, can be added as valuable information for the farmer.
Could you elaborate on your strategy for collaborative partnership within the project to contribute towards the development of a comprehensive knowledge lake for early prediction of crop growth?
STELAR brings together experts from the computing sciences with domain experts, which is an incredible advantage for both parties. We at VISTA are domain experts with extensive knowledge on agricultural production, on the analysis of satellite imagery, and on crop growth modeling. In the KLMS, we get support from the data science experts, who are, for example, deeply into data imputation issues, or are experts on artificial intelligence and bias detection. Of course, we aren’t “speaking the same language” in the beginning – meaning data is often stored in different formats and languages. So in the first part of STELAR, we focus a lot on co-defining and co-analyzing the various challenges and possible solutions, not only in technical terms, but also in terms of domain vocabulary. We need to get really specific, so that we all understand each other well and can then work on bringing together our expertise to develop better, innovative solutions.
In which countries is the pilot planned to be conducted for testing and validation purposes?
We will establish the pilot in three regions: one in France, one in Austria and one in Germany. This way, we will have enough variety in climatic and soil conditions to be sure that our methods will be transferable between regions. Nevertheless, we are still working in the EU under well-known conditions with good data availability. This ‘laboratory approach’ is important because we need data not only as input and for training, but also for validation of our results so that we can be certain that our outcomes reach the needed accuracy levels.
What active role will farmers assume within the project in terms of their contribution and involvement?
Early predictions are useful for many stakeholders in the agricultural market – from the farmers, players from agricultural extension, all the way to the agricultural industry and stakeholders such as agricultural insurances. Within the project, we will mostly focus on farmers, agricultural advisers and food security experts as end users of our services. We will present our results to our stakeholders, preferably already in a first version during the project and integrate their feedback in order to modify our approach and make sure that the results produced by the KLMS have the best benefit for the users.
If you want to find out more about STELAR’s Pilots, read our previous interview with the team spearheading Pilot A: Risk prevention in food supply lines.