EarthDaily Hackathon – ESA – Big Data and machine learning solutions
By Pedro Borges
In partnership with the European Space Agency EarthDaily recently hosted a Hackathon during the ESA BiDS 2023 Conference in Vienna, Austria. We will be hosting another Hackathon in early 2024 and would welcome your participation.
Hackathon Focus: Address urgent global challenges by applying the latest machine & foundational models in geospatial analytics to understand Earth’s dynamic changes. The aim was to combat environmental degradation, biodiversity loss, and natural disasters. The hackathon brought together participants in collaborative teams, tasked with building automated tools for classifying and analyzing land changes.
Focus areas included automated monitoring of rivers and water bodies, real-time deforestation detection, assessing forest fire impact, and evaluating green landscape infrastructures. Participants, regardless of skill level, were equipped with study sites, predefined data cubes of daily satellite imagery, and ready-to-use machine learning pipelines.
The test data was comprised of the VENµS satellite imagery dataset, a collaboration between the French Space Agency (CNES) and the Israeli Space Agency (ISA). This dataset, featured 12 spectral bands, daily and bi-daily coverage, and 4-5 meter resolution, proved instrumental with quality and spectral bands compatible with Sentinel-2 models.
We’re proud to announce the exceptional work and solutions showcased by our winning teams:
Best Concept: Team Coder Triad (Maximilian Wolf, Stefan Brand, Jayanth Siddamsetty)
The team tackled the crop-detection challenge where they embarked on a project to create an interactive app that enables users to draw polygons on a map and utilize extracted satellite time series data to deduce crop types. With the incorporation of the SoilGrids dataset (250m resolution), the model underwent training using LSTM, fine-tuning pre-existing models, and applying ensemble techniques to achieve outstanding results. They created a prototype of the map and created a data pipeline for the model training so the next steps would be to put all the pieces together to create an end-to-end solution. The outcome is a sophisticated prototype that elevates the elevated the team’s understanding of crop types through cutting-edge technology.
Introducing their project, one of the members shared “Data access via the EarthDaily API was very convenient with the Python tools provided by EarthDaily. It turned out that the on-the-fly geometric aggregation of satellite data for arbitrary input geometries would need performance optimizations to provide real interactivity. Besides, we faced challenges with integrating the various data sources with existing ML models. However, the EarthDaily team was very helpful and gave valuable tips to solve our problems.“
Another member said “I would like to say thanks to the organizers for putting a lot of effort in preparing the material for the hackathon. The API to access data was easy to use. My teammate had a bit of a trouble matching the structure with the training data provided, but the organizers helped us along. Would be happy to participate in the future events from Earthdaily Analytics.“
Best Visualization: Team lobi.ai (Alexander Lechner, Osorio Rodrigues)
The group, even though it was their first time using geospatial data, developed an algorithm designed to segment water areas within a given Region of Interest (ROI). They began with segmentation techniques, and were able to generate a dynamic GIF showcasing the segmentation of areas to discern differences in land, water, or vegetation. Eventually, the team decided to narrow their focus to water, calculating and leveraging the NDVI (Normalized Difference Vegetation Index) as a model feature.
Additionally, they prototyped an user-friendly interface, enabling users to pass the dataset from the UI—selected based on user inputs—for processing and receive the results back through the same interface. The group significantly advanced their understanding and visualization of water areas within the specified ROI through their creative approach.
Best Use of Existing Data: Team MetriaDataScientist (Johanna Skarpman Sundholm)
In their recent initiative, Johanna implemented a robust data filtering strategy to enhance the quality of training data by mitigating noisy inputs. The filtering methodology involved a process of sorting values, coupled with the exclusion of the lowest 1% and highest 10%. This approach underscores the team’s commitment to refining and optimizing the training data for superior outcomes.
“Well planned and thought through event that gave me the opportunity to get inspired and to see how others within my field do things. I only wish we could have extended the hackathon a couple of days to be able to explore the different parts of the prepared material in dedicated sessions.” said Johanna about this event.
The topics covered in this hackathon perfectly aligns with EarthDaily mission to harness technology for sustainability. This event exemplifies our dedication to leveraging Earth observation for sustainability.
In this Hackathon the Earth Data Store (EDS), a platform to aggregate and analyze geospatial and earth observation data, played a vital role in our projects, providing access to standardized terrestrial and spatial earth observation data to develop industry-specific applications that allowed participants to see how a region evolves over time using visual interactive maps and running deep-learning algorithms.
We extend our sincere gratitude to all participants, partners (European Space Agency), and sponsors (AWS) who played a vital role in making the hackathon a success.
We plan putting on another Hackathon in early 2024. Please visit https://pages.earthdaily.com/hackathon where we will post information on our next Hackathon soon.