A huge thank you to everyone who took part in the GeoAI Aquaculture Pond Identification Challenge.

Looking across the top solutions, a few strong themes emerged:

• The biggest challenge was not model complexity, but generalising to the test setup, where only 4-6 consecutive months were visible for each location.
• Top participants handled this by making the training data look more like the test data through window masking / augmentation.
• Strong solutions also made good use of satellite-specific features, especially water, vegetation and radar-derived indicators.
• Several top solutions accounted for train/test distribution shift, including the possibility of a higher proportion of ponds in the test set.
• The winning approaches relied mainly on gradient-boosted tree models such as CatBoost, LightGBM and XGBoost.
• Self-training / pseudo-labelling was also used effectively by multiple top teams to adapt models further to the test distribution.

The top three solutions took slightly different approaches: 1st place focused most strongly on transferability and train/test drift, 2nd place showed that a much simpler and very lightweight CatBoost pipeline could still perform extremely well, and 3rd place placed more emphasis on detailed satellite feature engineering and accurately modelling sensor-specific missingness.

We hope these solutions are useful references for everyone interested in working with real-world Sentinel-1 and Sentinel-2 data.

🏅 Participants can now also access and download their challenge certificates from the challenge page.

Congratulations again to everyone who took part!