In this challenge, the competitors' goal was to build a model to predict the yearly Aboveground Biomass (AGBM) for 2,560 x 2,560 meter patches of Finnish forests using satellite imagery from Sentinel-1 (S1) and Sentinel-2 (S2). AGBM is a widespread metric for the study of carbon release and sequestration by forests and is used by forest owners, policymakers, and conservationists to make decisions about forest management.
This repository contains code from winning competitors in the BioMassters DrivenData challenge. Code for all winning solutions are open source under the MIT License.
Winning code for other DrivenData competitions is available in the competition-winners repository.
| Place | User | Private Score | Summary of Model |
|---|---|---|---|
| 1 | kbrodt | 27.63 | Combined S1 and S2 images into a 15-band 12-month composite, used a UNET model with test time augmentation. |
| 2 | Team Just4Fun: qqggg, HongweiFan | 27.68 | Used a SWIN UNETR model adopted from the Medical Open Network for AI on satellite features represented in 3D. |
| 3 | yurithefury | 28.04 | Aggregated S1 and S2 data into 6 median composites per year, ensembled together 15 models using UNET++ architecture. |
| MATLAB Bonus Prize | Team D_R_K_A: kaveh9877, AZK90 | 31.08 | Combined S1 and S2 images into a 15-band 12-month composite, used a 1-D CNN to perform by-pixel regression. The resulting labels were added as a 16th band to the composites, which were then passed through a 3-D U-Net model. |
Additional solution details can be found in the reports folder inside the directory for each submission.
Winners announcement: Meet the BioMassters
Benchmark blog post: The BioMassters