Primary competition visual

Amini GeoFM Decoding the Field Challenge

Helping Africa
$8 500 USD
Completed (8 months ago)
Classification
798 joined
153 active
Starti
Jun 10, 25
Closei
Jul 20, 25
Reveali
Jul 21, 25
User avatar
Brainiac
Seeking clarification on Datasets, Models and Efficiency
Data · 19 Jun 2025, 09:23 · 8

Band Aggregation

How were the Sentinel-2 bands preprocessed? Were the pixel time-series aggregated using statistics like mean, median, or other temporal summarization methods?

Model Usage Constraint

Are we strictly required to use only PRESTO and Amini Earth FM, or are we allowed to use other models in combination

Efficiency Prize Criteria

What does "Most efficient solution in the top 10" mean?

Is it based on model size, inference time, resource usage, or a specific combination?

Use of Related Open Datasets

Can we use the Côte d’Ivoire Byte-Sized Agriculture Challenge Sentinel-2 data?

@Amy_Bray

Discussion 8 answers
User avatar
wildcard_

@AmyBray please clarify these concerns, specifically if we can use these models in combination with others

20 Jun 2025, 07:34
Upvotes 0
User avatar
Brainiac

Seems we can use other models, but still clarification is needed if we can create other features on top of the extracted embeddings, or we are only allowed to use only embeddings extracted from the two models

The only limitation is you can use the two listed models.

User avatar
Brainiac

Thanks for the clarification, @Jackline.

Just to confirm:

  • The starter notebook uses embeddings from PRESTO and Amini Earth FM, followed by a RandomForest classifier.
  • Are we allowed to use different classifiers (e.g., XGBoost, neural nets) on top of the embeddings?
  • And are we allowed to create additional features beyond the extracted embeddings — for example, handcrafted features from the input data — or are we restricted to only using the embeddings as input?

Appreciate your guidance!

Yes, the notebook was a guidance on one way to use the Model, you can use other classifiers, finetune the model etc.

User avatar
Brainiac

Got it, thanks for confirming!

User avatar
Brainiac

@Amy_Bray

Please also clarify how the Sentinel-2 bands in test.csv were scaled?

A quick summary shows that band values range from 0 up to about 3.2. That range is well below the raw 0 – 10 000 DN scale, yet above the usual 0 – 1 surface-reflectance range, so we’re unsure whether these numbers are:

  • Top-of-Atmosphere (TOA) reflectance divided by a constant,
  • Surface Reflectance divided by a constant, or
  • another scaling scheme altogether.

Knowing the exact divisor (e.g., 10 000, 4 096, etc.) will help us avoid double-scaling when we feed the data into PRESTO or Amini Earth FM.

Thanks a lot!

22 Jun 2025, 09:11
Upvotes 1

hey only scaling was done, of factor 0.0001