How amandeva/amini-geofm-decoding-the-field-challenge works
Observations from the file tree
• The repository contains only:
• Complete-project-doc.pdf
• README.md with a title and two embedded images (hosted on Google Drive)
Likely project type and purpose
• This is not a conventional codebase (no src/, requirements.txt, package.json, etc.).
• It’s most likely a research / ML project write-up (or competition submission report) documenting:
• a Geo-spatial Foundation Model (GeoFM) approach
• for a “Decoding the Field” challenge
In other words: you’re diagramming the *system described by the documentation*, not the structure of a software implementation.
How to extract the architecture from the PDF (your real “source of truth”)
Because there is no code, the key step is to mine the PDF for:
Inputs / data sources
• Satellite imagery? Multispectral? SAR? Time series? Field boundaries? Labels?
Preprocessing
• Tiling/chipping, resampling, normalization, cloud masking, band selection, temporal stacking, etc.
Modeling approach
• Which GeoFM? Frozen vs fine-tuned? Any adapters/LoRA? Decoder head?
Training loop & losses
• Objective (classification/segmentation/regression), loss functions, augmentations
Evaluation
• Metrics, cross-validation split, leaderboard evaluation format
Inference / submission format
• Batch inference pipeline, post-processing, formatting outputs
Compute + tooling
• GPU/TPU, framework (PyTorch, Lightning), experiment tracking (W&B), artifact storage, etc.
Practical instruction: skim the PDF and create a one-page bullet list with sections:
• Data → Preprocess → Model → Train → Evaluate → Infer → Deliver
For this kind of project, you want to explicitly show *both*:
Artifact flow (models/metrics/submissions)
• Data sources → Preprocessing → Train dataset
• Train dataset → Training → Checkpoints
• Checkpoints → Inference
• Inference → Submission artifact
• Validation → Metrics → Model selection (feedback loop)
• Metrics / error analysis → Preprocessing tweaks / model tweaks
This reflects how these challenge solutions are actually built.
Architectural patterns/design principles to highlight (GeoML specific)
• “Foundation model + task-specific decoder” architecture
• Separation of offline preprocessing vs online inference
• Reproducible experiments (data versioning, fixed splits, checkpointing)
• Geospatial leakage prevention (spatially-aware splits)
• Patch-based processing (tile/chip) due to large rasters
These are core to “Decoding the Field” style solutions and should be visually encoded.
Diagramming guidance: make it detailed and componentized (don’t overthink)
Because the repo contains only documentation, your goal is to over-separate into components and label them clearly. Use consistent visual semantics:
• External data sources, compute environment, outputs/submission
• End-to-end ML flow from ingestion to submission
• GeoFM encoder + decoder head + losses + outputs
Since you only have:
• Complete-project-doc.pdf → treat as the architecture spec
• “Architecture derived from Complete-project-doc.pdf”
• If the PDF includes specific model names, dataset names, metrics, or steps, use those exact terms in labels.