MM-RegNet v3 - Image2Biomass

MM-RegNet v3 is an applied research effort exploring how multimodal machine-learning methods can estimate pasture biomass from field imagery and associated metadata.

The project originated from the CSIRO Image2Biomass Kaggle competition and has grown into a testbed for transparent, reproducible experimentation. Its core goal is to connect image-based and tabular signals to predict five biomass components—green, dead, clover, GDM, and total—while maintaining strict scientific traceability.

🌱 Purpose and Approach

This repository demonstrates how a small, modular codebase can support:

Reproducible results across repeated runs.
Data versioning using DVC for complex data flows.
Light-weight MLOps concepts for persisting methods, models, and intermediate data.
Both ad-hoc experimentation and directed work toward competition goals.

All functionality executes locally through standard Python module calls—no external orchestration required.

📁 Repository Overview

Module	Description
`explorers/`	Exploratory data analysis and reporting utilities. Each script can be run individually or collectively via `explorers.data.report`.
`trainers/`	Baseline training modules for tabular and image models.
`fusers/`	Late-fusion utilities that combine predictions from separate models.
`workspace/outputs/`	Example artifacts produced by the explorers, trainers, and fusers.
`data/`	Training and test data (tracked via DVC).

Directories such as docker/ and orchestrator/ are internal and not part of the public MVP scope at this time.

🚀 Getting Started

Note: this is not open to the public at launch but this will be publicly available soon-ish. Reach out if you want to see it sooner, maybe we can talk.

Clone the repository:
git clone https://github.com/sophaime/img2bio.git, or git clone git@github.com:sophaime/img2bio.git

See https://github.com/sophaime/img2bio/blob/main/README_how-to-run.md for exact command-line examples to reproduce the included artifacts.

🔬 Guiding Principles

Reproducibility > Performance — every run should be reconstructible.
Isolation > Interference — keep experiments self-contained.
Transparency > Mystery — all transformations are visible and inspectable.

⚖️ License and Attribution

The project follows the licensing and participation terms defined by the CSIRO Image2Biomass competition.

Citation:
Liao, Q., Wang, D., Pirie, R., Whelan, J., Haling, R., Liu, J., Khokher, R., Li, X., Plomecka, M., & Howard, A. (2025). CSIRO - Image2Biomass Prediction. Kaggle. https://kaggle.com/competitions/csiro-biomass

Show BibTeX format

@misc{csiro-biomass,
    author = {Qiyu Liao and Dadong Wang and Rhys Pirie and Joshua Whelan and Rebecca Haling and Jiajun Liu and Rizwan Khokher and Xun Li and Martyna Plomecka and Addison Howard},
    title = {CSIRO - Image2Biomass Prediction},
    year = {2025},
    howpublished = {\url{https://kaggle.com/competitions/csiro-biomass}},
    note = {Kaggle}
}

📢 Status

Active development; incremental updates occur as methods evolve.