Image2Biomass - CSIRO Kaggle Competition

Competition Overview

This repository is for the CSIRO - Image2Biomass Prediction Kaggle competition, a collaboration between CSIRO, Meat & Livestock Australia (MLA), and Google Australia.

Objective: Build a model that estimates pasture biomass from top-view images of Australian pastures.

Prize Pool: $75,000 USD

Timeline:

Start: October 28, 2025
Entry Deadline: January 21, 2026

Evaluation Metric: Globally weighted coefficient of determination (R²)

Competition URL: https://www.kaggle.com/competitions/csiro-biomass

Prediction Task

Predict 5 biomass measurements (in grams) from pasture images:

```
Dry_Green_g
```
- Dry green matter
```
Dry_Dead_g
```
- Dry dead matter
```
Dry_Clover_g
```
- Dry clover matter
```
GDM_g
```
- Green dry matter
```
Dry_Total_g
```
- Total dry matter (highest weight in loss: 0.5)

Loss Weights:

[0.1, 0.1, 0.1, 0.2, 0.5]

- Dry_Total_g is weighted highest

Project Structure

image2biomass/
├── src/
│   ├── train.py           # Main training script with W&B integration
│   ├── models.py          # UnifiedModel class for flexible backbones
│   ├── utils.py           # Model configs and helper functions
│   ├── image_processing.py # BiomassDataset and transforms
│   ├── data_preprocessing.py # Data wrangling utilities
│   ├── submission.py      # Inference script for Kaggle submission
│   └── eda.ipynb          # Exploratory data analysis
├── data/
│   ├── train/             # Training images (357 images, 2000x1000px)
│   ├── test/              # Test images for submission
│   ├── train.csv          # Raw training data (long format)
│   ├── y_train.csv        # Processed targets (wide format)
│   └── sample_submission.csv
├── models/                # Saved model checkpoints (.pt files)
├── scripts/               # Cloud setup scripts
├── results/               # EDA visualizations
├── config.yaml            # Default training configuration
├── sweep.yaml             # W&B hyperparameter sweep config
└── pyproject.toml         # Dependencies (uv/pip)

Quick Start

Environment Setup

# Using uv (recommended)
uv sync

# Or pip
pip install -e .

Dependencies: PyTorch 2.9+, timm 1.0+, wandb, pandas, matplotlib

Training

# Standard training with config.yaml
python src/train.py

# Hyperparameter sweep with W&B
python src/train.py --sweep

Inference

The

src/submission.py

script loads a traced model and generates predictions. Update

MODEL_PATH

to point to your trained model.

Model Architecture

Uses

UnifiedModel

from

src/models.py

- a flexible wrapper around any timm backbone:

Input Image (2000x1000)
    -> Crop center (500-1500 px)
    -> Resize to model-specific size (224-518px)
    -> timm backbone (feature extraction)
    -> Regression head (Linear->ReLU->Dropout->Linear->ReLU)
    -> 5 biomass predictions

Supported Backbones (18 total)

Defined in

src/utils.py

MODEL_CONFIGS:

ViT: vit_tiny_patch16_224, vit_small_patch16_224, vit_small_patch32_224
DINOv2: vit_small_patch14_dinov2.lvd142m (518x518 input)
ConvNeXt: convnext_tiny, convnext_nano, convnext_pico
Swin: swin_tiny_patch4_window7_224, swin_s3_tiny_224
ResNet: resnet18, resnet34, resnet50
EfficientNet: efficientnet_b0, b1, b2
MobileNet: mobilenetv3_large_100
RegNet: regnetx_016, regnetx_032

Configuration

config.yaml (default training)

lr: 0.0001
n_epochs: 1
train_split: 0.8
batch_size: 4
bf16: true           # Mixed precision (CUDA/MPS only)
patience: 5          # Early stopping patience
model_class: UnifiedModel
model_name: vit_tiny_patch16_224

sweep.yaml (hyperparameter search)

Bayesian optimization over model architectures with Hyperband early termination.

Data Format

Training data: 357 unique images with 5 targets each

Images: 2000x1000 JPG (70cm x 30cm pasture quadrat)
Metadata: State (NSW/Tas/Vic/WA), Species, Height_Ave_cm, Pre_GSHH_NDVI, Sampling_Date

Image preprocessing: Center crop (500-1500px) -> Resize -> Normalize

ViT models: mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]
ConvNets/DINOv2: ImageNet normalization

Experiment Tracking

Uses Weights & Biases (wandb):

Project:
```
image2biomass
```
Logs: train_loss, val_loss, val_R2 per epoch
Artifacts: dataset directory, trained models (TorchScript traced)

Key Files Reference

File	Purpose
`src/train.py:31`	Main `train()` function
`src/models.py:6`	`UnifiedModel` class
`src/utils.py:9`	`MODEL_CONFIGS` dict
`src/utils.py:55`	`get_model()` factory
`src/image_processing.py:102`	`BiomassDataset` class

Development Notes

Models are saved as TorchScript (.pt) for Python 3.14 compatibility
Train/val split: 80/20 with seed 42
Early stopping monitors val_loss with configurable patience
R² computed with loss weights for consistency with competition metric

Image2Biomass - CSIRO Kaggle Competition

Competition Overview

This repository is for the CSIRO - Image2Biomass Prediction Kaggle competition, a collaboration between CSIRO, Meat & Livestock Australia (MLA), and Google Australia.

Objective: Build a model that estimates pasture biomass from top-view images of Australian pastures.

Prize Pool: $75,000 USD

Timeline:

Start: October 28, 2025
Entry Deadline: January 21, 2026

Evaluation Metric: Globally weighted coefficient of determination (R²)

Competition URL: https://www.kaggle.com/competitions/csiro-biomass

Prediction Task

Predict 5 biomass measurements (in grams) from pasture images:

```
Dry_Green_g
```
- Dry green matter
```
Dry_Dead_g
```
- Dry dead matter
```
Dry_Clover_g
```
- Dry clover matter
```
GDM_g
```
- Green dry matter
```
Dry_Total_g
```
- Total dry matter (highest weight in loss: 0.5)

Loss Weights:

[0.1, 0.1, 0.1, 0.2, 0.5]

- Dry_Total_g is weighted highest

Project Structure

image2biomass/
├── src/
│   ├── train.py           # Main training script with W&B integration
│   ├── models.py          # UnifiedModel class for flexible backbones
│   ├── utils.py           # Model configs and helper functions
│   ├── image_processing.py # BiomassDataset and transforms
│   ├── data_preprocessing.py # Data wrangling utilities
│   ├── submission.py      # Inference script for Kaggle submission
│   └── eda.ipynb          # Exploratory data analysis
├── data/
│   ├── train/             # Training images (357 images, 2000x1000px)
│   ├── test/              # Test images for submission
│   ├── train.csv          # Raw training data (long format)
│   ├── y_train.csv        # Processed targets (wide format)
│   └── sample_submission.csv
├── models/                # Saved model checkpoints (.pt files)
├── scripts/               # Cloud setup scripts
├── results/               # EDA visualizations
├── config.yaml            # Default training configuration
├── sweep.yaml             # W&B hyperparameter sweep config
└── pyproject.toml         # Dependencies (uv/pip)

Quick Start

Environment Setup

# Using uv (recommended)
uv sync

# Or pip
pip install -e .

Dependencies: PyTorch 2.9+, timm 1.0+, wandb, pandas, matplotlib

Training

# Standard training with config.yaml
python src/train.py

# Hyperparameter sweep with W&B
python src/train.py --sweep

Inference

The

src/submission.py

script loads a traced model and generates predictions. Update

MODEL_PATH

to point to your trained model.

Model Architecture

Uses

UnifiedModel

from

src/models.py

- a flexible wrapper around any timm backbone:

Input Image (2000x1000)
    -> Crop center (500-1500 px)
    -> Resize to model-specific size (224-518px)
    -> timm backbone (feature extraction)
    -> Regression head (Linear->ReLU->Dropout->Linear->ReLU)
    -> 5 biomass predictions

Supported Backbones (18 total)

Defined in

src/utils.py

MODEL_CONFIGS:

ViT: vit_tiny_patch16_224, vit_small_patch16_224, vit_small_patch32_224
DINOv2: vit_small_patch14_dinov2.lvd142m (518x518 input)
ConvNeXt: convnext_tiny, convnext_nano, convnext_pico
Swin: swin_tiny_patch4_window7_224, swin_s3_tiny_224
ResNet: resnet18, resnet34, resnet50
EfficientNet: efficientnet_b0, b1, b2
MobileNet: mobilenetv3_large_100
RegNet: regnetx_016, regnetx_032

Configuration

config.yaml (default training)

lr: 0.0001
n_epochs: 1
train_split: 0.8
batch_size: 4
bf16: true           # Mixed precision (CUDA/MPS only)
patience: 5          # Early stopping patience
model_class: UnifiedModel
model_name: vit_tiny_patch16_224

sweep.yaml (hyperparameter search)

Bayesian optimization over model architectures with Hyperband early termination.

Data Format

Training data: 357 unique images with 5 targets each

Images: 2000x1000 JPG (70cm x 30cm pasture quadrat)
Metadata: State (NSW/Tas/Vic/WA), Species, Height_Ave_cm, Pre_GSHH_NDVI, Sampling_Date

Image preprocessing: Center crop (500-1500px) -> Resize -> Normalize

ViT models: mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]
ConvNets/DINOv2: ImageNet normalization

Experiment Tracking

Uses Weights & Biases (wandb):

Project:
```
image2biomass
```
Logs: train_loss, val_loss, val_R2 per epoch
Artifacts: dataset directory, trained models (TorchScript traced)

Key Files Reference

File	Purpose
`src/train.py:31`	Main `train()` function
`src/models.py:6`	`UnifiedModel` class
`src/utils.py:9`	`MODEL_CONFIGS` dict
`src/utils.py:55`	`get_model()` factory
`src/image_processing.py:102`	`BiomassDataset` class

Development Notes

Models are saved as TorchScript (.pt) for Python 3.14 compatibility
Train/val split: 80/20 with seed 42
Early stopping monitors val_loss with configurable patience
R² computed with loss weights for consistency with competition metric

Image2Biomass - CSIRO Kaggle Competition

Image2Biomass - CSIRO Kaggle Competition

Competition Overview

Prediction Task

Project Structure

Quick Start

Environment Setup

Training

Inference

Model Architecture

Supported Backbones (18 total)

Configuration

config.yaml (default training)

sweep.yaml (hyperparameter search)

Data Format

Experiment Tracking

Key Files Reference

Development Notes

Related Skills

Markdown Converter

Nano Banana Pro

1password

Image2Biomass - CSIRO Kaggle Competition

Competition Overview

Prediction Task

Project Structure

Quick Start

Environment Setup

Training

Inference

Model Architecture

Supported Backbones (18 total)

Configuration

config.yaml (default training)

sweep.yaml (hyperparameter search)

Data Format

Experiment Tracking

Key Files Reference

Development Notes