CSIRO – Image2Biomass Prediction
About
For the CSIRO Image2Biomass Prediction, my goal was to develop a machine learning model that could estimate pasture biomass directly from field images and supporting environmental data.
The problem I am trying to tackle is deceptively simple: farmers need to know whether there’s enough grass in a paddock to feed their herds. In reality, estimating pasture biomass, the amount of feed available, is critical for grazing management, animal welfare, and land sustainability, yet traditional measurement methods are slow, inconsistent, or costly to scale.
To replace the “clip and weigh” and unreliable sensor methods with a scalable, I aim to deliver data-driven solution through this task. The dataset provided included annotated pasture images from various regions of Australia, paired with ground-truth biomass values and NDVI (Normalized Difference Vegetation Index) readings.
The task was to train a predictive model capable of accurately estimating biomass across different vegetation types, seasons, and soil conditions. Success meant building a model that could generalize across these contexts, effectively supporting precision agriculture and sustainable land management.
Central to this task is a simple goal: to help farmers make smarter grazing decisions, enabling researchers to monitor pasture health more effectively, and advancing the use of AI and remote sensing for environmental resilience.
Method
Scoring
Model performance is evaluated using a globally weighted coefficient of determination (R²) computed over all (image, target) pairs together. Each row is weighted according to its target type:
| Target Type | Weight |
|---|---|
| Dry_Green_g | 0.1 |
| Dry_Dead_g | 0.1 |
| Dry_Clover_g | 0.1 |
| GDM_g | 0.2 |
| Dry_Total_g | 0.5 |
Instead of calculating R² separately for each target and then averaging, a single weighted R² is computed using all rows combined, with the above per-row weights applied.
