# CSIRO - Image2Biomass Prediction — My Approach (Top 2% / 82nd)

The task involves predicting multiple pasture biomass components from
natural field images
, which is challenging due to:
Strong dependency and coupling between targets
High variability across environments (lighting, species, density)
Imbalanced importance across targets (Dry_Total_g has high weight)
My approach focuses on combining
structure-aware multi-target regression
with large-scale vision models and hybrid ensembling for robust performance.
Key Idea
Structure-aware multi-target regression with constraint-based modeling and hybrid vision + tabular ensemble.
Instead of directly predicting all targets independently, I model fundamental components and enforce deterministic relationships to improve consistency on high-weight outputs.
Approach Overview
The solution is built as a
hybrid multi-branch system
:
A large-scale Vision Transformer (DINOv3 ViT-Huge) is fine-tuned for regression, with a dual-view input strategy (left-right image split) to better capture spatial heterogeneity in wide-field pasture images.
A structure-aware multi-target formulation is adopted, where only fundamental components (Green / Dead / Clover) are explicitly predicted, and aggregate targets are derived via deterministic relationships:
GDM = Green + Clover
Total = GDM + Dead
This enforces physical consistency and stabilizes predictions on critical targets.
A complementary semantic modeling branch is introduced using SigLIP embeddings and concept-level feature engineering (e.g., vegetation density, semantic similarity), followed by GBDT models to improve robustness under distribution shifts.
Ensemble Strategy
Predictions from multiple models and branches are combined using
weighted ensembling
, including:
4-fold cross-validation ensemble of DINO-based models
SigLIP + semantic feature + GBDT branch
This hybrid ensemble improves stability and generalization, especially in challenging or out-of-distribution scenarios.
Result
Final Score:
Weighted R² ≈ 0.62
Rank:
82 / 3802 (Top 2%)
The solution achieves strong performance on both:
High-weight target (Dry_Total_g)
Component-level predictions with consistent structure
Full Implementation
For full details, code, and system design:
GitHub Repository:
[https://github.com/WoshengDeng/pasture-biomass-prediction](https://github.com/WoshengDeng/pasture-biomass-prediction)
This writeup is intentionally concise. The full pipeline and implementation details are available in the GitHub repository.