CSIRO Image2Biomass - 33rd Place Silver Medal Solution

33rd place out of 3000+ teams on the private leaderboard. Score: 0.74 R2

About the Competition

CSIRO Image2Biomass was a Kaggle competition focused on predicting pasture biomass from images. The dataset contains images of Australian pastures collected across 19 sites in NSW, Victoria, Tasmania, and Western Australia from 2014-2017. The goal was to predict 5 biomass components (Dry Green, Dry Dead, Dry Clover, GDM, Dry Total) which are important metrics for livestock grazing management.

Solution Overview

The solution combines two approaches:

1. Fine-tuned DINO v3 Huge (vit_huge_plus_patch16_dinov3) - 75% weight
2. SigLIP embeddings with GBDT ensemble - 25% weight

The key insight was retraining the DINO backbone on this specific dataset instead of using frozen embeddings like most public solutions.

Repository Structure

training.py     - Full DINO training code. Fine-tunes vit_huge_plus_patch16_dinov3 
                  (1.1B parameters) with 4-fold cross validation, early stopping, 
                  and cosine annealing scheduler.

inference.py    - Ensemble inference combining DINO and SigLIP. Loads trained 
                  checkpoints, runs prediction with post-processing, and generates 
                  submission.csv file.

Results

Stage	Score	Rank
Public LB	0.73	105th
Private LB	0.74	33rd

CV Mean R2: 0.83

Complete Writeup

For a detailed explanation of my approach, what worked, what did not work, and lessons learned:

https://www.kaggle.com/competitions/csiro-biomass/writeups/33rd-place-silver-medal-solution

Links

• Competition: https://www.kaggle.com/competitions/csiro-image2biomass
• Training Notebook: https://www.kaggle.com/code/ibrahimqasimi/csiro-biomass-33rd-rank-training-0-74-0-63
• Inference Notebook: https://www.kaggle.com/code/ibrahimqasimi/csiro-biomass-33rd-rank-inference-0-74-0-63
• Model Checkpoints: https://www.kaggle.com/datasets/ibrahimqasimi/dino-huge-retrain-checkpoints-zul
• SigLIP Model: https://www.kaggle.com/models/aishikai/google-siglip-so400m-patch14-384
• Data Split: https://www.kaggle.com/code/samu2505/csiro-datasplit

Requirements

torch
torchvision
timm
albumentations
transformers
scikit-learn
lightgbm
catboost
pandas
numpy
opencv-python

---

Muhammad Ibrahim Qasmi ( Youngest 3x Kaggle Grandmaster)

Follow me on kaggle :https://www.kaggle.com/ibrahimqasimi