The task involves predicting multiple pasture biomass components from natural field images, which is challenging due to:
\nMy approach focuses on combining structure-aware multi-target regression with large-scale vision models and hybrid ensembling for robust performance.
\nStructure-aware multi-target regression with constraint-based modeling and hybrid vision + tabular ensemble.
\nInstead of directly predicting all targets independently, I model fundamental components and enforce deterministic relationships to improve consistency on high-weight outputs.
\nThe solution is built as a hybrid multi-branch system:
\nA 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:
\nA 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.
Predictions from multiple models and branches are combined using weighted ensembling, including:
\nThis hybrid ensemble improves stability and generalization, especially in challenging or out-of-distribution scenarios.
\nThe solution achieves strong performance on both:
\nFor full details, code, and system design:
\nGitHub Repository:
\nhttps://github.com/WoshengDeng/pasture-biomass-prediction
This writeup is intentionally concise. The full pipeline and implementation details are available in the GitHub repository.
", "messages": [ { "id": 3429704, "postDate": "2026-03-27T05:35:24.577Z", "content": "The task involves predicting multiple pasture biomass components from natural field images, which is challenging due to:
\nMy approach focuses on combining structure-aware multi-target regression with large-scale vision models and hybrid ensembling for robust performance.
\nStructure-aware multi-target regression with constraint-based modeling and hybrid vision + tabular ensemble.
\nInstead of directly predicting all targets independently, I model fundamental components and enforce deterministic relationships to improve consistency on high-weight outputs.
\nThe solution is built as a hybrid multi-branch system:
\nA 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:
\nA 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.
Predictions from multiple models and branches are combined using weighted ensembling, including:
\nThis hybrid ensemble improves stability and generalization, especially in challenging or out-of-distribution scenarios.
\nThe solution achieves strong performance on both:
\nFor full details, code, and system design:
\nGitHub Repository:
\nhttps://github.com/WoshengDeng/pasture-biomass-prediction
This writeup is intentionally concise. The full pipeline and implementation details are available in the GitHub repository.
", "rawMarkdown": "The task involves predicting multiple pasture biomass components from **natural field images**, which is challenging due to:\n\n- Strong dependency and coupling between targets \n- High variability across environments (lighting, species, density) \n- Imbalanced importance across targets (Dry_Total_g has high weight) \n\nMy approach focuses on combining **structure-aware multi-target regression** with large-scale vision models and hybrid ensembling for robust performance.\n\n---\n\n### Key Idea\n\n**Structure-aware multi-target regression with constraint-based modeling and hybrid vision + tabular ensemble.**\n\nInstead of directly predicting all targets independently, I model fundamental components and enforce deterministic relationships to improve consistency on high-weight outputs.\n\n---\n\n### Approach Overview\n\nThe solution is built as a **hybrid multi-branch system**:\n\n- 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. \n\n- 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: \n - GDM = Green + Clover \n - Total = GDM + Dead \n This enforces physical consistency and stabilizes predictions on critical targets. \n\n- 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. \n\n---\n\n### Ensemble Strategy\n\nPredictions from multiple models and branches are combined using **weighted ensembling**, including:\n\n- 4-fold cross-validation ensemble of DINO-based models \n- SigLIP + semantic feature + GBDT branch \n\nThis hybrid ensemble improves stability and generalization, especially in challenging or out-of-distribution scenarios.\n\n---\n\n### Result\n\n- Final Score: **Weighted R² ≈ 0.62** \n- Rank: **82 / 3802 (Top 2%)** \n\nThe solution achieves strong performance on both:\n- High-weight target (Dry_Total_g) \n- Component-level predictions with consistent structure \n\n---\n\n### Full Implementation\n\nFor full details, code, and system design:\n\nGitHub Repository: \nhttps://github.com/WoshengDeng/pasture-biomass-prediction\n\n---\n\nThis writeup is intentionally concise. The full pipeline and implementation details are available in the GitHub repository." } ], "comments": [], "raw_markdown_by_id": { "3429704": "The task involves predicting multiple pasture biomass components from **natural field images**, which is challenging due to:\n\n- Strong dependency and coupling between targets \n- High variability across environments (lighting, species, density) \n- Imbalanced importance across targets (Dry_Total_g has high weight) \n\nMy approach focuses on combining **structure-aware multi-target regression** with large-scale vision models and hybrid ensembling for robust performance.\n\n---\n\n### Key Idea\n\n**Structure-aware multi-target regression with constraint-based modeling and hybrid vision + tabular ensemble.**\n\nInstead of directly predicting all targets independently, I model fundamental components and enforce deterministic relationships to improve consistency on high-weight outputs.\n\n---\n\n### Approach Overview\n\nThe solution is built as a **hybrid multi-branch system**:\n\n- 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. \n\n- 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: \n - GDM = Green + Clover \n - Total = GDM + Dead \n This enforces physical consistency and stabilizes predictions on critical targets. \n\n- 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. \n\n---\n\n### Ensemble Strategy\n\nPredictions from multiple models and branches are combined using **weighted ensembling**, including:\n\n- 4-fold cross-validation ensemble of DINO-based models \n- SigLIP + semantic feature + GBDT branch \n\nThis hybrid ensemble improves stability and generalization, especially in challenging or out-of-distribution scenarios.\n\n---\n\n### Result\n\n- Final Score: **Weighted R² ≈ 0.62** \n- Rank: **82 / 3802 (Top 2%)** \n\nThe solution achieves strong performance on both:\n- High-weight target (Dry_Total_g) \n- Component-level predictions with consistent structure \n\n---\n\n### Full Implementation\n\nFor full details, code, and system design:\n\nGitHub Repository: \nhttps://github.com/WoshengDeng/pasture-biomass-prediction\n\n---\n\nThis writeup is intentionally concise. The full pipeline and implementation details are available in the GitHub repository." } }