Machine Learning Yield Prediction & Computer Vision Plant Phenotyping
Wheat contributes to ~20% of global calories and over A$10 billion annually to the Australian economy, yet yield outcomes are highly sensitive to complex, non-linear interactions between climate, cultivars, and management practices.
This data science project bridges the gap between regional environmental data and micro-level plant traits. I developed a dual-pipeline system featuring a Machine Learning yield forecasting model and a Computer Vision organ segmentation tool, wrapped into an interactive dashboard to drive actionable agricultural decision-making.
- Languages: Python
- Computer Vision: PyTorch, Torchvision, Segmentation Models PyTorch (SMP)
- Machine Learning: Scikit-Learn (Random Forest, Gradient Boosting, SVR, PCA)
- Data Engineering & Analysis: Pandas, NumPy, Matplotlib, Seaborn
- Deployment & Cloud: HPC (Rangpur) for privacy-compliant model training
Traditional simulation models often fail to capture complex environmental relationships or require immense computational overhead. This pipeline scales predictive analytics to regional levels.
- Objective: Forecast wheat yield (kg/plot) based on tabular environmental data, daily weather records, and cultivar types.
- Models Evaluated: Random Forest (RF), Support Vector Regression (SVR), Gradient Boosting (GB), and Principal Component Analysis (PCA) with Linear Regression.
- Key Achievements: * Handled severe data imbalances and high-dimensionality weather data through robust preprocessing and feature engineering.
- Identified Random Forest as the optimal baseline model, providing highly accurate predictions while maintaining critical feature interpretability for stakeholders.
To understand yield at the biological level, we need to extract traits (like head-leaf ratio or stem density) directly from field imagery.
- Objective: Perform precise semantic segmentation on field images to classify pixels into four categories: Background, Wheat Head, Stem, and Leaf.
- Architectures Built: U-Net, SegFormer, and DeepLabV3+.
- Key Achievements:
- Fine-tuned DeepLabV3+ delivered the highest overall accuracy.
- Successfully overcame severe class imbalance, proving highly robust at identifying thin, difficult-to-segment minority classes like wheat stems.
The sample outputs below highlight the DeepLabV3+ model's accuracy on unseen test data.
Understanding the Visuals:
- Left Panel (Input): The raw RGB field image of the wheat crop.
- Middle Panel (Ground Truth): The human-annotated mask indicating the exact location of the wheat heads (green), stems (orange), and leaves (blue).
- Right Panel (Prediction): The model's generated output. As demonstrated, the model successfully distinguishes between overlapping plant organs and effectively isolates the thin stem structures, which is traditionally a highly challenging task in agricultural computer vision.
To translate these complex agronomic models into actionable insights, I developed an interactive Tableau dashboard. The dashboard allows stakeholders to explore yield predictions alongside environmental factors and cultivar performance visually.
Below is a walkthrough of the interactive features, including regional yield filtering and cultivar performance analysis:
This project utilizes high-fidelity agricultural and meteorological datasets under institutional and open-data agreements:
- Yield Data: Sourced from the National Variety Trials (NVT), providing plot-level harvest metrics across diverse Australian growing regions.
- Weather Data: Climate records, including solar radiation and evapotranspiration, were provided by the SILO database (Scientific Information for Land Owners), managed by the Queensland Government.
- Privacy & Ethics: All data processing was conducted in compliance with UQ data governance policies. Sensitive datasets were processed locally or via the Rangpur HPC cluster to ensure secure data handling and privacy.