Machine Learning Proof of Concept

Estimating grass curing from ordinary field photographs.

Luniqex developed an image-based AI proof of concept for estimating grass and fine-fuel curing percentage from field images. The model predicts a continuous curing value from 0-100%, where higher values indicate drier, more cured vegetation.

DINOv3 ViT-S/16+Image regressionVLM-assisted reviewGrouped benchmark MAE 7.85

Field worker capturing pasture imagery for AI-assisted grass curing assessment

Problem

Grass curing is operationally important because it affects how dry, flammable, and responsive fine fuels may be under fire-weather conditions. In practice, visual curing estimates can vary by observer, lighting, season, scene type, camera angle, and location.

The project asked whether a modern computer-vision model could learn this signal directly from ordinary photos without requiring a complex masking or segmentation workflow.

Approach

The project evolved from an initial computer-vision experiment into a curated image-regression system using DINOv3, VLM-assisted dataset review, human label auditing, and metadata calibration. Dataset quality work mattered as much as model choice: poor images were removed, corrected labels replaced original labels, and weak scene types were targeted for further review.

Conceptual diagram of a grass curing machine learning pipeline from field image to calibrated estimate

Conceptual system view: field image input, DINOv3 vision backbone, regression head, raw estimate, metadata calibration, and final curing estimate.

A Vision Language Model review pipeline used Gemini, OpenAI, and Claude to generate scene and quality metadata. Where providers disagreed, a breaker step adjudicated the result. This helped identify unsuitable training examples and understand where the model struggled.

Conceptual interface showing multi-provider visual review of pasture imagery

Conceptual VLM review interface showing how multiple visual interpretations can be compared before final dataset decisions.

Model

The current model family used a DINOv3 ViT-S/16+ backbone adapted into a regression model. Instead of classifying an image, the model outputs a single numeric curing estimate. A post-model calibrator then adjusts the estimate using month, season, and coarse region metadata.

Result

The best recorded grouped benchmark improved across successive data and training stages: early DINOv3 fine-tuning reached 15.32 MAE, the larger v2 curated dataset reached 12.82, the v3 image-only grouped model reached 8.57, and the calibrated v3 model reached 7.85.

The main lesson was that practical AI for environmental field assessment depends on the full system: curated data, reliable labels, structured review, realistic validation, and careful deployment design.

Current status

This remains a proof of concept. Accuracy depends on image quality, label quality, metadata reliability, and whether the deployment setting resembles the training data. The work demonstrates a credible pathway rather than a finished operational product.