Project
FlashForest drone seeding microsite machine learning
Applied machine-learning work with FlashForest on classifying drone-imagery microsites for aerial seeding and reforestation decision support.
This project develops a reproducible machine-learning pipeline for classifying reforestation microsites from high-resolution drone imagery, supporting drone-based tree planting and seed-deployment decisions after disturbance.
The current workflow treats microsite selection as a four-class semantic segmentation problem: background, good, fair, and poor. RGB orthomosaic tiles are matched to pixel-wise segmentation masks prepared in Roboflow by forestry experts.
A key technical challenge is severe class imbalance: most labelled pixels belong to the poor class, while good and fair microsites are much rarer. The project therefore emphasizes spatial data splitting, per-class intersection-over-union metrics, and careful interpretation of aggregate accuracy.
The baseline model uses a U-Net encoder-decoder architecture on 512 by 512 RGB image tiles. Experiments have tested early stopping, geometric augmentation, photometric augmentation, weighted sparse categorical cross-entropy, Dice loss, filtering low-information tiles, and oversampling informative good and fair tiles.
The public ML4seeding repository contains notebooks for orthomosaic metadata review, pseudo-orthomosaic construction for A10 and A58 image segments, image tiling, grid overlays for annotation, U-Net training, predicted-mask stitching, drone flight-path preparation, and operations-research post-processing concepts.
The current finding is that the training pipeline is usable as a baseline, but future performance gains will depend heavily on data quality, annotation consistency, and collecting more examples of the minority good and fair microsite classes.
The longer-term decision-support goal is to turn georeferenced prediction layers into operational planning inputs, including candidate planting points, spatial thinning, and drone path or mission-planning models.
Status
- Ongoing applied research and software project
People And Roles
- Elaheh Ghasemi: past postdoctoral researcher and machine-learning pipeline lead
- Salar Ghotb: past postdoctoral researcher and project contributor
- Kailey: TRANSFOR-M program student contributor
- Gregory Paradis: FRESH lead and supervisor
Collaborators And Partners
- FlashForest
- Innovate BC / IgniteBC
Outputs
- ML4seeding public GitHub repository
- Reproducible U-Net semantic segmentation notebooks
- Orthomosaic review, pseudo-orthomosaic construction, and image-tiling workflows
- Evaluation workflow using accuracy, mean IoU, foreground mean IoU, and per-class IoU
- Predicted segmentation-mask and georeferenced output workflow prototypes
- Drone flight-path and operations-research post-processing prototypes