Verdant — Aerial Afforestation Survey

Land-restoration and greening programmes need to know, across large areas, how many planting pits were dug and how much vegetation actually took. Doing that by hand from drone footage is slow and inconsistent: pits, young plants and bushes are small, look similar to the sand around them, and appear in the thousands across many tiles. There was no off-the-shelf model for these classes, so it had to be trained from scratch on the client's own imagery.

Detect planting pits and vegetation (plant and bush) in drone imagery of arid land reliably enough to support afforestation monitoring, and stand up the full pipeline (annotation, dataset, custom training and inference) so the model can be retrained and rerun as new imagery arrives.

Drone tiles were labeled for three classes (pit, plant, bush) and organised into a versioned dataset on Roboflow, with augmentation and preprocessing to balance the classes and harden the model against lighting and terrain variation. A YOLOv11 detector was trained on that dataset and evaluated with the standard object-detection diagnostics: box / class / DFL loss curves, precision and recall, mAP50 and mAP50-95, and a per-class confusion matrix. The trained weights then run inference over new aerial imagery, drawing class and confidence boxes on each detected pit, plant and bush.

1. 1Drone tiles of the site are collected and labeled for pit, plant and bush
2. 2The dataset is versioned on Roboflow with augmentation and preprocessing
3. 3A YOLOv11 detector is trained on the exported dataset
4. 4Training is evaluated: loss curves, precision / recall, mAP and confusion matrix
5. 5The trained model runs inference over new aerial imagery, boxing each class with a confidence

• Custom YOLOv11 object detector (Ultralytics), three classes
• Roboflow dataset: annotation, versioning, augmentation and export
• Evaluation suite: mAP50 / mAP50-95, precision / recall, confusion matrix
• Class-balancing and preprocessing for small, low-contrast aerial objects
• Inference over drone tiles with class and confidence overlays

• Full pipeline delivered: annotation, training and inference
• Custom-trained on the client's own drone imagery
• Three afforestation classes: pit, plant and bush
• Roboflow-managed dataset, so it can be re-versioned and retrained
• Evaluated with standard detection diagnostics
• Runs over aerial tiles with per-object class and confidence

Data lives as a versioned Roboflow project (labeled drone tiles, augmentation and preprocessing), which exports in YOLO format for training. A YOLOv11 model is trained with Ultralytics and produces the usual run artifacts (weights plus loss, precision / recall, mAP and confusion-matrix plots). The exported weights load into a Python inference step that scores new drone tiles and writes class-and-confidence boxes back onto the imagery for review.

Verdant is delivered as a trained model plus an inference pipeline: it consumes drone tiles and produces annotated imagery, boxing each detected pit, plant and bush with a confidence, so afforestation coverage can be reviewed tile by tile and aggregated across a site.

• Roboflow (dataset, annotation, augmentation, export)
• YOLOv11 / Ultralytics (training + inference)
• Python + OpenCV (inference and overlays)
• Standard YOLO training artifacts (metrics, confusion matrix)

Training runs on a GPU with Ultralytics, and the exported weights run inference as a Python step over new drone imagery. Because the dataset is versioned on Roboflow, the model can be retrained and redeployed as new tiles are labeled, which keeps it current as the restoration site changes.