Castana
Login Create account

Model Limitations

Understanding the strengths and constraints of AeroCastaña

Note: The AeroCastaña pipeline is a decision-support tool. While highly accurate, model outputs should complement local knowledge and field validation.

Key Considerations

  • Fruit detections rely on known Brazil nut (Castaña) tree locations. Unknown trees are not currently automatically detected, though experiments are ongoing.
  • Environmental conditions, such as dense shadows or overlapping canopy, may lead to missed or false detections.
  • The pipeline estimates per-tree fruit counts, but conversion to actual harvest volume is still experimental and refined each year with field data.
  • Outputs should be used alongside field expertise to make management decisions.

Technical Model Details

For those interested in the technical side, AeroCastaña uses the YOLO11x architecture, trained via transfer learning on high-resolution drone imagery. Models were trained to detect small, partially occluded Brazil nut fruits (cocos) in canopy images.

Metric Value Explanation
Model Parameters 56,828,179 The "brain" of the AI – more parameters mean the model can capture more subtle patterns in the images.
GFLOPs 194.4 A measure of computational effort; higher numbers indicate more processing power needed.
Training Images 14 The number of unique image tiles used for validation (training data is thousands of tiles, this is a summary sample).
Fruit Instances 843 The total number of fruits labeled for the model to learn from.
Precision 0.82 Of all fruits the model predicts, 82% were correct.
Recall 0.726 The model correctly detected about 73% of all fruits in the images.
mAP50 0.832 A combined metric of precision & recall at a 50% overlap threshold; higher is better.
mAP50-95 0.463 Average precision over stricter overlap thresholds; shows performance on harder cases.

Full training logs, hyperparameters, and datasets are available in our GitHub repository for those who want to reproduce or evaluate the model in depth.

Future Improvements

  • Automatic detection of unknown Castaña trees without requiring pre-existing GPS points.
  • Enhanced small-object detection to reduce false positives and capture fruits partially hidden in dense canopy.
  • Integration of multi-year data to refine per-tree yield estimations.
  • Optional modeling of new prospective concession areas to help prioritize forest protection.