Swara's Weekly Update for National Loon Center: Turning Drone Footage into Conservation Data

Counting wildlife across vast habitats by hand doesn't scale. For the National Loon Center, the alternative is a computer vision pipeline that can look at drone footage and automatically detect loons, nests, and surrounding ecology, turning raw imagery into structured data researchers can actually use.

Machine learning versus hand-coded rules

The foundation of the approach is machine learning itself: teaching a computer to recognize patterns from examples instead of programming every rule by hand. Rather than writing out rules for what a loon looks like, the model is given many loon images, learns the patterns in them, and uses those patterns to make predictions on new, unseen images. That shift, from hand-coded rules to learned patterns, is what makes the system generalizable across the wide variety of conditions drone footage actually captures.

Object detection: the what and the where

The specific kind of machine learning at work here is object detection, which answers two questions at once: what is the object, and where is it. Instead of simply labeling an image as containing a bird, the model identifies the bird specifically, draws a bounding box around it, and attaches a confidence score to that prediction. For this project, that translates into detecting and counting loons directly in aerial imagery, rather than requiring a person to review every frame.

Annotation in CVAT

Before a model can do any of that, it needs labeled examples, and building those labels is called annotation. The process is straightforward in principle: take an image, find the object of interest, draw a bounding box around it, and assign it a label such as loon. This project annotates using CVAT, a computer vision annotation platform, uploading images and creating labels like loon, nest, fish, and vegetation, then marking each object directly in the image. The quality and variety of these annotations matter directly, since they're what the model learns from. Once enough labeled data exists, the annotations are exported in a format usable for training a detection model.

Why YOLO

The model family behind the detection itself is YOLO, short for "you only look once." YOLO is designed to detect objects efficiently, finding both an object's location and its class in a single pass over the image, rather than scanning repeatedly. The goal is that once enough high-quality annotated loon images exist, a YOLO model can be trained specifically for this use case, predicting where loons appear in new, unlabeled drone imagery.

The feedback loop

The full pipeline is a cycle rather than a one-time build: collect good images, prepare and annotate them in CVAT, use those annotations as training data, train a YOLO model, evaluate it, and then use its own predictions to help annotate new images faster. Better data leads to a better model, and a better model speeds up annotating the next batch of data. That loop is the actual engine of the project, and the end goal isn't just detecting an object in a single image, it's turning raw drone footage into structured conservation information researchers can rely on.

Key takeaways

• Machine learning here means training on labeled examples rather than hand-coding rules for what a loon looks like.
• Object detection identifies both what an object is and where it is, using a bounding box and a confidence score.
• CVAT is the annotation platform used to label loons, nests, fish, and vegetation directly in drone images.
• YOLO ("you only look once") detects an object's class and location in a single pass, making it efficient for this kind of imagery.
• The pipeline runs as a feedback loop: better annotations improve the model, and the model's predictions speed up future annotation.

Who this is for

This is a practical introduction to applied computer vision for anyone interested in conservation technology, drone-based ecological monitoring, or the CVAT and YOLO tooling itself. It's part of the ongoing work by Humanitarians AI Fellows building the National Loon Center's computer vision system from the ground up.