What Computer Vision Does
Classification assigns a label to an entire image, detection locates objects with boxes, segmentation identifies pixels belonging to a class, and tracking links an object across frames. A model does not understand an object in the same way as a specialist. It learns statistical patterns and produces a probability or estimate that may be reliable only under conditions similar to the training data.
Workflow, Environmental Applications, and Annotation
The workflow includes problem definition, data collection, cleaning, annotation, separation into training/validation/test sets, model training, evaluation, and deployment. Every transformation — resizing, compression, colour correction, cropping, georeferencing — can affect the result, so source-file versions and preprocessing steps should be reproducible. Computer vision can help count visible waste, classify land cover, detect clearing, assess vegetation, map water, identify smoke or damage, and select images for specialist review. It is especially useful as a filter for large data streams, ranking images by probability of an event while final confirmation, causal interpretation, and legal assessment remain with competent people. Annotation requires clear rules on what counts as an object and how to label partial visibility, shadows, mixed materials, and uncertain examples.
Metrics, Error Types, and Domain Shift
Overall accuracy can be misleading when the hazardous class is rare. Detection and segmentation are evaluated using precision, recall, F1, overlap measures, false alarms, misses, and per-class performance. The cost of an error depends on the use case: in early warning, a missed event may be more serious than an extra inspection; in an automated enforcement decision, a false alarm may have major consequences. Thresholds should reflect risk and the role of human review. A new camera, altitude, illumination, season, dust, snow, or lens damage changes image statistics. A model may remain highly confident even when it is wrong. After deployment, monitor data distributions, uncertain cases, regional errors, and changes after updates.
Human Oversight, Hazardous Materials, and Responsible Deployment
For environmental, safety, and legal decisions, a model should support rather than replace an accountable specialist. The interface should display the source image, marked region, confidence, model version, and known limitations. Users should be able to reject output, record reasons, and escalate difficult cases. Critically, an image may reveal a shape, label, colour, or context resembling a material, but it does not confirm chemical composition. Asbestos, lead, mercury, reagents, and many wastes cannot be reliably identified from a photograph alone. The system should present a preliminary signal: restrict access, avoid touching the object, and involve competent specialists. An organisation should document purpose, process owner, data sources, legal basis, risk assessment, metrics, limits of use, and data protection — testing the model before release and after every material change.
Computer vision scales image analysis, but it does not turn a photograph into infallible evidence. A reliable system combines representative data, meaningful metrics, monitoring for changing conditions, human review, traceability, and a clearly limited role for the model.