TerrAlert Team · 2026-09-09
Every detection system, no matter how well built, makes two distinct kinds of mistakes: flagging something that was not actually a meaningful change (a false alarm, or false positive), and missing something that was (a missed detection, or false negative). Understanding this trade-off, and how to reduce false alarms without simply going quiet, is one of the most practically important things to understand about any monitoring tool.
A change-detection model is comparing two images and deciding whether the difference between them is "real" change or noise. Several things can look like change without being the kind of change anyone cares about: a shadow falling differently due to sun angle, a temporary vehicle or object in frame, seasonal variation in vegetation color, a slight misalignment between two images, or sensor calibration differences between passes. A model tuned to catch every genuine change will, almost by necessity, also catch some of these look-alikes.
The opposite failure, missing a real change, happens when the difference between two images is subtle relative to normal noise: a small structure, a change that happened gradually across several image pairs rather than in one obvious jump, or a change partially obscured by cloud shadow or vegetation. A model tuned to avoid false alarms by requiring a very clear signal will, by the same logic, miss more of these subtle real changes.
This is the fundamental trade-off, usually described as precision versus recall: pushing a system toward fewer false alarms (higher precision) generally means missing more real changes (lower recall), and vice versa. There is no setting that maximizes both simultaneously. Any system has to choose a point on that trade-off, whether it makes that choice explicit or not.
Simply raising a confidence threshold trades false alarms for missed detections one-for-one. You catch less of everything. A better lever, where it is available, is corroboration: checking whether an independent source agrees with the finding before treating it as confirmed. If an optical model flags a possible fire and an independent thermal-based fire-detection product (like NASA's FIRMS) also shows activity in the same place at the same time, the combined finding is far more trustworthy than either source alone, because the two sources are wrong in different, mostly unrelated ways. A shadow that fools an optical model has no reason to also show up as thermal activity.
This is a meaningfully different strategy from just tuning one model more conservatively: it reduces false alarms without proportionally increasing missed detections, because it is not throwing away borderline cases. It is checking them against a different kind of evidence before deciding.
This is really a special case of a broader point. See Why Clear Information Matters During Important Decisions for why a precisely-scoped, checkable claim about accuracy beats an impressive-sounding but vague one.
Why won't a vendor just quote one accuracy number? Because a number measured on a benchmark dataset often does not transfer to arbitrary real-world locations, and quoting it as if it does is misleading. A system's real-world precision depends heavily on the specific terrain, imagery quality and change type involved.
Does corroboration slow alerts down? Slightly, in exchange for reliability. Waiting for or checking a second, independent source takes marginally longer than acting on the first signal alone, which is the honest trade-off behind faster-but-noisier alternatives.
TerrAlert corroborates optical detections against independent sources, FIRMS, GLAD, RADD, and Sentinel-1 radar, before most alerts fire. The per-area policy is explained on the FAQ.
All posts