TerrAlert Team · 2026-09-09
Satellite constellations now photograph most of the Earth's surface every few days. That is an enormous amount of imagery, far more than any team could review by eye, even for a modest number of areas. The problem monitoring software actually solves is not "can we get the picture" (that part is largely solved). It is "how do we find the handful of things in all this data that a person actually needs to see."
Consider watching just ten areas of interest, each re-imaged every few days. Over a year, that is well over a thousand individual images to review, most of them showing nothing unusual and some of them unusable due to cloud cover. Reviewing that by hand does not scale past a small number of small areas, and it gets worse fast as the number of areas or their size grows. This is the actual reason automated change detection exists, not because computers are inherently better at seeing than people, but because they can tirelessly compare a thousand image pairs and only surface the handful worth a person's attention.
A single filtering step is rarely enough to get from "all available imagery" down to "the few things worth reviewing." A well-designed pipeline filters in stages:
Each stage discards the large majority of what came into it. That is by design. The goal of the whole pipeline is to go from "everything" to "the short list," and a pipeline that surfaces too much has simply moved the original volume problem one step downstream instead of solving it.
It is tempting to think that adding more data sources, more satellites, more bands, more frequent imagery, automatically makes a monitoring system better. It helps only if the filtering keeps pace. More raw data with the same filtering just means more noise reaching the same overloaded reviewer. The actual bottleneck in most monitoring workflows is attention, not information, and the systems that help the most are the ones that respect that.
A well-designed monitoring product does not hand you a spreadsheet of every pixel that changed. It hands you a short, dated list: this area, this category of change, this confidence level, this size estimate, each item worth a specific decision (investigate further, dismiss, escalate). If reviewing your alerts feels like a new data-analysis job on top of your actual job, the filtering step somewhere upstream has not done its work.
Does more monitored area mean more noise? Not if the filtering pipeline scales with it. The whole point of automated filtering is that it does not get more tiring as volume grows, unlike manual review.
How do I know if a system's filtering is any good? Ask what proportion of alerts you receive turn out to matter on investigation. A system with well-tuned filtering keeps that proportion high.
See how TerrAlert structures this pipeline in How TerrAlert Works.
All posts