False Positives and False Negatives in Vegetation Screening: What a Co-op Line Walk Taught Us

A false positive sends a crew to a span that turns out to be fine. A false negative marks a span as fine when it isn't. Satellite vegetation screening produces both. Most false positives come from a short list of causes: the cleared strip is narrower than one pixel, something green under the line isn't a tree, the mapped line isn't quite where the real one is, or the images being compared were taken under different conditions. Most misses come from dead trees, which aren't green, from trees just outside the strip, and from imagery taken when the leaves are off.

You can't get either one to zero from orbit. You can cut both a lot, and you can set things up so your crews learn quickly where the screening is right on your system and where it isn't. Here's what we learned when a small western distribution co-op walked a stretch of line we had screened, and what we changed because of it.

What counts as a false positive in vegetation work?

In crew terms, a false positive is a truck rolled to a span that didn't need anything. A false negative is a span the screening called fine that had a tree in it.

False positives cost fuel and an afternoon, and if there are too many, people stop trusting the list. False negatives cost more, because the tree nobody looked at is the one that takes the line out. A useful screening keeps false positives low enough that crews keep using it, without hiding the spans that need a visit.

What did the line walk show?

We screened part of a small co-op's system with Sentinel-2 satellite imagery. Their staff walked a stretch of it, photographed 16 poles and sent the photos back with notes, so we could check the screening span by span. Two kinds of false positive showed up, and one result was right in a way we hadn't set out to test.

Forest beside a narrow strip

In the forested sections the cleared strip was narrow and well kept. Sentinel-2's sharpest pixels are 10 m, about 33 feet across, which is close to the width of the whole strip. Most pixels along the line mixed cleared ground with the trees beside it, so spans with a clean strip through thick forest scored as high as spans that had grown in. The screening was right that there were a lot of trees near the line. It couldn't say whether the strip itself was clear.

Green that isn't trees

On another part of the system, a span with no trees anywhere near it ranked high because it crossed irrigated cropland. NDVI, the greenness index most satellite screening uses, measures how green and leafy the ground is, not what's growing there. A watered crop in midsummer can read greener than a stand of pine.

A change it caught

We also compared the same season two years apart. One stretch showed a sharp drop in greenness that nobody had told us about. The co-op confirmed they had replaced poles and done heavy cutting there between the two images. Comparing a line with itself over time sidesteps some of the problems above, because the forest beside a narrow strip is there in both years.

Why does satellite screening flag spans that are fine?

The strip is narrower than a pixel. Many distribution rights-of-way are 20 to 40 feet wide. At 10 m, a satellite reading describes the corridor and its edges together, not the strip on its own.

Green that isn't woody. Irrigated crops, pasture, lawns, golf courses and wetlands can all be greener than trees in summer.

The mapped line is off. If the GIS line sits 10 or 15 feet to one side of the real one, which happens on systems digitized from older maps, the screening measures the forest edge instead of the strip.

Shadows. Tall trees throw shadows across a narrow strip, and shadow changes how a pixel reads. In the aerial images we tested, tree shadows reached 5 to 10 m across the strip.

Images that don't match. No two satellite passes or aerial flights are taken under the same light. In one pair of aerial images we compared, the same stand of trees read 0.53 on the greenness scale in one year and 0.31 in another. Nothing had changed on the ground. Comparing raw numbers across those years would have suggested the forest was dying. The two flights were also up to 2 m out of alignment, which on a narrow strip is enough to shift the measurement into the forest edge.

Why does it miss real problems?

Dead and dying trees. A dead tree has lost its green, so to a greenness index it looks like open ground. Standing dead trees beside the strip are among the trees most likely to come down on a line, and greenness screening is weakest exactly there. We've tested spotting grey crowns in high-resolution aerial imagery, and it isn't reliable enough yet to count on. Patrols still matter for this.

Trees outside the strip. A screening that only measures the cleared strip says nothing about a tall tree 15 feet outside it. That's why LineGuard measures one pixel beyond each edge of the corridor. It's also part of why forested spans score high, which is the trade-off the line walk showed.

Height. Satellite imagery shows how much vegetation is there, not how tall it is. Young regrowth and mature trees can read about the same. Clearance to the wire takes LiDAR or eyes on the ground.

Leaf-off and snow. Deciduous trees in late fall or early spring barely register, and snow hides what's under it. A screening from the wrong month will show a clean line.

Small things. A single tree inside a 10 m pixel of grass may not stand out from its neighbours.

How do you cut false positives and false negatives?

1. Use leaf-on imagery from the same window each year. Mid to late summer works for most of the US. Comparing July with July is fair. Comparing July with October isn't.
2. Check your line against imagery before screening. Lay your GIS lines over an aerial photo and fix any stretches that sit off the real line. It takes an afternoon and removes a whole class of false positives.
3. Measure trees, not just green. A land cover map separates trees and shrubs from crops, grass and water. After the line walk we changed LineGuard to count greenness only on pixels mapped as tree cover or shrubland, remove mapped roads and cloud, and report spans with no trees or shrubs near the line as not ranked, instead of ranking a crop field against a forest.
4. Compare a line with itself, and match the images first. Change over time is often more telling than one year's reading, but only after the two images are adjusted to the same scale and lined up. Then ignore small changes. In our aerial tests, after matching two flights, changes of less than about 10 percentage points in how green a span's strip was were within the noise.
5. Use sharper imagery where strips are narrow. In the US, USDA's National Agriculture Imagery Program (NAIP) flies aerial imagery at about 0.6 m, or 2 feet, every few years. At that size a 30-foot strip is about 15 pixels wide, so you can see the strip and the forest edge as separate things. We're testing it on narrow forested strips now.
6. Record what crews find. The only way to know your own false positive rate is to write down, at each flagged span, whether it needed work, was fine, or wasn't yours to deal with. Ask crews to note work they find on spans the screening didn't flag too, because that's your miss rate. After one season you'll know which kinds of spans the screening gets right on your system.
7. Keep patrols for what satellites can't see. Dead trees, clearance distances and hardware problems still need people or LiDAR.

Where does satellite screening fit?

Use it to decide where to look first, not where you never need to look. It's cheap enough to run across a whole system every year, which LiDAR usually isn't, so it's good at pointing your patrol and LiDAR budget at the stretches that changed or carry the most trees. It doesn't measure clearance, it can't reliably see dead trees, and on narrow strips it describes the corridor rather than the strip.

For more on how it works, see Satellite Vegetation Monitoring for Electric Utilities. For when LiDAR is worth paying for, see LiDAR Vegetation Survey Cost Per Mile. If you run a co-op, our co-op vegetation management guide covers programs, cycles and costs.