Infrastructure

Watch the cliff between the surveys.

Rock cut tombs, carved facades, and ancient cliffs lose material to heat, wind, and water a little at a time, and then a lot at once. A survey team visits once. We compare repeat imagery, from drones or satellites, and evolve a change detector for each site so the next survey goes where the rock is moving.

The problem

Thousands of sites, one survey, and a clock that does not stop.

Heritage agencies are registering sites faster than anyone can monitor them.

138monumental rock cut tombs at a single ancient site, every one assessed for rockfall, sliding, or weathering risk
14,000+registered archaeological sites in one national register, with no stated monitoring technology behind the register
Oncethe number of times a full drone survey has been flown at one of the most visited sites, as a one off project

A sandstone facade carved two thousand years ago is a structure with a decay rate. Daily heating and cooling open cracks. Wind carries sand that abrades. Rare rain finds the cracks and widens them. Tourism adds vibration and footfall. Geologists can assess a site and rank its risks, and they have. What they cannot do is come back every month to every site. The assessment is a snapshot; the decay is a film.

Repeat imagery is now cheap. A drone can fly a facade in an afternoon and a satellite passes over every few days. Comparing two images of the same cliff for change sounds simple. It is not: the sun moves, shadows move, sand drifts, cameras differ, and a one millimeter crack has to be found in a wall the size of a building. A general change detector flags the shadows. The one that matters flags the crack.

In plain terms

It is like comparing two photographs of your house taken a year apart to see if a crack has grown. If one was taken at noon and one at sunset, almost everything looks different and nothing is. What you need is a way to ignore the light and the dust and see the crack. Then you need to do it for fourteen thousand houses and send the inspector to the ones that changed.

Why the standard approach fails

Change detectors see the sun move. They need to see the rock move.

01

Light and shadow swamp the signal

On a carved facade, the difference between morning and afternoon is larger than the difference between this year and last. Generic change detection flags the time of day.

What we do. The detector is evolved per site on image pairs where the true change is known, so it learns to ignore that site’s light and keep that site’s cracks.

02

Every site is a different material

Sandstone, limestone, mud brick, and basalt weather differently and photograph differently. One model does not fit the register.

What we do. A model family evolved per material and per site, starting from the nearest neighbor in the portfolio rather than from scratch.

03

Real change is rare, so there is little to learn from

Most image pairs show no structural change. Models trained on that learn to say nothing.

What we do. Detection of every known past change event at the site is a hard floor, and the loop weights the rare true changes so the detector commits on them.

04

A flag is not a plan

A list of ten thousand possible changes is not useful to a team of five geologists.

What we do. Flags are ranked by the site’s existing risk assessment and the size of the change, into a survey schedule the team can actually fly.

Same facade, a year apart. A general change detector flags the shadows that moved with the sun and the sand that drifted at the base. The evolved detector ignores all of that and flags the one crack that grew.
General detector Evolved detector 4 flags, 3 of them light and sand 1 flag, the crack that grew
How we solve it

Repeat imagery in. A ranked survey list out.

The same evolutionary loop as our medical imaging, pointed at a cliff instead of a lung.

In plain terms

We make dozens of versions of a program that compares two photographs of the same rock face and marks what changed. We test them on pairs where geologists already know the answer, including pairs taken at different times of day with nothing changed. Versions that flag shadows are thrown out. Versions that miss a known rockfall are thrown out. The survivors are combined and varied, and after a few rounds the program sees the cracks and not the sun. Then it runs on every new flight or satellite pass and tells the team where to go.

Not one model. A population that is bred. Every generation is made from the last: the strongest candidates become parents, their settings are combined, small variations are introduced, and the children are scored on your own cases. Weak lines die out. Strong ones carry forward.
Generation 1Generation 6Generation 14 48 candidates, all differentWeak ones gone, strong ones combinedYour model, chosen by your team
This is an evolutionary algorithm, not a leaderboard. We do not train many models and keep the best. Each generation is bred from the last. The strongest candidates are selected as parents, their settings are combined to make children, small random changes are introduced, and the children are scored on your own scans. Settings that work together survive together.
1. Scoreon your own scans 2. Selectthe strongest as parents 3. Combineparents into children 4. Varysmall random changes Nextgeneration How a child is made Parent A Parent B Crossover: segments from each parent Child Mutation: one setting changed at random Child Each segment is one setting, such as window or threshold
  1. 01

    Build the baseline from what exists

    Past drone surveys, satellite archives, the existing risk assessments, and any photographs with a date. Each site gets a reference model of its facades and slopes, and a record of what has already been documented as changed.

  2. 02

    Align repeat imagery

    New flights and satellite passes are registered to the reference so that the same pixel means the same square of rock across years, cameras, and seasons.

  3. 03

    Evolve the change detector per site

    Candidates are scored on known true changes caught, on false alarms from light and sand, on the minimum change size detected, and on consistency across seasons. Candidates that miss a documented past event are discarded. The conservation team chooses how sensitive the deployed detector should be.

  4. 04

    Rank and schedule

    Detected changes are ranked by the site’s risk assessment, the size and rate of the change, and visitor exposure, into a prioritized survey list. The team flies the top of the list. Every confirmed or dismissed flag feeds the next generation.

  5. 05

    Extend across the register

    Sites with no drone survey start with satellite imagery and a detector evolved from similar material elsewhere in the portfolio. As surveys accumulate, each site gets its own.

What the evolution changes

What counts as change, and what counts as noon.

What we evolveWhy it matters on a carved cliff
Illumination normalizationHow the detector corrects for sun angle and shadow is evolved per site and per orientation of the facade.
Minimum change sizeThe smallest crack or loss worth flagging is a tradeoff against false alarms, chosen by the conservation team from the Pareto set.
Material specific texture modelSandstone grain, limestone pitting, and mud brick erosion each look different. The texture features are evolved per material.
Rare event weightingTrue structural change is rare in the record. The loss weighting is evolved so the detector commits on it.
Documented event floorEvery past rockfall or crack growth on record must be detected in replay. Candidates that miss one are discarded.
Ranking ruleHow risk assessment, change rate, and visitor exposure combine into survey priority is evolved against the team’s own past decisions.
What you get

Every site watched, the right sites visited.

Change maps per facade

Where the rock moved, how much, and since when, on every repeat flight or pass.

A ranked survey list

The sites and faces the team should visit next, with the evidence.

Rate of change over time

Crack growth and material loss tracked as a trend, not a snapshot.

Register wide coverage

Satellite monitoring for sites that have never been flown, improving as they are.

Visitor safety input

Changes on faces above paths and viewing points flagged first.

A permanent record

Every image, every flag, and every team decision archived against the site.

Where this fits

Built for agencies with more sites than surveyors.

  • National heritage commissions and site authorities with large registers
  • Rock cut and carved monuments in arid climates with daily thermal stress
  • Destinations scaling tourism onto fragile sites
  • Operators of cliffs, quarries, and slopes above public paths

Finding a growing crack in a cliff the size of a building, across years of different light, is the same kind of problem as finding a small nodule in a chest scan across different scanners. It is the problem of a small true signal in a large, shifting background. We evolve for that signal, and we measure ourselves on the changes that were actually there.

Send us two flights of the same facade. We will show you what changed between them, what did not, and what the team should look at first.
Partner With ALBA