Feature fileAnnual refreshAccess by pilot
An imagery feature for your valuation model.
One embedding per property, learned from the parcel and its surroundings, in a vector file keyed to your record IDs.
01Delivery
What is in the delivery
- Deliverable
Per-property embedding file
- Format
- vector file, keyed to your record IDs
- Contents
One row per property in the list you submit. Each row holds your record ID and a fixed-length vector learned from imagery of the parcel and a buffer around it.
Join it to the training table on the ID and treat the columns as one feature block. Exterior view only. The same encoder runs over every property in a file, so vectors compare across your whole book. Once a year.
- Deliverable
Coverage and quality flags
- Format
- per-row flags
- Contents
Every row says whether the property had a clean pass. Cloud, shadow or canopy over the lot sets a flag, and that vector may be lower confidence or arrive late in the cycle.
Filter on the flag at training and scoring time. Nothing is quietly filled in.
- Deliverable
Structure note
- Format
- written note
- Contents
A short document on how the vector is laid out, what it was learned from, and how to use it as a model input: scaling, dimensionality, and what the dimensions do and do not mean.
They are learned directions, not named attributes, so no column is "pool" or "lot condition". Written so it can sit in your model inventory entry.
- Deliverable
Annual refresh
- Format
- new dated file each year
- Contents
A new file for the same record IDs, timed to land before your refit. Each file is dated, so with two or more you can build the feature as a series and see change over time.
02Method
How it works
01
Input
You send record IDs, each with a coordinate or parcel geometry, and the area they cover. Before anything is computed we tell you which properties fall outside the imagery.
02
Capture
The latest very high resolution RGB pass is selected for each parcel, with a buffer around it for the street and neighbors.
03
Encode
One fixed encoder turns each parcel image into a vector. No per-region tuning, so a row in one county means the same thing as a row in another.
04
Return
The vector file comes back with the flags and the structure note. You join it, refit, and run your own validation on your own hold-out sales.
03Limits
Limits
- Condition
L1Exterior only
- Effect on the file
Condition, finishes, kitchens and furnaces are not visible from above, and nothing in the file should be read as them.
- Condition
L2Lift is yours to measure
- Effect on the file
This is one more input. Whether it moves your MdAPE or PPE10 depends on your model, your segments and your hold-out set. We promise no lift. Test it on sales the model has not seen.
- Condition
L3Resolution floor
- Effect on the file
At 0.5 m a pixel, a house, a driveway, a pool and a stand of trees resolve cleanly. A missing shingle row or a cracked walkway does not. The vector carries the broad look of the parcel and its context, not a defect list.
- Condition
L4Cloud, shadow and canopy
- Effect on the file
A cloudy pass, a long shadow or a canopy over the roof makes that property's vector less reliable, or late. The row carries a flag.
- Condition
L5Annual, not current
- Effect on the file
The image is a once-a-year view. A renovation or demolition after capture shows at the next refresh, not before.
- Condition
L6Not a valuation
- Effect on the file
The vector is an input. The value stays with your model.
04Reason
Why this exists
Your model sees living area, lot size, year built and the comps. It does not see the overgrown side yard, the pool, or the rail line behind the back fence. Those are the properties where model value and later appraisal drift apart, and you hear about it afterward.
Building the feature in house means licensing imagery, training a vision model and then defending it through model risk review. Meanwhile the model has nothing to tell two houses on one street apart by how they look.

05Fit
Who it's for
- Data science leads at AVM vendors who want to test an imagery feature against their own validation set.
- Lender analytics teams that run their own valuation model and want parcel context in it.
- Teams with a feature store, a refit cycle and a validation process, who can join a vector file on a record ID.
Who it's not for
- Anyone looking for a finished valuation, an appraisal or a condition grade. This is an input, not an output.
- Teams that need interior condition. Imagery from above cannot supply it.
- Teams without a model to put it into. The file only helps if you can retrain and validate.
06Queries
Questions from model and procurement reviewers
Does this replace an appraisal?
No. It is an input to your own model and produces no value.
Does the embedding capture interior condition?
No. Only what is visible from above: the exterior parcel and its surroundings.
How do we validate lift?
On your side. Add the vector, refit, and compare error on a hold-out set of sales the model has not seen. The structure note explains the layout so the test is clean. The result is yours.
What happens with cloud or occlusion?
The property is flagged, and its vector can be lower confidence or late. Filter on the flag if your model wants clean inputs only.
How often is it refreshed?
Once a year, as a new dated file for the same record IDs.
What do the dimensions mean?
They are learned, not named. No single column is "roof" or "pool". Use the vector as a block and check feature importance in your own model.
How does it fit model governance?
The structure note says what the vector is, what it was learned from and what it cannot see, so your reviewers have a plain description of the input. Approval follows your own process.
Who owns the vectors, and what about the imagery?
Use and ownership terms for the vectors are settled in the contract before the pilot starts. The imagery stays under its own license. You receive vectors, not pictures.
How is a pilot priced?
It is scoped on its own: property count, coverage area and how many refreshes you want to test. You get a written scope to take through procurement.
07 Routing slip