Capture PWA
Phone photos, optional card, context, private GPS
Methods in the open
How phone images, measured spectra and laboratory evidence connect without overstating what they prove.
Capture and release architecture
Phone photos, optional card, context, private GPS
RLS, quarantine, immutable grant
SoilGrids, SSURGO, Open-Meteo
Sanitize, sharpness, optional card, GPS, duplicate
Eligible structured data; photos held pending screening
Different evidence, connected by one sample
Ordinary phones record RGB images, not wavelength-resolved spectra. Extracted soil RGB is uncalibrated, even with the reference card. No printed card is needed. A photograph does not automatically create an extracted optical record or register a library sample. Sample registration is optional and needs your real declared depth.
A separate instrument records reflectance by wavelength. Pairing it with the same physical sample requires verified identity, method, calibration and uncertainty.
Verified assays provide separate reference labels. Biological relationships require actual assays and expert-reviewed terms; images or spectra alone do not establish soil microbial composition.
The local RGB-to-reflectance demo uses invented samples and spectra. It demonstrates the software, not field accuracy. No independently validated soil reconstruction model is provided; real models need verified field pairs and independent spatial and device evaluation.
Existing maps are priors
SoilGrids provides global soil context. SSURGO adds higher-resolution United States context. Open-Meteo contributes elevation and recent precipitation. These sources are stored as weak labels with uncertainty rather than treated as measured truth.
Confirmed laboratory measurements are modeled separately as gold labels. Future estimators must report a basis and confidence interval for every output.
Deterministic photo QA
Orientation and colour profiles are applied to the pixels, then EXIF, GPS, XMP, ICC, comments, and residual segments are removed.
Sharpness and exposure are checked on the candidate soil region. A detected physical card is masked out; without one, the candidate region is the whole frame.
Missing a card alone never fails QA. Marker and patch extraction support geometry and image checks, not calibrated reflectance.
Accuracy and gross prior mismatches can trigger review.
Perceptual hashes flag repeated or near-repeated images.
Each submission settles to pass, fail, or flagged review from its two required photos.
Public by design, private by default
Exact geometry remains protected operational data. Public map and dataset records carry only H3 cells: resolution 8 (about 0.74 km²) where at least three people have contributed, otherwise the coarser resolution 6 or 4 cell. Anonymous public reads are limited to aggregate cells—not per-submission coordinates.
Read the governing specification ↗