Skip to content
Agronix

Method and validation

Measured figures, with the data to check them

Every Agronix feature is compared with an external reference before it reaches customers. Here are the results, with date, data used and limits.

How we validate

How to read the figures: precision says how many of the items found are correct, recall how many of the real ones were found, F1 is an average of the two (1 = perfect).

Validations carried out

  • Vegetation indices

    Mean NDVI 0.801 against 0.800 for the reference

    Data and comparison
    Public dataset of an 11 ha winter barley field (DJI Mavic 3 Multispectral, flight of 5 April 2024, 1,236 multispectral images), compared with the report of the photogrammetry software supplied with the data. Area within the boundary 10.78 ha against 10.77; standard deviation 0.105 against 0.106.
    Limit
    One field, one crop and a single sensor.
    Measured on 14/10/2026Go to the feature

  • Thermography

    Map and source photos agree within 0.03 °C

    Data and comparison
    167 radiometric thermal photos from a DJI sample dataset (H30T camera): median difference between the map and the nearest photo 0.03 °C, with 90% of points within 0.17 °C. Against another program's map: 26 cm offset and −0.12 °C.
    Limit
    Sample dataset: validation on one of our own field flights is still to be done.
    Measured on 16/10/2026Go to the feature

  • Points to check

    10 zones out of 10 found, median error 0.4 m

    Data and comparison
    On the barley survey, 10 zones with vigour lowered by 15% inserted at random positions, in two series: all found, each point inside its zone; with no zones inserted, no point further than 25 m from the edge. Weedy rice against the hand-drawn masks of the WeedyRice dataset: precision 88%, recall 71%.
    Limit
    With milder drops, of 8%, it finds only 1 to 3 out of 10.
    Measured on 15/10/2026Go to the feature

  • Plant and weed counting

    F1 0.80 on 3,343 plants

    Data and comparison
    Test photos never used in training: plants in 6 fields of the UAVmulticrop dataset (F1 0.800, total counted −8.1% against the real figure) and 946 weeds from the USU dataset (F1 0.799).
    Limit
    Rice panicles are still weak (F1 0.67) and remain in testing; validation on one of our own fields is missing.
    Measured on 18/10/2026Go to the feature

  • Trees and canopies

    Precision 0.92 and recall 0.85 on olive trees

    Data and comparison
    Olive grove in Meknes (Morocco), public dataset OliveTreeCrownsDb (CC BY 4.0): compared with 2,426 hand-labelled canopies. Heights agree with the authors' point cloud, with a median difference of −0.01 m.
    Limit
    Recall target of 0.90 not reached: some low olive trees are mistaken for ground.
    Measured on 15/10/2026Go to the feature

  • Diseases and pests

    Median accuracy of 94% across 33 models

    Data and comparison
    33 models for diseases, pests and invasive plants, tested on 19,225 photos from public datasets never used in training: from 76% (maize) to 100%. Brown marmorated stink bug searched in whole photos: F1 0.65 on 59 photos from 2022-2023.
    Limit
    Close-up archive photos: in the field the result is an indication, and confirmation remains with the agronomist.
    Measured on 4-19/10/2026Go to the feature

  • Boundaries proposed from the survey

    F1 0.63 on the test tiles

    Data and comparison
    1,139 test tiles from the public AI4Boundaries dataset: precision 0.635, recall 0.617.
    Limit
    It is a proposal: every boundary is checked on the map before importing it.
    Measured on 19/10/2026Go to the feature

  • Irrigation advice

    Evapotranspiration identical within 0.01 mm

    Data and comparison
    Reference evapotranspiration (Penman-Monteith, FAO-56) over 99 days from the Arpa Piemonte station in Vercelli, compared with an independent implementation of the same method.
    Limit
    It validates the calculation: actual water needs also depend on soil, crop and irrigation system.
    Measured on 17/10/2026Go to the feature

  • Volumes and levelling

    Within 1% on a known volume

    Data and comparison
    Cone-shaped stockpile of 418.9 m³ built on the computer: volume calculated within 1%. Levelling on a constructed terrain with a 0.20% slope and a 10 m³ mound: 5 tests out of 5. Canopy volume on constructed shapes: within 3%.
    Limit
    Shapes built on the computer: validation on a stockpile measured on the ground is still to be done.
    Measured on 16/10/2026Go to the feature

Still to be measured

Tests already planned: the results will be published here, even if below target.

  • Positional accuracy of surveys on ground check points: the first test, on a public flight, stayed above the thresholds set (12.2 cm horizontal and 24.7 cm vertical, against 5 and 10 cm) and must be repeated.
  • Post-processed positions (PPK) with the regional GNSS network, on a real flight.
  • Thermography, plant counting and volumes on our own surveys with the DJI Matrice 4T in Piedmont fields.
  • Disease recognition on photos taken in Italian fields.

Surveys with technical value

No software makes a survey “certified”. Accuracy is demonstrated on each job with check points measured on the ground: Agronix compares them with the map and reports the residuals on the survey page. The ASPRS standard (2023) requires at least 30 for full compliance; Italian cadastral filings must be signed by a qualified professional.

See Agronix on your own farm

An online presentation, built around the way you work.

Request a demo