Computer vision and multispectral imagery for soil analysis

AI soil analysis: where field diagnostics breaks down

Manual sampling is simple: one or two points per several hectares and one visit per season.

The accuracy of field measurements, however, rests not on the instrument but on the heterogeneity of the plot — within one field, zones with different particle-size distribution, water-holding capacity, and residual nutrients coexist.

An average from sparse points describes none of them, yet it is used to build an application map for the entire area.

The second failure is the incompatibility of measurements between shifts and contractors. Different sampling depth, different moisture at the time of measurement, different containers, and different time to the laboratory produce scatter that is later read as dynamics.

Without a reference sample and a unified calibration procedure, two reports from adjacent months cannot be compared: the trend turns out to be noise, and manual sampling errors accumulate in the database and carry over into decisions.

The third is decisions based on incomplete data. Some points are lost, some fields are covered by clouds at the time of imaging, and some plots were simply not walked.

Gaps are filled by interpolation, and zones appear on the map that do not exist in the field; material is applied according to them, and overapplication occurs where it is not needed.

Soil condition diagnostics breaks not on the model but on the input data: sparse grid, mismatched calibrations, coverage gaps.

AI soil analysis stops being guesswork only when measurements are tied to coordinates, reconciled by moisture and depth, and the map honestly shows where data is missing. We break down how this is addressed in practice.

Measurable effect: speed, accuracy, coverage

We calculate the effect before start: we record baseline values for three metrics — field survey time, share of area under regular monitoring, and convergence of the condition map with control samples.

After commissioning, we compare against the same metrics, so the debate "it got better" is translated into numbers rather than impressions from a demonstration.

What changes in numbers

Field survey speed is determined by the survey route, not the number of visits. One pass along the contour provides continuous coverage for the plot, whereas manual sampling stretches the work over weeks, and some fields are not included in the plan at all.

We check the accuracy of condition maps on control samples that did not participate in model building.

If the discrepancy for key indicators exceeds tolerance, the plot is re-surveyed in another mode — therefore the share of repeat measurements decreases rather than grows along with the area.

Survey area coverage no longer rests on labor intensity: the entire contour is under monitoring, not selected 10–15% of fields. Ground samples remain for calibration and disputed zones, not as the primary way to obtain data.

The currency of soil data rests on regular surveying: open multispectral imagery Sentinel-2 supplements field flights where freshness is needed between campaigns.

The error class "decision made from last year's map" goes away — plot condition is visible for the current season.

Comparison of approaches to surveying and data processing

The choice of survey method is determined by three parameters: field area, required map detail, and update frequency. A laboratory sample gives reference composition but describes a point, not a field.

Remote methods cover the entire area, but their readings are tied to ground measurements — otherwise the map remains pretty but unprovable.

In practice, one chooses among four working schemes and their combination.

The hybrid scheme usually wins: satellite imagery of fields provides a regular background and overview, drone surveying refines problem zones in a matter of hours, and manual sampling calibrates the model and confirms conclusions.

After calibration, the discrepancy between forecast and laboratory decreases severalfold — the map becomes the basis for decisions, not an illustration. Describe the area and monitoring frequency: we will return with an assessment and survey plan.

Implementation stages: from soil audit to model handover

The order of work is built from data, not from the model: first we understand what is already known about the plots, and only then do we select sensors and survey modes.

This order saves weeks on re-surveying and prevents training the model on features that are not present in the field. Below are the stages by which field monitoring implementation proceeds.

1. Plot and data audit. We check field boundaries, crop history, existing surveys, and yield maps. We record exactly what needs to be measured and at what frequency.
2. Plot survey scheme. We select the carrier, altitude and frame overlap, spectral channels and ground control points, and tie surveying to the phenological phase. The result is a route reproducible in the next season.
3. Collection and labeling of field data. We bring images to a single projection and calibrate against a reference, then label field contours, suppression zones, and sampling points. A second specialist checks the labeling — discrepancies are resolved before training.
4. Model training and tuning. We train the model on the labeled set and separately tune sensitivity thresholds for the crop and soil type. For each field, we save quality metrics to see the boundaries of applicability.
5. Validation on control plots. We run the model on fields not included in training and compare the result with laboratory samples. Error is calculated for each indicator separately and recorded in the report.
6. Handover of the solution to the client. We hand over the model, survey scheme, processing scripts, and update regulations, and set up map export to your system. We train agronomists and the IT team — the solution remains operational without us.

Documentation, models, and processing regulations

A model without documentation quickly turns into a black box: after a season no one remembers what imagery it was trained on and why it gives a confident forecast on one plot but not on the neighboring one.

Therefore, along with the result, we hand over a package by which your team or another contractor can continue the task.

• Model passport — which survey channels are fed as input, how values are normalized, what confidence threshold is considered operational, and on which plots errors are predictable.
• Image preparation rules — calibration, clipping of cloud cover and shadows, bringing to a single coordinate grid and channel order: a new image goes through the same path as the training set.
• Soil condition maps — layer by layer by indicator (moisture, organic matter, compaction), with problem zone contours, legend, and units of measurement.
• Repeat survey regulations — the season window and crop development phase at which comparison with the previous season is valid, and the permissible variation in survey conditions.
• Instructions for the field team — pre-trip checklist, reference points on the plot, row marking, actions under cloud cover and schedule shifts.
• Model update regulations — when accumulated data is sufficient for retraining and how it is verified that forecast quality has not dropped.
• Version log — what changed, on which set it was tested, what metrics were obtained, and to which plot a specific version applies.

The package is assembled so that after a season you won't have to reconstruct the logic from correspondence: using these documents, a new engineer repeats the processing, compares the result with reference maps, and continues surveying without us.

Case study: diagnostic accuracy before and after implementation

A farm with fields of different run lengths and heterogeneous fertility: some plots maintained yield, others failed at edges and in depressions. Diagnostics were performed by manual sampling — one composite sample per several dozen hectares, once per season.

Problem zones were found after the fact, when plants were already lagging in development, and individual patches were not included in the sample at all.

What was changed

Switched to regular surveying: multispectral imagery Sentinel-2 was supplemented by flights over disputed squares, images were brought to a single grid and calibrated against ground measurements.

Then — calculation of vegetation indices from spectral channels and classification into zones: stable, suppressed, requiring check. The output is maps for the entire field area, not individual points, and layer-by-layer comparison of images between dates.

What the metrics confirmed

• Coverage: point samples on a fraction of a percent of the area → map across the entire field contour.
• Detection time: problem zones became visible 2-3 weeks earlier, before external signs of suppression.
• Convergence: discrepancy between forecast and ground control decreased from about a quarter to 8-10%.
• Dynamics: the condition of each plot is visible in a single series of dates — a stable trend is separated from seasonal fluctuations.
• Agronomist visits became targeted: zones with discrepancy are checked, not the entire field.

The key in the case is not a one-off image but regularity: it is the series of dates that gives the dynamics of plot condition and allows planning work by zones, and then checking whether the applied decision worked.

How does multispectral surveying differ from conventional?

A conventional camera captures three broad channels — red, green, blue — and mixes them into the familiar picture.

Such surveying confidently answers questions like "where is the field boundary," "what is the leaf contour," "is there a shadow," but plant condition is almost unreadable in it: two areas with different chlorophyll content produce similar color, and the difference drowns in white balance and variable lighting.

Multispectral surveying divides the same part of the spectrum into narrow ranges and adds near-infrared and red edge.

This is where vegetation features appear that are absent from the color image: red edge shift, reflectance ratio in infrared and red channels, depth of water absorption bands.

A suppressed plant changes these ratios noticeably earlier than it becomes visually yellow, so stress is caught at a stage when the decision is still cheap.

With soil, the logic is the same. Organic matter, moisture, salinity, and mineral composition manifest differently in individual spectral channels and are almost indistinguishable in three broad ones.

Therefore, comparison of survey types is not about the "best camera" but about which physical properties of the surface are captured in measurements at all and which remain invisible.

For the model, what is critical is not the images themselves but data consistency: channels co-registered geometrically, radiometric calibration for the changing sun, and tying to ground agrochemical samples.

Without this, features shift between flights, and the network learns from lighting instead of soil condition. Color surveying is not discarded — it is used to build planting masks and geometry onto which spectral layers are overlaid.

Technical concerns before project start

Concerns at the assessment stage usually relate not to algorithms but to the stability of the result in a real field. Below are brief answers on points that most often come up before start. If something does not match your contour, we will address it separately.

What to do with cloud cover and shadows?

Surveying is planned in windows with minimal cover; frames with cloud shadows are rejected using a mask. The map is assembled from several passes over the season — only clean areas are taken.

If the window is not closed, uncovered zones are marked on the map rather than filled in by guesswork.

How are images matched with laboratory samples?

Sampling points are tied to coordinates; each sample receives a zone index and date. Spectral values are taken from the image closest in time to sampling — otherwise moisture and growth phase introduce bias. The output is a correspondence table by which the calculation can be rechecked.

How often is surveying needed?

For fertility maps, two or three passes per season are enough: before sowing and at key growth stages. For zone-specific top dressing, surveying is done more often, but calibration still relies on ground samples — survey frequency does not replace them.

Who owns the data and model?

Data, labeling, and the trained model remain yours and are handed over in open formats with feature descriptions. The cloud environment is set up under your accounts; our access is closed after project handover.

Let's discuss the task: what data to prepare for start

Before the conversation, it is useful to collect what will determine the scope of work: the completeness of source data determines whether we can immediately calculate the stages and avoid extra iterations for clarification.

Below is the minimum set that is usually enough to move from task discussion to a work plan and estimate.

• Plot boundaries — vector field contours in any convenient form: GeoJSON, shapefile, export from an agricultural platform, drive tracks, or corner points. If there are no exact boundaries, we will reconstruct them from imagery.
• Field treatment history over several seasons: crops, sowing and harvesting dates, applied fertilizers and crop protection products. This is what the model compares spectral features against.
• Past measurements: agrochemical analysis reports, sampling points and dates, results for individual fields. Even disparate data provide calibration.
• Image archive, if available: multispectral drone surveys or satellite scenes from past seasons with dates. Not critical — if absent, we will collect them ourselves.
• Related layers: soil maps, relief, moisture zones, data from nearest weather stations.
• People on site: who will confirm boundaries, perform sampling and accompany in the field, who is responsible for agronomy and will accept the result.

From this set, we assemble a work plan: stages of surveying and model calibration, list of output maps, timelines, and labor estimate. Describe the task — we will return with a plan and estimate.