Soil compaction detection
On this page 10
Read the tramlines the crop is drawing for you
Growth patterns, imagery and machine data together show where heavy traffic has closed the soil up.
| Category | Detect |
| Slug | soil-compaction |
| Cadence | Once or twice per season, best read at peak biomass |
| Payload | Multispectral, high resolution RGB, machine telemetry import |
| Needs a visit first | No |
The problem
Compaction is invisible from the surface and expensive to find with a penetrometer, so it usually gets diagnosed years late by a crop that keeps underperforming in the same strips. Headlands, gateways and the lines a full trailer took in a wet harvest are the usual suspects, but suspicion is not a map.
How it works
| Step | What happens |
|---|---|
| Overlay | Crop performance from multispectral flights is laid over the machine traffic your terminals recorded |
| Correlate | Persistent underperformance that follows traffic lines is separated from underperformance that follows soil type |
| Rank | Suspected areas are ranked by size and yield cost so the penetrometer goes to the worst first |
| Confirm | Your ground measurements are recorded against each area and close the loop on the model |
What you get
- Suspected compaction map
- Ranked verification list
- Traffic intensity layer
- Remediation plan by area
What the dashboard measures
- Suspected area as a share of the field
- Overlap with recorded traffic
- Yield gap inside suspected areas
- Areas confirmed on the ground
Where it matters most
Winter wheat, sugar beet, potatoes, cooperatives.
Sugar beet is the clearest case: a heavy crop moved by heavy machines, so last year's lifting shows up in this year's stand.
Build it
No new hardware at all. This capability is a data integration and a correlation, running over imagery you already fly for fertilisation and disease. It is the cheapest capability in the catalog to add and one of the more defensible, because the customer's own machine logs are half the evidence.
The machine data problem
Getting traffic data out of a farm's terminals is the real work.
| Source | Format | Route in |
|---|---|---|
| ISOBUS task controller | ISOXML TLG log files on the terminal's USB stick | Parse directly. This is the open path |
| John Deere | Operations Center | Their developer API, with the customer's consent |
| CLAAS | Telematics | Their API |
| Trimble, Müller, Kverneland | Varies | ISOXML export is usually available |
| Nothing at all | Fall back to inferring tramlines from the imagery itself |
The fallback is worth building first: tramlines and headland turns are visible in the imagery, and a Hough transform over the vegetation index finds them without any telemetry integration. That gets the capability working on farms whose machine data you cannot reach, which is most of them.
The correlation, and its honest limits
suspect(pixel) = persistent_underperformance(pixel)
AND follows_traffic(pixel)
AND NOT explained_by(soil_texture, elevation, wetness)Every term is a raster. The third is what stops you selling a soil map as a compaction map, and it needs the elevation model from field mapping plus whatever soil data the farm has.
Be explicit that the output is a ranked list of places to put a penetrometer, not a diagnosis. Compaction is confirmed in the ground. The value is that the farm tests 6 places instead of 60, and the confirmations feed back as labels.
In-house software
| Stage | What we run | Licence | Replaces |
|---|---|---|---|
| Multi-year performance | Per-field index stack, normalised per flight | rasterio BSD | |
| Traffic layer | ISOXML TLG parser, then rasterise the swaths with the machine's width | own code | telematics platform modules |
| Tramline inference | Hough transform over the index raster, no telemetry needed | OpenCV BSD | |
| Confounder removal | Regress out elevation, wetness index and soil texture | scikit-learn BSD | |
| Ranking | Connected components, scored by area times yield gap | scikit-image BSD | |
| Ground truth loop | Penetrometer readings recorded per area, closing the model | this repo |
Writing an ISOXML TLG reader is a couple of days of work and it unlocks
precision fertilisation's as-applied
reconciliation and
sustainability reporting at the same
time. It is the single highest-leverage parser in the platform. See
Prescriptions.
Cost efficiency
- Zero marginal hardware cost. Sell it as an add-on to a fertilisation subscription, not as a flight.
- Read it at peak biomass. One flight at the right growth stage beats four at the wrong ones. For cereals that is flag leaf to early grain fill.
- Use the free archive for the "persistent" part. Five years of Sentinel-2 at 10 m from Copernicus establishes which strips underperform every year. The drone flight resolves what the satellite cannot: whether the pattern follows a wheel track or a soil boundary.
- Sell the remediation plan, not the map. Subsoiling costs real money per hectare and is damaging where it is not needed. "These 3.4 ha, to 35 cm, after harvest" is a decision. A red raster is not.