Rain gauges and soil-moisture probes are proven instruments, and nothing here argues for replacing them. What cameras add is a different kind of evidence: continuous, visual, and covering an area rather than a point. Paired with a small machine-learning model running on the device, a camera becomes a sensor that can tell you when it started raining, how hard, and how the surface of a paddock responded over the following days. This article explains what that looks like in practice and where it goes wrong.
Why cameras at all
A tipping-bucket gauge measures one point and fails quietly. A spider builds a web in the funnel, a leaf blocks it, or the mechanism sticks, and it reports zero rain through a storm. A soil probe reports one depth at one location, and installing enough of them to characterise a variable paddock gets expensive and disruptive.
A camera looking across the same area gives you something the point sensors cannot: context. You can see the rain arrive. You can see standing water form in the low corner and drain from the high one. You can see the surface colour of the soil shift as it wets and dries. And when a gauge reads zero during an obvious downpour, you have the evidence to know which instrument to trust. That cross-checking role alone justifies a camera on most sites.
What the models look for
The detection tasks are narrower than general computer vision, which is what makes them tractable on small hardware.
For rainfall, the model looks at change between frames rather than any single image. Rain shows up as streaks against a contrasting background, as droplet impacts and ripples on any standing or collected water, as a change in surface reflectance when a dry surface becomes wet, and as a general loss of contrast in the far field. A reference surface in the frame, something flat and matte with a known appearance, makes this more reliable.
For soil moisture, the signal is slower and mostly about colour and texture. Wet soil is darker than dry soil of the same type, cracking patterns open and close as the surface dries and wets, and the way vegetation holds itself changes with water stress. These are relative signals. A camera can tell you the surface is wetter than it was yesterday and roughly how that is trending; it does not give you a volumetric water content at depth. We treat camera-derived moisture as an index that is calibrated against a probe at each site, and we say so in the reporting.
False positives, and there are many
Anyone who has run a motion-triggered camera on a farm knows the problem. The environment is full of things that look like the thing you are watching for.
- Dust raised by wind, vehicles or stock produces streaks and contrast loss that resemble rain.
- Insects near the lens, especially at night under infrared illumination, look like large droplets.
- Glare and sun angle change surface reflectance hour by hour, which can mimic wetting.
- Animals walking through the frame trigger change detection and can darken or disturb the reference surface.
- Dew and lens fogging wet the reference surface and the lens without any rain.
- Shadows from moving cloud alter soil colour across the whole frame at once.
Reducing these takes a mix of model work and plain engineering. On the model side, we train on footage from the same kinds of sites, including the failure cases, and we tune decision thresholds per site rather than shipping one global setting. On the engineering side, a lens hood, a fixed reference surface, an illuminator mounted away from the lens, and a requirement that an event persist for more than a few seconds remove a large share of the noise before the model ever sees it. What remains is handled by fusing the camera output with whatever conventional sensors exist. A rain event that the camera reports and the gauge does not is flagged for review, not silently accepted or discarded.
Why the processing stays at the edge
It would be simpler to stream everything to a cloud service and run the models there. On a remote site it rarely makes sense.
Backhaul is the first reason. Continuous video from several cameras will exhaust a satellite or 4G data allowance quickly, and the links are not reliable enough to depend on for a sensing function. Running inference on the device means only events, summaries and a few sample frames need to leave the site.
Resilience is the second. If the link drops during a storm, which is exactly when it is most likely to, an edge node keeps observing and keeps a local record. It uploads when the link returns. A cloud-dependent pipeline has a gap in the data at the moment of most interest.
Power is the third, and it cuts the other way. Inference costs energy, so the models are kept small and quantised, they run on hardware with a modest accelerator, and they are scheduled so the heavy work happens when there is something to look at rather than on every frame all day.
Being honest about what it is
Camera-based sensing is a complement, not a replacement. It is strongest at detecting events and trends, at covering area, and at auditing point sensors. It is weakest at absolute measurement. Anyone presenting camera-derived rainfall totals or soil-moisture values without a co-located physical reference is overstating what the method can do, and we would rather explain the limits up front than have them discovered later.
Where Bizix Agritech fits
We build and deploy camera networks with custom models for rainfall and soil-moisture analysis, running on low-power edge hardware and tuned per site to keep false positives down. The models are treated as one instrument among several, calibrated against physical sensors, and the outputs are reported with their limits stated. Designed and supported in Australia.