I’m with you, @pknoot, I’m fine with accurate (enough) with a higher price point given the climbing price of water. @franz have you found ones that are more accurate than others?
@franz Sensors are the missing link in creating the necessary ground truth for training the system on actuals rather than relying on estimates. Accuracy is certainly a big issue (garbage in, garbage out).
As @pknoot alluded to, given an initial on site calibration, maintaining accuracy is a matter of time series analysis and spatiotemporal analysis across sensors. The predictive regression model would take weather and flow meter data as input and sensor data as output. The perfect sensor would have 3 or more independent readings per device. 3 spikes would give 3 whereas 4 would give 5 assuming each pair could be electrical isolated. Spikes at different heights would be even better.
We’d be alerted to sensors that are out of distribution with regards to the prediction model. Mitigation would be ensuring the sensor is still calibrated and working properly (against ground truth [1] and/or a hand-held moisture meter) or making sure the sensor is properly seated to take accurate measurements.
At any rate, that’s how I’d build it.
[1] Soil samples (provided or local samples measured by volume or weight) with different moisture levels determined by volume.