Down the Plot, Up the Stack: A Season of LoRaWAN Sensors on a British Allotment

Allotments are an underrated engineering test bed. They offer mains-free power, intermittent human attention, a hostile moisture environment, and a radio path obstructed by exactly the sort of clutter that datasheets describe as suburban. Between March and October 2026, CavLab instrumented a half-plot on a site in the south-east of England with eight low-power soil sensors and a single solar-powered LoRaWAN gateway, and left it largely to its own devices. This is the field report.

The brief we set ourselves

The requirement was deliberately modest: report volumetric soil moisture at each bed every fifteen minutes, survive a British summer without maintenance, and present the data on a dashboard a plot-holder could read without a manual. Eight nodes covered four raised beds, a greenhouse border, two open beds and one control node left on a shelf in the shed, which turned out to be the most informative sensor of the lot.

Hardware was deliberately commodity-grade. Capacitive moisture probes rather than resistive ones, because the resistive type corrodes within a season in wet soil and quietly drifts as it does so. Each node used an SX1276-class radio on 868 MHz, an OTAA join to a self-hosted network server, an eleven-byte payload at spreading factor 9, and two AA lithium thionyl chloride cells. The gateway sat at 2.2 m on the shed roof, powered by a 20 W panel, a small charge controller and a 12 V 6 Ah LiFePO4 pack.

Range, and what a hedge costs

On paper the link budget was generous. On the plot it was more interesting. The nearest node, forty metres away with line of sight, reported around -71 dBm RSSI and a healthy positive SNR. The worst node — 640 m away at the far corner, behind two mature hedges, a brick shed and a stack of water butts — settled at -119 dBm with an SNR of about -7.5 dB, which is right at the edge of what SF9 will demodulate. It joined on the first attempt in March and needed four attempts after the hawthorn came into leaf in May.

The lesson is not that LoRa is disappointing. It is that vegetation is a seasonal attenuator, and a network commissioned in bare spring will not behave the same way in July. Anyone planning a site-wide deployment should survey in leaf, not in bud, and should budget for a spreading factor two steps higher than the link calculator suggests.

Power, and the limits of a small panel

Average gateway consumption came out at about 1.4 W, which is a hard number to argue with until you look at a British September. The panel carried the load comfortably from March through August, including a fortnight of unbroken sunshine in June when the battery sat at absorption voltage for days. Then came three consecutive overcast days at the end of September, and the pack drifted to 11.9 V, at which point the gateway's own brown-out protection rebooted it twice in an evening.

The fix was unglamorous: a second 10 W panel angled more steeply, and a rule that the gateway should stop forwarding routine telemetry below 12.1 V rather than wait for the hardware to decide. Node-side power was never a problem. At one transmission every fifteen minutes, the AA lithium cells showed an estimated three-year life on paper; measured voltage droop after seven months was consistent with something in the region of two and a half years, derated for cold mornings.

Water, water, everywhere

Two of the eight enclosures let water in. Both were IP67-rated, and both failed at the same place: the cable gland, where the probe lead entered. A gland sized for 4–6 mm cable mated to a 3.5 mm lead will seal on the bench and weep in a week of drizzle driven sideways by wind. One node lost a fortnight of readings before the corrosion on the header pins became obvious.

Mitigations were cheap and effective. Silica desiccant sachets inside every enclosure, replaced monthly. A drip loop taped into the probe cable so water runs off rather than toward the gland. Conformal coating on the one board that had already suffered. The two recovered nodes then ran without incident through the wettest October week of the season, which is a small but genuine result.

Soil probes themselves held up better than expected. The capacitive pair that spent the season in a permanently damp bed showed a slow upward drift of about 4% volumetric moisture content by October, probably from soil compaction around the blade rather than electronics. The shed control node, sitting in dry air on a shelf, drifted by less than 1%.

The data, and the shape of its gaps

Overall packet delivery was 96.4% across the season, which sounds excellent until you disaggregate it. The six healthy nodes delivered 99.1%. The two marginal nodes delivered 78% and 84%, and their losses clustered precisely where you would expect: in heavy rain, in full leaf, and in the last hour before a battery-driven reboot.

There was also a subtler failure of interpretation. Moisture readings tracked rainfall with an obvious lag, but the lags differed by bed, and the dashboard's default fifteen-minute resolution made slow drainage look like noise. Replotting daily means turned the data into something a gardener could actually use. A civic-tech project that ships only a raw time series has shipped a database, not a tool.

Where the weak links were

Not the radio protocol, which did what the standard promises. The weak links were mechanical and meteorological: glands, gaskets, desiccant, panel angle, and the assumption that a link budget calculated in March still holds in July. The second-order weak link was human. Two nodes were unplugged during a working-party day and reinserted without anyone noting it, producing a fortnight of readings that were true, plausible and attached to the wrong bed.

For a hobbyist, the honest summary is that this is deployable in the UK on a modest budget, provided you treat enclosures as the hard part and the radio as the easy part. Label every cable, survey in leaf, oversize the panel, and design the dashboard for the person who will actually look at it on a wet Sunday in October.