LoRaWAN (Long Range Wide Area Network) provides long-range communication with low energy consumption for precision-irrigation sensing, whereas periodic reporting occupies radio airtime even when measurements can be processed locally. This paper studies conformal risk-controlled uplink transmission using an INT8 (8-bit integer) TinyML (Tiny Machine Learning) classifier. A sensor node predicts whether an irrigation event will occur within the next hour from a two-hour multivariate sensing window. The local prediction is retained when a calibrated confidence score meets or exceeds a threshold; otherwise, a fixed sensor payload is transmitted to a gateway classifier. Conformal risk control (CRC) selects the threshold to control the expected value of the monotone retained-decision error loss under exchangeability. Evaluation uses the 2024 automated-irrigation season of a public LoRaWAN tomato testbed and 20 grouped data splits. At target levels ϵ = 0.01 and 0.02, the Tiny multilayer perceptron (MLP) activates the LoRaWAN uplink for 10.88% and 5.45% of windows, with mean retained-decision error risks of 1.11% and 2.16%, respectively. The INT8 model occupies 51.55 kB Flash and 1.57 kB peak RAM on an ESP32-S3, with 0.170 ms latency and 0.0111 mJ per inference. Adaptive prediction sets (APS), regularized adaptive prediction sets (RAPS), and posterior-probability thresholding yield comparable low-risk operating points. Cross-season evaluation shows a pronounced loss of positive-event recall under distribution shift, delimiting the applicability of the static 2024 calibration.

TinyML,LoRaWAN,precision irrigation,conformal risk control,conformal prediction,edge inference,uplink transmission