Intelligent wireless sensing network for precision frost management in sweet cherry (Prunus Avium)

Srikanth Gorthi

Tree fruit growers commonly rely on either nearby open field weather station or in-orchardair temperature (Ta) measurements to activate frost-mitigation systems including wind machines and over- or under-tree sprinklers. However, Ta often diverges from true bud tissue temperature (Tb), leading to premature or delayed frost mitigation. This dissertation aimed to integrate sensing, modeling, and forecasting framework to improve spring frost management in sweet cherry orchards. Following objectives were undertaken to realize a LoRaWAN- enabled wireless sensing network (WSN) to capture high-resolution microclimate variability within the blocks, to monitor Tb using direct, indirect sensing methods, and develop reliable machine learning models to predict Tb as well as nowcast localized input attributes for effective frost risk assessment and mitigation. Objective 1 quantified Tb–Ta offsets and developed a physics-informed machine learning framework using WSNs data derived feature sets from three topographically ( 5 to 15% grade) different orchard blocks. Mean differences between actual Tb and direct (thermocouple) and indirect sensing (radiative frost sensor (Trf )) were respectively –0.75 ◦C and –1.75 ◦C. Sup- port vector regression, among six machine learning models, was found the best model to predict Tb with the highest accuracy (RMSE = 0.8 ◦C; R2 = 0.98). This model used feature set that incorporated radiative-cooling metrics, sky-temperature estimates, and nighttime- weighting functions along with Ta. Predicted bud temperature (cTb) closely matched mea- sured Tb for chilling accumulation (67.7 vs. 68.8 h) and growing-degree hours (4426.9 vs. 4397.8), demonstrating its suitability for real-time frost-risk assessment. Objective 2 advanced WSN design and optimization to support site-specific frost monitoring. Multi-season deployments (2023–2025) across sweet cherry, grape, and apple orchards showed that the R2 radio (STM32WLE5JC) module achieved low packet-loss (1.8–10.2%) and approximately 115× lower sleep current than the R1 (RFM95W) configuration. Unmanned aerial vehicle-based thermal imagery was used to strategically deploy WSN sensing nodes and pertinent data was used to quantify spatiotemporal variability in Ta and Tb. Using QR-POD and compressed sensing, the number of required sensing nodes were optimized by 66% for Ta and 50% for Tb with reconstruction RMSE of 0.27 ◦C (Ta) and 0.45 ◦C (Tb). Results confirmed that single-point measurements are inadequate and that optimized WSNs substantially improve frost-risk assessment. Objective 3 developed a multi-task deep learning nowcasting framework leveraging 2.5 km Unresolved Mesoscale Analysis reanalysis, topographical embeddings, and high-frequency AgWeatherNet (15-min) and in-orchard (5-min) observations. The model achieved RMSE of 2.48 ◦C (air temperature), 14.40% (relative humidity), and 1.23 m s−1 (wind speed) for weather station-level forecasts (hourly, 10 days), and 1.71 ◦C, 7.62%, and 0.49 m s−1 at orchard sites (15-min, 48 hours). This architecture effectively captured mesoscale–microscale dynamics and shows strong potential for incorporation into adaptive decision-support systems for frost management. Overall, these contributions establish a cohesive framework that integrates sensing, sensor-fusion modeling, network optimization, and localized weather nowcasting to enhance frost-risk assessment and improve the resource use planning and precision mitigation strategies in sweet cherries and can be translated to other perennial specialty crops.

Intelligent wireless sensing network for precision frost management in sweet cherry (Prunus Avium)

Files and links (1)

Abstract

Metrics

Details