Farm-scale minimum temperature mapping for strategic and tactical frost management

Abstract

Frost causes millions of dollars of damage to broad acre cereal crops and is an important climate risk that presents a management challenge to grain growers in Australia. Management options to mitigate impact of frost include pre-season cropping decisions such as timing of sowing, choice of crop type and cultivar, and post-frost event management such as cutting for hay, or grazing frost-damaged crops. This study, conducted in close collaboration with growers and farm consultants, contributes to both pre-season and post-frost management through improved spatio-temporal understanding of frost risk. This study aims to identify combinations of meteorological and terrain variables suitable for generating high spatial resolution (approx. 30 m) maps of minimum temperature (Tmin). This mapping used Multivariate Adaptive Regression Splines (MARS) models to combine terrain indices with temperature data. Modelling and mapping were undertaken for two sites in the SE Australian cereal-cropping zone, Mintaro, South Australia and Hopetoun, Victoria. To facilitate the development of the mapping methodology, iButton® temperature loggers were deployed at each site during June-September 2016. A clustering-based sampling scheme, for 96 loggers at each site, sought to capture terrain-driven temperature variability at landscape scale (4 x 4 km at the Mintaro site; 5 x 3 km at the Hopetoun site). The logger dataset from the Mintaro site included data from 71 loggers over 25 cold nights (<2°C) and from the Hopetoun site 87 loggers over 20 cold nights, with temperatures recorded at 30 minute intervals. MARS models were developed to predict the logger Tmin values. We compared models using three separate sources of temperature data, a local weather station at each site, an official weather station and remotely sensed Moderate Resolution Imaging Spectroradiometer (MODIS) night-time land surface temperatures. We also developed models using temperature loggers themselves as predictors for the entire site, both individually, and with all 3-logger combinations for each site. Comparing models which used local, distant, and remotely sensed temperature predictors showed that locally measured temperatures generate the best models. Then, using the individual loggers as predictors showed that wind speed and humidity were important predictors in many cases. Additionally, by modelling every three-logger combination for each site, we demonstrated that some logger locations were not used in any models, and further, some models using a single logger performed well. There are several practical findings from this work. Firstly, local temperature measurements show the best potential for generating high-resolution maps of Tmin at farm scale. Our results also suggest that remotely sensed temperature is not acquired reliably enough to make it suitable for this purpose. Models based on temperature data from one to three logger locations fitted well, but showed that location of loggers may be important. The improvement to models by including wind-speed and humidity measurements shows that these may be important parameters to measure. With further validation and refinement of the methods, we believe this method has the potential to be applied across the broad acre cropping areas of Australia.