Machine learning reveals that leaf temperature extremes drive shifts in plant photosystem heat thresholds across marked microclimatic variation

Understanding relationships between plant heat tolerance thresholds and the environment currently is hampered by significant variation around the means, which masks potentially important information. Rather than relating to broad-scale climate measures, local adaptation of heat thresholds might occur at finer microclimatic scales, which are particularly variable in thermally extreme, heterogeneous environments, found in many alpine systems. Further, air temperatures frequently over or underestimate leaf temperatures, which are known to influence heat thresholds. Yet, a clear relationship between microclimatic conditions and heat tolerance thresholds has yet to be established. We aimed to determine the influence of prior leaf heat load on leaf photosystem heat tolerance thresholds (T_crit) for two co-occurring alpine plant species in Kosciusko National Park, Australia: Grevillea australis and Dracophyllum continentis. Measurements were taken on five consecutive days across eight paired sites contrasting in aspect (NW, SE) at Schlink Pass (ridge line) and Mt Stilwell (cold air drainage valley). We found that T_crit and its relationship with leaf temperature parameters, did not differ between species, locations or aspects. Traditional statistical models found that T_leaf parameters explained some variation in T_crit; however, when pooling across sites and species, machine learning identified that 85% of the variation in T_crit was explained by not only maximum, but also minimum leaf temperatures in the four days prior to measurement. This finding suggests that exposure to cold extremes could be conferring cross-tolerance, promoting heat tolerance acclimation. Microclimatic variation is complicated, potentially obscuring patterns that maybe present. To uncover these complex relationships between environmental conditions and plant acclimatory responses, we recommend integrating machine learning techniques with traditional statistical methods.