[Insert Fig. 3 here]
In theory, window sizes of 80 s can be used, but in practice the quality of recorded temperature data deteriorates as temperatures level off and noise thus becomes increasingly prominent. For example, if we exclude temperature rises of less than 0.015 K (3 times the standard deviation of temperature noise due to sensor resolution), sap velocities > 80 cm h-1 could not be deciphered using a 100 s window. Optimal window sizes generally decrease as flow rates increase. To simplify, we suggest a fixed window of 40 s forV DRM calculations – the same window size as recommended for the HRM.
In minimizing SEs, the DRM algorithm identifies both the optimal size and timing of the averaging window. There is a distinct transition of optimal window size and timing in the range of velocities at which the intrinsic uncertainties of V 12 andV 23 are similar, which occurs at around ~ 20 cm h-1. At lower sap velocities,V 12 is less uncertain thanV 23 (hence V DRM =V 12, Eqn. 7 ), giving an optimal averaging window 40 s wide and centred between 60 and 70 s (Fig.4 ) – very similar to the averaging window commonly used for the HRM (40 s wide and centered at 80 s, see Chen et al. (2012)). At intermediate sap velocities, V 23 is less uncertain than V 12 (henceV DRM = V 23), and the optimal averaging window is centered at > 200 s – roughly when the temperature traces for Probe #2 and #3 intersect (see blue and red lines in Fig. 4 ). Note that the velocity at which the transition from V 12 to V 23occurs will depend on conditions and will differ with probe positions and heat pulse strength and length. For example, greater heat intensity (Q /t 0) will lead to a transition at greater flux.