SAMPLING Sampling Requirements: 1) Instrument of Measurement 2) Scales to Measure
Should produce reliable and useful data 1) Instrument of Measurement Should produce reliable and useful data Accuracy vs. Precision true repeatable Precise but not accurate! (repeatable)
2) Scales to Measure Measurements should be collected often enough in space and time to resolve the phenomena of interest t, x u
Sampling Interval Choice of sampling increment t or x is important. Sample often enough to capture the highest frequency of variability of interest, but not oversample For any t the highest frequency we can hope to resolve is 1/(2 t) Nyquist Frequency (fN) fN = 1/(2 t) ; if t = 0.5 hrs fN = 1 cycle per hour (cph)
This means that it takes at least 2 sampling intervals (or 3 data points) to resolve a sinusoidal-type of oscillation with period 1/ fN if 1/ fN = T = 12 hrs, then 1/(1/2 t ) = 12 hrs and t = 6 hrs, i.e., t = T/2 t 12 hrs t
In practice f = 1/(3 t) due to noise and measurement error. If there is a lot of variability at frequencies greater than f we cannot resolve such variability aliasing For example, ■ sampling every month regardless of the tidal cycle ■ sampling for tidal currents every 13 hours Then, we should measure frequently!
■ We should sample long and often! Sampling duration We should sample to resolve the fundamental frequency (fF) fF = 1/(N t) = 1/T ■ We should sample long and often!
Δf × LOR ≤ 1 → Rayleigh Criterion LOR = length of record Sampling duration To resolve two frequencies separated by ( Δf ) Δf × LOR ≤ 1 → Rayleigh Criterion LOR = length of record e.g. Δf = 2π/12 h - 2π/12.42 h = 0.0177062 h-1 T = 2π/0.0177062 h-1 = 354.858 h = 14.786 days
Continuous sampling vs. Burst sampling Continuous sampling mode: at equally spaced intervals Burst sampling mode: burst embedded within each regularly spaced time interval 0 2 4 6 8 10 12 hrs
Regularly vs. Irregularly sampled data Regular if unknown distributions Irregular if looking for specific features
Independent Realizations If correlated not independent do not contribute to statistical significance of measurements
(prepared by Lonnie Thompson – Ohio State University)
(prepared by Lonnie Thompson – Ohio State University)