NEP, D*, NETD

These three quantities are of importance in optical sensing systems and concern the measurement of radiation intensity.  NEP and D* may be applied to the whole spectrum of electromagnetic radiation; NETD is only relevant in the thermal infrared region.  NEP, D*, NETD are applied to both image sensors and single detectors.  When applied to an image sensors, one considers a single pixel from the array.

Definitions in words

NEP

noise equivalent power: the signal [V] of that optical power [W] impinging on a pixel that corresponds to the RMS value of the noise [VRMS].

D*

Specific detectivity; noise equivalent power, normalized to unity detector (pixel) area (to compare pixels of different sizes) and to unity bandwidth (to compare systems of different speeds).

NETD

noise equivalent temperature difference: the signal [V] of that scene temperature difference [K] impinging on a pixel that corresponds to the RMS value of the noise [VRMS]. This is of course only relevant for thermal IR detectors.

As you are an engineer, you want something more solid than a definition in words.  Herebelow follow the formulas.

Let
R = responsivity of detector channel (e.g. a pixel) in [V/W]
Z = optics conversion factor in [K/W]
A = pixel area [m2]
Vn = RMS noise voltage in [V]
BW = signal bandwidth, as seen by the final output, in [Hz]
Fa = acquisition frame rate in [Hz] (clearly distinguished from BW!)
SVn = spectral density of noise voltage
in [V2/Hz]
In general Vn is the integral over bandwidth of (SVn(f).df). For white noise: (Vn)2 = SVn * BW.

then:
NEP = Vn / R in [W]
NETD = NEP * Z in [K]

According to this definition, NEP and NETD are not normalized quantities; their meaning is "the accuracy of a single measurement of a pixel of such-and-such camera". Some authors consider NEP and NETD as normalized . To avoid confusion, we will postfix the normalized qualtities with a "*" (as with D*) [1]

NEP* = NEP / sqrt(Fa) in [W/sqrt(Hz)] [2]

NETD* = NETD / sqrt(Fa) in [K/sqrt(Hz)]

D* = sqrt(A)/NEP* in [m.sqrt(Hz)/W]

 

The NEP of a chopping image sensor

A chopping image sensor takes alternating illuminated and dark (reference) frames. The resulting "real" image is obtained from the difference (demodulation) of two consecutive frames. Chopping removes many sources of non-uniformity and low frequency (LF) noise as 1/f noise. As the resulting images (pixel signals) are uncorrelated in time, and have thus white noise, it is again possible to calculate NEP*, NETD* and D*.   Chopping and demodulation remind of correlated double sampling.

The chopping and differencing operations decrease the signal to noise ratio [for a white noise limited system] with a factor 2, compared to a situation where 100% of the time is used for useful observation

consider the following case:

Note: Fa is the well defined image acquisition frequency, in this case 40Hz. BWa is the signal bandwidth as supposed for a, which may be much higher, e.g. in case of a bolometer readout, it is the bandwidth of the pixel sampling. Clever readers will have seen different definitions for NEP* and D*, featuring a bandwidth and not a sampling frequency. Image sensors and pixels are sampled data systems. In cases that Fa and BW do strongly differ, a image sensor is indeed inferior to a continuous time system. Think it over.

Note: chopping decreases the S/N by a factor 2 if the signal is white noise limited. even if the unchopped signal contains 1/f or other LF noise, it will produce a signal that is free of 1/f noise, and thus can be averaged to improve S/N. Although the chopped signal is free of LF noise, NEPb as given above does retain its 1/f contribution! The 1/f is just aliased to higher frequencies. It would be completely wrong to assume that 1/f noise sources may be neglected when using chopping.

Note: capacitor (dis-)charge noise (also called reset noise or kTC-noise) may be cancelled by chopping, if and only if the circuit operation is such that the capacitor is not electrically (dis-)charged in between the chopper halve periods. This principle is similar to correlated double sampling.


From blackbody power to temperature difference

A 300K blackbody emits about P = 1 kW/m2. This amount follows a 4th law, thus P ~ T4. The differential power per kelvin is thus

 
dP      4  1kW/m2
--  =  ---.------  =  13 W/m2K
dT     300    K

Caveat: here we assumed the full optical bandwidth, and a perfect blackbody. A real object has a lower emission (typ 50%), does not follow the 4th power law precisely, and the spectrum is typically limited by the detector response, filter transmission or atmospheric transmission. In literature a value of 2.5 W/m2K for the 8-12 um band seems to be popular.


From scene radiation to radiation on pixel

The object emits power in all directions, i.e. over a halve sphere. At a distance r of the source, the power density is 2.PI.r2. The pixel in the camera sees the outside through the opening of the lens. It is of no importance whether the lens is in place or just an opening, as the average power through the opening is the same. With a lens diameter d and a focal distance F, one defines the "f-number" of the lens as f = F/d. "Good" (sensitive) lenses have low f. The highest qualities are in the range 0.8...1.5. The lens only transmits a part of the total radiation of the hemisphere. In the case of uniform, omnidirectional radiation, the ratio between the energy passgin through the lens opening and the total radiation density is 1/(4.f2). The attenuation of the radiation flux from object radiation flux to detector incoming power is thus:

power_pixel[W] = flux_scene[W/m2] . 1/(4.f^2) Area_detector [W/K] [W/m2K]

EXAMPLE suppose 300K blackbody scene, f=1 optics, 20x20 um detector. For the 300K scene we calculate 13 W/m2K differential power. thus power_pixel[W/K] = 13 [W/m2K] . (1/4) . 400E-12[m2] = 1 [nW/K]. Btw, this example shows that any reasonably good NEP should be better than 1E-10W.



[1] this "*"-notation is not generally accepted, except for D*.

[2] division by sqrt(Fa) is ONLY ALLOWED if the noise is essentially white, thus uncorrelated over time. If the noise is dominated by 1/f noise (the pixel suffers from low frequency noise), NEP* and NETD* can simply not be defined.