SWIR imaging · 12 min
Every silicon camera goes blind at 1100 nm — not by design, but by bandgap. On the other side of that line sit 1064 nm lasers, see-through silicon, moisture bands, and the NIR-II tissue window. This is how to choose the InGaAs sensor that gets you there.
The governing physics
Whether you need a SWIR camera at all, and which one, follows from three pieces of semiconductor physics. Two of them are the reason to buy; the third is the bill.
λ_cutoff = hc / E_g → Si: 1100 nm · InGaAs: ~1700 nm
A photon with less energy than the bandgap cannot free an electron, so it passes through the pixel unrecorded. Silicon's 1.12 eV gap ends its vision near 1100 nm; InGaAs at 0.75 eV carries on to about 1700 nm. No exposure time argues with this.
Si transparent >1150 nm · H₂O absorbs at 1450 nm · scattering ∝ λ⁻ᵇ
The same bandgap that blinds silicon cameras makes silicon wafers transparent, so SWIR looks through them. Water develops a strong absorption band, so moisture turns dark. And scattering keeps falling with wavelength, opening the NIR-II window through tissue, haze, and some plastics.
I_dark ∝ exp(−E_g / 2kT) → ~2× every 8 °C
A gap small enough for 1550 nm photons is small enough for thermal energy too, so InGaAs pixels fill with hundreds of dark electrons per second where cooled silicon collects a fraction of one. That is why every serious SWIR camera ships with a TEC — and why exposure time, not preference, sets how deep the cooling must go.
400–1700 nm · 3.45–5 µm px · TEC ← vs → 900–1700 nm · 15 µm px · deep TEC
Sony's SenSWIR sensors thin the InGaAs layer so it responds from 400 nm upward: one camera sees the visible and infrared worlds simultaneously, in registration, with pixels small enough for real resolution. That is the SWIR5000 / 3000 / 1300 / 330 line.
The deep-cooled 900–1700 nm class spends the budget the other way: 15 µm pixels with nine times the collecting area, on large formats, refrigerated 40 °C below ambient. It cannot see the visible band and does not care — it exists for measurements where every photon is precious. That is the SWIR1302 / 331 line, and the fork between the two families is the first real decision.
Interactive
The simulation puts real QE curves and datasheet dark-current figures behind three sliders. Park the wavelength at 1064 nm and watch silicon collapse; stretch the exposure to seconds and watch deep cooling earn its price.
Pick a working wavelength and a light level, and see which sensor delivers a usable signal-to-noise ratio — and which noise source is actually in charge of it.
wavelength (nm)
QE
79%
SenSWIR at 1310 nm
Signal
791 e⁻
photoelectrons / px
Dark charge
38 e⁻
383 e⁻/s × 100 ms
SNR
19
per pixel, single frame
1310 nm — The NIR-II window — tissue scattering and autofluorescence collapse, silicon is blind.
10^4.0 photons/px/s — moderate: lab imaging, active illumination
100 ms per frame
What limits you now
Read-noise limited: frames are short enough that dark current barely registers. More exposure per frame helps; so would binning or a higher-gain readout mode.
QE curves are digitised from typical published response data; dark current and read noise come from the SWIR series datasheets at their shipped cooling points, with pixel areas of 6.5, 5, and 15 µm respectively. Real scenes add stray light, lens transmission, and fixed-pattern effects this model ignores. Use it to understand which regime you are in, then confirm the numbers for your configuration with us.
At a glance
| Sensor class | Spectral range | Pixels · format | Cooling | Typically chosen for |
|---|---|---|---|---|
| Cooled silicon sCMOS | 400 – 1000 nm (usable) | 6.5 µm · large formats | Deep TEC · <1 e⁻/s dark | Everything below 1000 nm — cheaper, lower noise, higher resolution. The reference to beat. |
| SenSWIR InGaAs, 3.45 µm — SWIR5000 / SWIR3000 | 400 – 1700 nm | 3.45 µm · up to 5 MP | TEC 25 °C below ambient | Resolution-first SWIR: wafer inspection, material sorting, one camera covering visible and infrared at once. |
| SenSWIR InGaAs, 5 µm — SWIR1300 / SWIR330 | 400 – 1700 nm | 5 µm · 1.3 / 0.33 MP | TEC 25 °C below ambient | Sensitivity and speed: laser profiling, 200–753 fps process monitoring, general lab SWIR. |
| Deep-cooled InGaAs, 15 µm — SWIR1302 / SWIR331 | 900 – 1700 nm | 15 µm · large 1.5″ / 3/4″ formats | Deep TEC 40 °C below ambient | Photon-starved work: NIR-II in vivo fluorescence, long exposures, nightglow, 1000 fps CameraLink variants. |
Most SWIR purchases are driven by one specific measurement. If yours is on this list, the first-choice column is the camera to price first.
| Application | First choice | Why |
|---|---|---|
| 1064 nm laser beam profiling & alignment | SWIR1300 | Silicon keeps a few percent QE at 1064 nm — enough to glimpse a bright beam, useless for quantitative profiles. InGaAs sits near 75% there. |
| Through-silicon wafer & IC inspection | SWIR5000 · BSM microscope | Silicon turns transparent past its own bandgap, so >1150 nm light images die attach, voids, and cracks through the wafer. Resolution matters, so take the 5 MP sensor. |
| NIR-II / SWIR in vivo fluorescence | SWIR1302 | Tissue scattering and autofluorescence collapse in the 1000–1400 nm window, but the photons are few: 15 µm pixels and deep cooling are the requirement, not a luxury. |
| Moisture detection & material sorting | SWIR3000 | Water absorbs strongly at 1450 nm, so moisture reads dark. Visually identical plastics separate cleanly by SWIR reflectance — through packaging. |
| Telecom & silicon photonics (1310 / 1550 nm) | SWIR1300 | Waveguide and fibre work lives exactly where InGaAs QE peaks; 200 fps helps active alignment loops converge. |
| High-speed inline inspection | SWIR330 · SWIR331 | 400–753 fps over USB3, and up to 1000 fps via CameraLink on the SWIR331 — SWIR contrast at production-line rates. |
| Astronomy J and H band | SWIR1300 | The atmosphere is transparent in J (1100–1400 nm) and H (1500–1800 nm); large pixels and TEC cooling suit long guided exposures. |
| UAV / embedded night observation | SWIR330 UMV | Atmospheric nightglow illuminates scenes in SWIR even on moonless nights, and the 33 × 33 × 38 mm, 70 g UMV body flies where cooled cameras cannot. |
A defensible order
SWIR cameras are priced like instruments, not webcams, so the expensive mistake is buying capability the measurement never uses — or missing the one specification it cannot live without.
Below 1000 nm, a cooled silicon sCMOS is cheaper, lower-noise, and higher-resolution — SWIR buys you nothing. Between 1000 and 1100 nm, deep-depletion silicon can sometimes stretch. Beyond 1100 nm the bandgap has decided for you, and the only question left is which InGaAs.
SenSWIR sensors see visible and SWIR at once — one camera, one optical path, registered images across the whole range. The 900–1700 nm deep-cooled class gives up the visible half to spend everything on large pixels and low dark current. Whether you need the visible band answers this in one question.
Milliseconds: dark current is irrelevant — take an uncooled UMV or standard TEC and enjoy the frame rate. Hundreds of milliseconds to seconds: dark charge grows linearly while signal shot noise grows as its square root, and deep cooling becomes the specification that decides the measurement.
Only now pick megapixels against per-pixel sensitivity (3.45 µm vs 5 µm vs 15 µm), the frame rate the process needs, and the interface the system can carry — USB3, GigE, 10GigE, CameraLink, CXP, or the 70 g UMV module for airborne and embedded builds.
The cameras
Four SenSWIR models covering 400–1700 nm and two deep-cooled 900–1700 nm models, each in USB3, GigE/10GigE, CameraLink, CXP, or compact UMV variants — the interface changes, the sensor decision does not.
The resolution flagship. Five megapixels of full-band SWIR for wafer inspection and material sorting where feature size, not photon count, is the constraint.
The 3 MP middle ground in a compact 1/1.8″ format — process monitoring and QC lines that want resolution and rate from the same camera.
The lab default. Larger 5 µm pixels collect more light each, 58.7 dB dynamic range handles laser work, and 200 fps at full resolution covers almost every scientific use. Start here.
The speed-and-size pick: 400 fps over USB3, and a 70 g uncooled UMV variant that turns a drone or an embedded gantry into a SWIR instrument.
The photon-starved specialist. Nine times the pixel area of the SenSWIR line and deep cooling for the long exposures of NIR-II in vivo fluorescence and low-radiance imaging.
Large pixels at extreme rates — up to 1000 fps via CameraLink with deep cooling intact, for fast NIR-II tracking and semiconductor process inspection.
The BSM modular microscope systems pair these cameras with infinity-corrected M Plan Apo NIR objectives (5× to 50×, NA to 0.65) and 1200–1550 nm LED illumination — a complete through-silicon inspection and NIR-II microscopy bench where objectives and cameras swap without realignment.
Around the camera
A SWIR sensor only reports what the optics deliver and the illumination provides. These are the tools and companion guides for the rest of the chain.
Filter the full camera range — silicon and InGaAs — by sensor, cooling, speed, and interface against your application.
The silicon side of this decision: QE, read noise, and pixel size below the 1100 nm line.
Where the silicon cutoff first bites most labs: 785 and 1064 nm Raman push the Stokes bands into exactly this territory.
Pairing an InGaAs array with a spectrograph turns this guide’s wavelength range into NIR spectroscopy.
1064, 1310, 1444, and 1550 nm sources — active illumination for the wavelengths these cameras see.
Bandpass and longpass filters for SWIR bands — including blocking the visible response of 400–1700 nm sensors when only the infrared matters.
The quiet failure mode of a new SWIR setup is the optics in front of it: visible-band AR coatings ghost and lose transmission past 1100 nm, and lenses corrected for visible light focus SWIR at a different plane. If the image looks soft or dim on day one, suspect the lens before the camera — and specify SWIR-corrected optics for quantitative work.
Common questions
A photon can only create a photoelectron if it carries more energy than the sensor material’s bandgap. Silicon’s bandgap is 1.12 eV, which corresponds to a cutoff near 1100 nm — beyond that wavelength the photon passes through the pixel without being absorbed, no matter how long you expose. InGaAs has a bandgap of about 0.75 eV, moving the cutoff to roughly 1700 nm. The cutoff is physics, not engineering: no firmware, cooling, or exposure setting moves it.
Marginally. A back-illuminated silicon sensor keeps a few percent quantum efficiency at 1064 nm — often enough to locate a bright laser spot for alignment, and some deep-depletion CCDs stretch this usefully further. But for beam profiling, power mapping, or anything quantitative, working at 3% QE against 75% on InGaAs means fighting a twenty-five-fold signal handicap plus etaloning artefacts. If 1064 nm is the experiment rather than a nuisance, it is a SWIR camera’s job.
The 1000–1400 nm band, also called SWIR or NIR-II, is where biological tissue becomes unusually cooperative: scattering falls steeply with wavelength, autofluorescence nearly vanishes, and absorption stays low between the water bands. Fluorophores emitting there can be imaged millimetres deep in vivo with resolution that visible-band imaging cannot approach. The price is photon scarcity — emission is weak, which is why NIR-II cameras lead with large pixels and deep cooling rather than megapixels.
InGaAs cannot be made with standard CMOS fabrication. The photodiode layer is grown on indium phosphide and must be hybridised to a silicon readout circuit, pixel by pixel, with yields far below silicon norms — and the market volume is a fraction of the consumer imaging that subsidises silicon sensors. Sony’s SenSWIR line has pushed prices down meaningfully by shrinking pixels to 3.45–5 µm and industrialising the process, which is precisely why the 400–1700 nm series exists at its price point.
Decide by exposure time. The small bandgap that lets InGaAs see to 1700 nm also generates dark current — hundreds of electrons per pixel per second near room temperature, roughly doubling every 8 °C. At machine-vision exposures of a few milliseconds that charge is negligible and an uncooled UMV module is fine. At the 100 ms–10 s exposures of fluorescence or astronomy, dark current becomes the dominant noise source, and the 40 °C-below-ambient deep TEC of the SWIR1302/331 class is what keeps it below the signal. The lab above lets you find the crossover for your numbers.
Optically, most glass transmits well to 1700 nm and beyond — the C-mount thread fits and an image appears. Two things degrade: anti-reflection coatings designed for visible light lose effectiveness, costing transmission and adding ghosting; and chromatic correction no longer applies, shifting focus between visible and SWIR and blurring broadband images. For the 400–1700 nm sensors used across their whole range, or for quantitative work, SWIR-corrected lenses are worth specifying alongside the camera.
Tell us the wavelength, the light level, and the exposure time you expect. Those three numbers decide the sensor — and sometimes the honest answer is that a cooled silicon camera does it for a third of the price.