Yes—at a basic level, you illuminate the sensor with known narrow-band light at many wavelengths, record each channel’s response, and plot response versus wavelength.

The y-axis can be shown in a few common ways:

• normalized response (unitless, often peak set to 1)
• quantum efficiency (percent of photons converted to signal)
• electrical/digital response per incident irradiance or radiant power

That’s why many published curves appear unitless: they are often normalized to compare shape rather than absolute sensitivity.

In practice, you would sample the pixel values from calibrated monochromatic captures for each channel, subtract dark/black level, keep exposure and illumination known, and then either:

• normalize the resulting values to get relative spectral sensitivity curves, or
• divide by the known incident light level to get an absolute responsivity curve.

If your goal is color/vision modeling, the normalized per-channel spectral sensitivity is often the key quantity, provided the measurement conditions are controlled and the optical path (filters, CFA, sensor stack) is included consistently. Absolute units matter more if you need true radiometric response, not just spectral shape.

So the short answer is: yes, the curves can come from measured pixel values, but whether they are unitless or physical depends on whether you normalize them or calibrate them against known input energy.