Generating hyper-spectral images with discrete wavelength samples

[DnBaral]

I am simulating radiance images taken of space debris with a telescope. For further analysis this images need multiple spectral channels representing discrete wavelengths or small bands. I already had some smaller success with a classic path integrator and the specfilm. While I provide spectral data for the BSDFs and emitters, the sensor response functions are each defined for a small spectral band. This produces an image with the shape (H, W, C), where C is the number of discrete wavelength bands.

for i in range(num_bands):
    wl0 = wl_subset[i, 0]
    wl1 = wl_subset[i, 1]
    samples = np.linspace(wl0, wl1, samples_per_band, endpoint=False) + (
        wl1 - wl0
    ) / (2 * samples_per_band)
    vals = np.ones_like(samples)  # TODO: replace with Sensor Response Function
    spd_path = os.path.join(data_dir, f"{prefix}_{i+1:03d}.spd")
    self.write_spd(spd_path, samples, vals)
    film[f"band_{i+1:03d}"] = {"type": "spectrum", "filename": spd_path}

For small numbers of wavelength bands this works fine. But if I increase the number of bands (i. e. 120) the rendering becomes very slow.

I found no standard plugins that seems to produce such "real" hyper-spectral radiance cubes. To accelerate the rendering I tried to create my own custom integrator and film, without success. Ideally, the integrator should evaluate the surface interaction for each discrete wavelength. Maybe it is possible to create a nested parallelism to evaluate each wavelength. This would result, that each ray-wavelength pair is calculated in parallel accelerating the calculation.

Are there some examples of a hyper-spectral renderer in Mitsuba3 that I can use as a starting point? Is such a concept even feasible in Mitsuba3 or do i have to change plans completely? I appreciate any help or tips. Maybe someone had a similarly project that succeeded?

PS:
For my understanding the native Mitsuba3 spectral rendering uses a Montecarlo integral to sample over multiple wavelengths/spectra. This integral somewhat "destroys" the discrete sampling approach. It accelerate the creation of spectral images with the loss of discrete wavelength slices.