SPECPR Users’ Manual Page 8-f17.1 F17: High To Low Resolution Convolution Alias: conv[olv] Routine f17 convolves a high resolution data set to a low resolution data set. It has 2 modes: 1) use Gaussian profiles whose widths and centers are specified by the user, or 2) use user-specified filter profiles. The routine outputs two data sets: (1) the resultant wavelength values (center values) and (2) the convolved spectrum. The output of the center values is optional. A typical use would be to use mode 1 (Gaussian profiles) for convolving one spectrometer to another, and mode 2 for convolving a spectrum to a specific filter response, such as a LANDSAT TM system. To run the function, go to the math routines and enter the file ID and number for the input spectrum and f17. Example: v23f17 or v23conv The routine starts with a query for the use of the default wavelength assignment, which is displayed at the top of the screen. To change the default assignment, enter the wavelength data set file ID (upper case) and the record number. If you don’t want to change the wavelength set, just type return. Next you will be asked whether you want to normalize the data. This is a convolution normalization. Normally you want to do this. If you normalize,the resulting spectrum is divided by a convolved spectrum whose input data values are all one. If you don’t normalize, the convolution output will vary as a function of the area under the curve of each response function. Unnormalized convolutions can be used for tracking voltages from a sensor. If you simply want to compare the spectral response, calibrated to, say, radiance or reflectance, and your input data are calibrated in those units, you want to normalize the convolution. Next you are asked if you want to use the Gaussian mode or the filter bandpass mode. Type: g Gaussian mode. <return> bandpass filter mode.

Gaussian Profile Convolution You are first asked for a data set describing the center values of the Gaussians for the output data set (these are the wavelengths you are convolving to). Enter the file ID and record number of the data set. Next you are prompted to enter the Gaussian bandwidths. These are the Gaussian Full Width at Half Maxima (FWHM) for each Gaussian center value entered in the previous step. Enter the file ID and the record number of this data set. The Gaussian bandwidth data set and the Gaussian center values must have the same number of channels. Next your are asked to input the bandwidths of the input data set. The spectral resolution of the input data set must be known to properly convolve to the output data set. Enter the file ID and record number of the input spectrum resolution. This data set must have the same number of channels as the input wavelength data set in use. The convolution routine computes the convolution and then asks if you want to output the center wavelengths. See below.

Bandpass Filter Convolution If you selected the bandpass filter convolution mode, the routine prompts you for the first filter profile and asks how many filters will be convolved. Each filter profile produces one output channel. The filter profile must be sampled at the wavelengths of the input data set (if it is not, you could reinterpolate them using a routine like the cubic spline, f12). For example, say you were convolving a laboratory spectrum of kaolinite to LANDSAT TM filters (7 bands). Say the filter response function sampled at the wavelengths of the laboratory spectrum was at w46 (the other 6 filters must be sequentially located in the same data file). You would then type:

w46 7 |

The input filter response functions are read in and the output spectrum is calculated. Next the user is asked if the output of center wavelengths is desired.

Center Wavelengths After the convolution is complete (either Gaussian or bandpass mode) you are queried on whether or not to output the center value wavelengths. If you wish to write the center values, type in the file id and record number of the file where you wish to write the center values, or type return to ignore writing the center values. Next enter the title for the center values when prompted. Because this is an output data set (like any other), use a lower case file id(not upper case).

Note that f17 cannot include errors in its calculations. However, you can do error propagation by simply convolving the error data set in the same manner as the regular data set.

The equation used to calculate the spline is

Rj = ∫xxl S Bj dx, (eqn 8.f17.1) |

where Rj is a data point in the resultant spectrum, xo is the first value of the input spectrum, xl is the last value of the input spectrum, S is the input spectrum, and Bj is the spectral bandpasses. The center values are calculated by the following:

Cj = ∫xxl x S Bj dx, (eqn 8.f17.2) |

where Cj is the center value for that particular spectral bandpass, xl, S, Bj and xo are the same as in equation 8.f17.2, and x is the wavelength. The integrals are approximated by using numerical integration by summing the rectangular segments. Each rectangle is computed by using the channel separation of the input spectrum:

n Rj = Σ Si Bj,i δxi, (eqn 8.f17.3) i=1

where n is the number of channels and δxi is the spectral bandpass of channel i

╭ x2 - x1, i = 1 │ δxi = ┥ xi+1 - xi-1, 1 < i < n (eqn 8.f17.4) │ ╰ xn-1 - xn, i = n

Similarly, the center values (wavelength) of each convolved channel are computed from the equation

n Σ xi Si Bj,i δxi (eqn 8.f17.3) i=1 Cj = ————————————————— Rj

During the calculations, divide operations are checked for a zero denominator. If this happens, an error message is printed for the user, and the denominator is reset to 1.0×10-36 instead of crashing the program. If no file has been input to the program at the start, an error message will be printed, and the routine exits. Note that if the number of bandpass files times five is greater than the number of channels, a warning message will be printed explaining that, with data arranged in such a manner, the resulting accuracy of the data is questionable. This is due to using the summations (equations 3 and 5) and the equation 4 to approximate the integrals. The message is only a warning--you must determine if the bandpasses are complex enough that accuracy may actually be in error due to a small sample. IMPORTANT CONSIDERATION: if the filter bandpass or the Gaussian profile is close to the input spectral resolution, you may want to interpolate the input spectra to finer channel spacing. Because the numerical integration rectangles are controlled by the channel spacing, interpolating to finer spacing will increase the accuracy of the numerical integration. You must interpolate both the input spectra and the resolution "spectrum" to the finer wavelength set. Remember, for example, if you had 0.001 μm spacing and 0.001 μm resolution, and you interpolated to 0.0001 μm spacing, the resolution is still only 0.001 μm, thus the spectrum is oversampled. This oversampling is only useful in terms of making the numerical integration more accurate. If your convolved spectra show a "staircase" effect, it is an indication of the limitation of the numerical integration and you should resample the input data to finer spacing.