Tanager band spectral wavelengths?

Ok, I just found that the metadata for the individual STAC items does include information about the band wavelengths. Can I suggest, though, that it would be useful to have this information located elsewhere also?
 

BTW, here’s the list of Tanager center wavelengths I extracted from the metadata file above, and which I’ll incorporate into some reusable code:

var wl_tanager =
[376.40, 381.40, 386.40, 391.40, 396.30,
401.30, 406.30, 411.20, 416.20, 421.20, 426.20, 431.10, 436.10, 441.10, 446.10, 451.10, 456.00, 461.00, 466.00, 471.00, 476.00, 481.00, 486.00, 490.90, 495.90,
500.90, 505.90, 510.90, 515.90, 520.90, 525.90, 530.90, 535.90, 540.90, 545.80, 550.80, 555.80, 560.80, 565.80, 570.80, 575.80, 580.80, 585.80, 590.80, 595.80,
600.80, 605.80, 610.80, 615.80, 620.80, 625.80, 630.80, 635.80, 640.80, 645.90, 650.90, 655.90, 660.90, 665.90, 670.90, 675.90, 680.90, 685.90, 690.90, 695.90, 
700.90, 705.90, 710.90, 715.90, 721.00, 726.00, 731.00, 736.00, 741.00, 746.00, 751.00, 756.00, 761.00, 766.00, 771.10, 776.10, 781.10, 786.10, 791.10, 796.10, 
801.10, 806.10, 811.10, 816.20, 821.20, 826.20, 831.20, 836.20, 841.20, 846.20, 851.20, 856.30, 861.30, 866.30, 871.30, 876.30, 881.30, 886.30, 891.30, 896.40, 
901.40, 906.40, 911.40, 916.40, 921.40, 926.40, 931.50, 936.50, 941.50, 946.50, 951.50, 956.50, 961.50, 966.60, 971.60, 976.60, 981.60, 986.60, 991.60, 996.60, 
1001.60, 1006.70, 1011.70, 1016.70, 1021.70, 1026.70, 1031.70, 1036.70, 1041.80, 1046.80, 1051.80, 1056.80, 1061.80, 1066.80, 1071.80, 1076.80, 1081.90, 1086.90, 1091.90, 1096.90, 
1101.90, 1106.90, 1111.90, 1116.90, 1122.00, 1127.00, 1132.00, 1137.00, 1142.00, 1147.00, 1152.00, 1157.00, 1162.10, 1167.10, 1172.10, 1177.10, 1182.10, 1187.10, 1192.10, 1197.10, 
1202.10, 1207.20, 1212.20, 1217.20, 1222.20, 1227.20, 1232.20, 1237.20, 1242.20, 1247.20, 1252.30, 1257.30, 1262.30, 1267.30, 1272.30, 1277.30, 1282.30, 1287.30, 1292.30, 1297.30, 
1302.30, 1307.40, 1312.40, 1317.40, 1322.40, 1327.40, 1332.40, 1337.40, 1342.40, 1347.40, 1352.40, 1357.40, 1362.40, 1367.50, 1372.50, 1377.50, 1382.50, 1387.50, 1392.50, 1397.50, 
1402.50, 1407.50, 1412.50, 1417.50, 1422.50, 1427.50, 1432.50, 1437.60, 1442.60, 1447.60, 1452.60, 1457.60, 1462.60, 1467.60, 1472.60, 1477.60, 1482.60, 1487.60, 1492.60, 1497.60,
1502.60, 1507.60, 1512.60, 1517.60, 1522.60, 1527.60, 1532.60, 1537.60, 1542.60, 1547.60, 1552.60, 1557.70, 1562.70, 1567.70, 1572.70, 1577.70, 1582.70, 1587.70, 1592.70, 1597.70,
1602.70, 1607.70, 1612.70, 1617.70, 1622.70, 1627.70, 1632.70, 1637.70, 1642.70, 1647.70, 1652.70, 1657.70, 1662.70, 1667.70, 1672.70, 1677.70, 1682.70, 1687.70, 1692.70, 1697.70,
1702.70, 1707.60, 1712.60, 1717.60, 1722.60, 1727.60, 1732.60, 1737.60, 1742.60, 1747.60, 1752.60, 1757.60, 1762.60, 1767.60, 1772.60, 1777.60, 1782.60, 1787.60, 1792.60, 1797.60,
1802.60, 1807.60, 1812.50, 1817.50, 1822.50, 1827.50, 1832.50, 1837.50, 1842.50, 1847.50, 1852.50, 1857.50, 1862.50, 1867.50, 1872.40, 1877.40, 1882.40, 1887.40, 1892.40, 1897.40,
1902.40, 1907.40, 1912.40, 1917.40, 1922.30, 1927.30, 1932.30, 1937.30, 1942.30, 1947.30, 1952.30, 1957.20, 1962.20, 1967.20, 1972.20, 1977.20, 1982.20, 1987.20, 1992.10, 1997.10,
2002.10, 2007.10, 2012.10, 2017.10, 2022.00, 2027.00, 2032.00, 2037.00, 2042.00, 2047.00, 2051.90, 2056.90, 2061.90, 2066.90, 2071.90, 2076.80, 2081.80, 2086.80, 2091.80, 2096.70,
2101.70, 2106.70, 2111.70, 2116.70, 2121.60, 2126.60, 2131.60, 2136.60, 2141.50, 2146.50, 2151.50, 2156.50, 2161.40, 2166.40, 2171.40, 2176.40, 2181.30, 2186.30, 2191.30, 2196.20,
2201.20, 2206.20, 2211.20, 2216.10, 2221.10, 2226.10, 2231.00, 2236.00, 2241.00, 2246.00, 2250.90, 2255.90, 2260.90, 2265.80, 2270.80, 2275.80, 2280.70, 2285.70, 2290.70, 2295.60,
2300.60, 2305.60, 2310.50, 2315.50, 2320.50, 2325.40, 2330.40, 2335.30, 2340.30, 2345.30, 2350.20, 2355.20, 2360.20, 2365.10, 2370.10, 2375.00, 2380.00, 2385.00, 2389.90, 2394.90, 2399.90,
2404.80, 2409.80, 2414.70, 2419.70, 2424.60, 2429.60, 2434.60, 2439.50, 2444.50, 2449.40, 2454.40, 2459.40, 2464.30, 2469.30, 2474.20, 2479.20, 2484.10, 2489.10, 2494.00, 2499.00];
 

​@Emil Cherrington you are a machine! Thanks for sharing this back and for the suggestion. I will pin the Planet docs (which has a really good AI search feature) and the Tanager Product Spec to this group!

Adding here as well: 

• User Documentation
• Tanager Docs

I’ll add to the FAQ as well!

BTW, I also put the list of bands (spectral bands + the additional band masks, etc.) in a Google Sheet and made it public, if it’s helpful to folks. I also marked which thin bands are roughly equivalent to the broad bands of the Operational Land Imagers of Landsat-8 and Landsat-9. There are two spreadsheet tabs. See: https://bit.ly/tanager_bands

And from the second tab of the spreadsheet listed above:
 

Please keep in mind that the wavelengths are different for every scene. Imaging spectrometers don’t have “bands” in the sense that multispectral satellites do. Instead of filters, they regularly sample the spectrum at evenly spaced points, but the exact location of those points shifts due to thermal stresses and other factors. 

That’s the reason that the wavelengths are in the metadata for each file rather than in a central location. They’re not constant and are continually changing.  That’s also why it’s important not to treat them like bands and instead select the band index range based on the wavelengths you want from each file. Band 11 in one file could be closer to band 12 in another. 

The examples you’ve given are useful in general, but are only correct in detail for a single scene. The shifts are currently small, but it’s dangerous to assume they’re constant.  If you’re reading the data in via gdal/rasterio/etc, they’ll be available as a “wavelengths” metadata item on the reflectance or radiance cube subdataset.  If you’re reading the data in using HDF5 (e.g. h5py), they’ll be an attribute on the radiance cube subdataset.

Point taken, and that’s good to know, ​@joe.kington. Are there any plans to do any harmonization across the collected data to make it such that spectra extracted from time 1 can be compared to spectra from time 2?

Sorry for the delay!

No, there are no plans that I’m aware of for that.  “Harmonization” for this is usually more harmful than helpful.  If you’d like to interpolate to new spectral sample locations, it’s easy to do, but doing so will distort key features that hyperspectral remote sensing focuses on like the exact location of spectral adsorption features.  (e.g. an adsorption peak being shifted by even 1 nanometer indicates different composition in many cases, and interpolation onto different spectral sampling locations will shift/alias those)  It’s the type of thing that should only be done on a case by case basis and not done automatically.

Put simply, these aren’t bands, and if you’re trying to use each spectral sample as an independent band, your analysis will yield poor results. 

It’s important not to do band math or similar approaches on raw sample points.  If you’d like to use a band math like approach, you’ll want to first combine multiple samples to get something similar to a “band” from another satellite.   Otherwise, you’re going to have poor SNR and often incorrect values.  Typically, you’d do that by convolving the spectral response of the band you’d like to approximate with the spectral sample points from the hyperspectral data.  In other words, a “band” is a weighted average of many different samples in Tanager data.  

If you’re doing more typical hyperspectral analysis, you’d be looking for spectral signatures, rather than bands. You’d be looking for the spectral location of certain dips and trying to measure where the peak of that dip is exactly between different Tanager samples.  Alternatively, you might be trying to match spectra to features in a spectral library. Either way, you’d be analyzing things in a very different manner than you typically would for multispectral data. 

In a sense, hyperspectral is more like vector data than raster data.  Each pixel is a linestring through spectral space. The exact vertex locations aren’t expected to be identical, and instead you’re looking at the overall shape.  Typically, you’d work with the basic products to avoid resampling as well, and so each “pixel” is a different size and represents a vector region rather than an evenly-spaced raster pixel.