Bad spectral data
is invisible.
Until now.

MicaSense RedEdge-MX Dual · 10-band multispectral

Sampled at the field spectrometer’s recorded GPS positions, the orthomosaic differs in spectral shape from the field spectra at the five brightest points by 33.6°, against a 12° limit, so SpectrIQ set that comparison aside rather than report it. That is consistent with the roughly 12 m horizontal accuracy the release documents for those GPS fixes, against 4 cm pixels; it is not a defect in the release. Extra calibration work did not beat the manufacturer default on this dataset. A reduced run: 3 of the 46 imaging checks were selected for this capture.

Itres CASI-1500 · 48-band hyperspectral

Each sidecar carries three extent records. Two locate the raster correctly; the third is an unpopulated template default: the same fixed box near Stanwood, Washington in every sidecar inspected, across two deliveries in different years and different states. Because the box is fixed and the imagery is not, how far wrong it reads depends on the delivery: about 3,040 km for the Alaska tiles, about 3,520 km for the New York ones. It is the GeoBndBox at dataIdInfo/dataExt/geoEle with no esriExtentType attribute (the standard ISO 19115 path, which is the one a consumer reads), and is checkable in about ninety seconds. A packaging defect in the discovery metadata, not an imagery one: the GeoTIFF georeferencing is correct and the imagery is fully usable. And the same square kilometre, published twice, disagrees by a median 16.5% per band, in relative digital number; no radiometric units are declared. That is SpectrIQ’s figure, from a blind run; a separate manual analysis of the same pair landed at 16.2%. A reduced run: 11 of the 46 imaging checks were selected for these tiles.

RGB framing camera + DSS natural colour · 3-band

No defects found in the imagery. Worth passing on: the two Portland years ship in different coordinate reference systems: same zone, same pixel, different datum declaration.

Surface temperature · single-band thermal

No defects found in the data: five windows over Phoenix and Lake Mead, all passing, including one placed deliberately across the swath edge to see whether any check would read fill as measured ground. On a calibrated satellite product 13 of the thermal suite’s 41 checks can execute; the rest decline, each with its reason, because the sensor state they need is not in a delivered product.

NEON Imaging Spectrometer · 426-band hyperspectral

The widest product yet: 426 bands, 380–2510 nm, and three different spectrometers across two collection years. All six captures pass. One check, Spectral Smoothness, flags every capture as advisory: 15–21% of pixels carry spectral spikes, concentrated in the bluest bands near 380–410 nm, and the cause is not yet classified. The same ground in two collection years differs by a median 6.9–11.4% per band in reflectance; the years were flown by different instruments, so that difference is measured but not attributed. NEON publishes more of what the checks need than any product here: band centres, per-band FWHM, a declared scale factor, a declared fill value and a per-pixel atmospheric classification, all inside the delivered file. A small number of metadata observations go to NEON before they are published anywhere.