A Study of Planet’s Dove Satellites’ Radi A Study of Planet’s Dove Satellites’ Radiometry ometry against Sentinel against Sentinel -2 over U.S. Farmland 2 over U.S. Farmland Joshua Greenberg, Planet Labs, Inc. Otago, New Zealand – July 31, 2019
Introduction Singapore Strait, Singapore – July 29, 2016
The Planet Fleet Flock Dove Classic RapidEye SkySat Dove-R Sensor Type Four-band frame Multispectral push- Multispectra / Four-stripe push- Imager with a split- broom Panchromatic push- frame Imager frame NIR filter frame Spectral Bands Blue: 455 - 515 nm Blue: 440 - 510 nm Blue: 450 - 515 nm Blue: 490 nm Green: 500 - 590 nm Green: 520 - 590 nm Green: 515 - 595 Green: 565 nm Red: 590 - 670 nm Red: 630 - 685 nm Red nm Red: 605 - 695 Red: 665 nm NIR: 780 - 860 nm Edge: 690 - 730 nm nm NIR: 740 - 900 NIR: 865 nm NIR: 760 - 850 nm nm Pan: 450 - 900 nm Orbit SSO SSO SSO SSO GSD ~3.0 m ~6.5 m ~1.0 m (~0.8 pan) ~3.0 m Frame Size/Swath ~ 24.6 km x 16.4 km 77 km ~ 3.2 km x 1.4 km ~ 26 km Width (single camera) Crossing Time 9:30 - 11:30 am 11:00 am 10:30 - 13:00 9:30 - 11:30
The Planet Fleet Flock Dove Classic RapidEye SkySat Dove-R Sensor Type Four-band frame Multispectral push- Multispectra / Four-stripe push- Imager with a split- broom Panchromatic push- frame Imager frame NIR filter frame Spectral Bands Blue: 455 - 515 nm Blue: 440 - 510 nm Blue: 450 - 515 nm Blue: 490 nm Green: 500 - 590 nm Green: 520 - 590 nm Green: 515 - 595 Green: 565 nm Red: 590 - 670 nm Red: 630 - 685 nm Red nm Red: 605 - 695 Red: 665 nm NIR: 780 - 860 nm Edge: 690 - 730 nm nm NIR: 740 - 900 NIR: 865 nm The present study will look at NIR: 760 - 850 nm nm Pan: 450 - 900 both Dove Classic and Dove-R. nm Orbit SSO SSO SSO SSO GSD ~3.0 m ~6.5 m ~1.0 m (~0.8 pan) ~3.0 m Frame Size/Swath ~ 24.6 km x 16.4 km 77 km ~ 3.2 km x 1.4 km ~ 26 km Width (single camera) Crossing Time 9:30 - 11:30 am 11:00 am 10:30 - 13:00 9:30 - 11:30
Dove-R Radiometric Calibration Planet’s current radiometric calibration methodology primarily makes use of well-characterized desert pseudo-invariant sites, a regime far removed from many actual use cases -- e.g. Ag!
Motivation for Farmland Crossover Study Planet must understand the behavior of our agricultural data so we can communicate to customers. Moreover, we must understand the behavior of our radiometry over the full dynamic range, not just bright desert sites.
Motivation for Comparison to Sentinel -2 In Blue, Green, and Red, Dove- R has similar central Dove-R wavelengths and band widths to Sentinel-2a’s and -2b’s B2, B3, and B4 respectively. This study compares the Doves’ NIR band to Sentinel-2’s B8 . A comparison to B8A is planned, too. Sentinel-2a
Motivation for Comparison to Sentinel -2 Dove Classic (0f44) The comparisons between Dove Classics’ RSRs and Sentinel-2’s are not as close for RGB (though the comparison of NIR to B8 is more reasonable). Sentinel-2a
Identifying farmland: USDA’s Cropland Data Layer “ The Cropland Data Layer (CDL) is a crop-specific land cover data layer created annually for the continental United States using moderate resolution satellite imagery and extensive agricultural ground truth. The CDL is created by the USDA, National Agricultural Statistics Service (NASS), Research and Development Division, Geospatial Information Branch, Spatial Analysis Research Section.” -- developers.google.com (and it’s free)
Cropland Data Layer (USDA provides an awesome website for accessing the CDL.) CropScape: A Web service based application for exploring and disseminating US conterminous geospatial cropland data products for decision support Weiguo Han, Zhengwei Yang, Liping Di , Richard Mueller
Extracting Farmland Polygons from CDL ● Planet’s crossover infrastructure requires polygon AOIs to search for crossovers ● CDL is polygonized by finding contiguous, majority-crop regions of over 45 km 2 area. ● Identity of crops is discarded (for now) .
Note on CDL vs 2008 CLU Sublette County, WY -- 2008 CLUs Zoomed in The official title of this talk references a plan to use the USDA Common Land Units (CLUs). Because I encountered difficulties working with the 2008 CLU dataset, I opted to work with the 2018 CDL instead.
Example Crossover Tile May 2, 2019 (RGB TOA-rad shown). 1057 (Dove-R) at ~ 10:30 AM local S2B at ~ 10:37 AM Tile 1055524 in UTM zone 10 (25km ⨉ 25km)
Sampling Subtiles Each UTM tile was broken down into 100 2.5km ⨉ 2.5km subtiles, numbered 00 to 99. Here is subtile 48 : 1057 (Dove-R) S2B
Sampling Subtiles, cont. Ideally the 2018 CDL itself (or CLU dataset) 1057 (Dove-R) S2B would be used at this point to attempt to sample only crop pixels (and to guess at their identity). That’s not yet implemented; for this analysis I used an NDVI filter but it’s imperfect. :) Products after NDVI filtering and sampling each to 4m GSD
Comparing to 2018 CDL for Fun 2018 CDL 1057 (Dove-R) S2B Idle cropland Grapes Grass
2D Histogramming Pixel value pairs are histogrammed: density scatter plots are shown, for this subtile, for B, G, R, N.
2D Histogramming: Summary Statistic Zooming in... Fitting a Gaussian to the histogram “peak” allows one to find a mode, an x-y pair that expresses the relationship between the band values for this subtile. Here, the Sentinel-2B mode value is 1522.4 which corresponds to ~ 65.8 W/sr/μm/m 2 radiance in its red band, vs. ~ 63.6 for the Dove sat.
All Subtile Modes for a Crossover Event Each subtile contributes a single x-y point, its 2d histogram mode. A single crossover event -- a crossing of the swaths of one Dove and Sentinel-2 -- can cross multiple UTM tiles, each of which may have many subtiles with usable pixels. The 1057 vs. S2B crossover event on 2019-05-02 contributed over 2000 points (before subsampling...).
Subsampling I analyzed over 28000 distinct crossover events, representing 6 million subtiles (1.6 trillion pixels). This is the largest study I’ve ever attempted on Planet data. Unfortunately, I have not been able to scale our statistical analysis code to process this amount of data before JACIE! To accelerate the analysis, I chose 39 Doves (34 Classic, 5 Dove-R), and for each crossover event, randomly chose just 5 subtiles. This decreased the amount of data to analyze by a factor of 600.
All Crossover Events for a Single Dove Sat All crossover events for 0f44 after subsampling: Btw, note that high variance appears within data f or a single sat.
All Measurements for All Dove Classic, Plotted Together All crossover events for Dove Classic vs Sentinel-2 A and B, after subsampling:
All Measurements for All Dove-R, Plotted Together All crossover events for Dove-R vs Sentinel-2 A and B, after subsampling:
Dove-to-Sentinel-2 Crossovers, counts by month Counts distinct crossover events that contributed at least one usable data point after subsampling: approx 2000 total
Fitting Method A hierarchical linear model is used. It accounts for: ● Within-event clustering (by using random per-crossover- event intercepts) ● Correlation of those intercepts with event mean radiance ● Similarity of sats to each other - - the flock mean slope can be a prior for each sats’ slopes.
All Dove Results Slopes of fit to all data (grand mean) 1.14 1.12 1.28 0.99
Distribution of Sat Slopes Slopes for each sat processed so far (34 Classic, 5 Dove-R) 1.14 ± 0.02 1.12 ± 0.02 1.28 ± 0.05 0.99 ± 0.02 (...at 1 SD, across population of sats)
Dove Classic Percent Differences vs. Sentinel -2 Radiance
Dove -R Percent Differences vs. Sentinel-2 Radiance
Summary London Array Wind Farm, United Kingdom – April 17, 2016
Summary ● A study has been conducted comparing Dove sats to Sentinel-2 in near-simultaneous crossovers over farmland AOIs derived from USDA’s Cropland Data Layer. ● This enables us to say something about the relative and absolute errors in the Dove top-of-atmosphere radiance products over a more diverse set of landcover, and dimmer landcover, than heretofore. ● At the bottom of the dynamic range, relative errors are high, although the behavior is better with Dove-R than Dove Classic.
Summary, continued ● Future work will use specific crop information from CDL, and will take additional effects into account (e.g. BRDF). ● Eventually this analysis will be able to be automatized and provide regular feedback on how Planet stacks up to other data providers when it comes to cropland.
go.planet.com/explore19
Recommend
More recommend
