Earth observation by constellations of CubeSats: New opportunities and challenges

Rasmus Houborg Author

11/15/2018 Added

90 Plays

Description

CubeSats

Searchable Transcript

Search:

[00:00:00.690]Thank you.
[00:00:02.400]Let's see how long I last in South Dakota with the winters.
[00:00:07.270]All right, so this talk is gonna be about CubeSats.
[00:00:12.020]It's really kind of an exciting new area I think,
[00:00:13.940]and I started looking into those new kind of sensors
[00:00:17.610]about two or three years ago,
[00:00:19.020]and it really, really excited my passion for doing science.
[00:00:23.450]This talk is really about how to exploit
[00:00:27.132]this type of new resource using an agnostic sensor data
[00:00:32.540]fusion approach that takes into consideration
[00:00:36.860]other more traditional sensor types like Landsat
[00:00:40.470]and Sentinel Two.
[00:00:42.230]So, it's really an exciting area
[00:00:44.170]for remote sensing and accu-culture monitoring
[00:00:45.780]as I see it with these new observational paradigms
[00:00:49.460]and the general proliferation of sensor data.
[00:00:53.221]There's so much new data coming alive.
[00:00:55.150]And then who would imagine that today we are able
[00:00:59.290]to do daily imaging at three minute solution globally?
[00:01:02.260]So, that wasn't something that you would imagine feasible
[00:01:04.410]just a few years ago.
[00:01:05.960]Obviously, these new data sources
[00:01:07.570]doesn't come without limitations.
[00:01:09.840]And as I see, you need to use it.
[00:01:12.000]With other more conventional resources.
[00:01:20.150]This talk is about kind of presenting this technique,
[00:01:23.200]for integrating these different data sources,
[00:01:25.690]but it's also kind of...
[00:01:28.396]I also want to show or demonstrate,
[00:01:30.130]how to evaluate how good our SmallSat data really is.
[00:01:32.680]Is it really useful for doing scientific analysis?
[00:01:36.023]Delivering products of sufficient quality,
[00:01:38.670]for doing science and application for case and studies.
[00:01:47.120]There we go
[00:01:48.380]So the end idea, the CubeSat.
[00:01:51.377]And for those who don't know it,
[00:01:54.645]it's a small satellite that has a single unit size
[00:01:57.130]of 10 times 10 centimeters
[00:01:59.690]and a mass list of one point three kilograms.
[00:02:04.070]The basic concept is really to expand the observation
[00:02:07.330]potential by launching flocks of these small CubeSats
[00:02:11.461]that are based on a very cheap commercial
[00:02:14.820]after-shelf components, which makes constellations
[00:02:18.330]of hundreds of these in space economically feasible.
[00:02:23.170]I'm doing configure these in a number
[00:02:25.356]of comparable sizes to increase
[00:02:27.380]potential application fields,
[00:02:29.280]including the three U, six U, and 12 U sizes;
[00:02:34.060]and the last number of companies that kind of
[00:02:35.760]taking up on this new technology,
[00:02:37.860]including Planet Labs and Astro Digital
[00:02:40.120]and a number of other kind of companies.
[00:02:43.060]My work has focused on using data from Planet.
[00:02:52.800]So, they operate the largest constellation
[00:02:55.270]of CubeSats in space.
[00:02:57.473]They adopted three U standard
[00:03:01.200]and about four kilograms per CubeSat,
[00:03:04.100]currently have around 150 active CubeSats in space
[00:03:09.100]and that kind of facilitates daily imagining
[00:03:12.820]at three to four meter resolution.
[00:03:16.590]So for instance, when we look at commercial satellites,
[00:03:19.640]they will take years to produce and cost you a fortune.
[00:03:23.250]Planet can produce 40 satellites in just one week
[00:03:26.690]at a tiny fraction of that cost.
[00:03:28.940]So, this is a completely new, really a game changing
[00:03:31.430]kind of resource that we're looking at here.
[00:03:35.580]Obviously, the special kind of resolution
[00:03:38.480]is a bit limited.
[00:03:40.330]It's only four bands that are quite broad
[00:03:42.640]in the red, green, blue, and the infrared.
[00:03:51.527]All right.
[00:03:56.400]I just want to show this animation
[00:03:57.798]just to show how they orbit in space.
[00:04:01.304]The cubes assign a center corner as orbit
[00:04:04.343]and they orbit in a train.
[00:04:06.380]It's pretty unique and it gives you this very high,
[00:04:09.332]unprecedented spatiotemporal frequency.
[00:04:19.301]And the spatial advantage is superior
[00:04:21.702]to that what you can achieve
[00:04:23.194]from conventional system like Landsat.
[00:04:25.370]This compares the Landsat 30 meter
[00:04:27.090]with the Planet Scope, which is the Planet's CubeSat data,
[00:04:30.820]three meters in data and you can really appreciate
[00:04:33.182]the enhanced spatial resolution.
[00:04:35.342]You're able to resolve individual trees
[00:04:38.750]and shadows from the planet data,
[00:04:41.410]as well as intra field variability
[00:04:45.460]that you won't be able to see with LandSat-like information.
[00:04:52.770]But at the same time, these CubeSats
[00:04:54.233]are not without limitations.
[00:04:56.580]As mentioned, they're more spatially limited,
[00:04:58.990]they only have four bands,
[00:05:01.029]and they have significant overlap in the visible domain.
[00:05:06.950]This shows an average spatial response
[00:05:08.963]of a selection of CubeSats
[00:05:11.000]compared to that of LandSat eight.
[00:05:12.560]Kinda the shading around each of these curves
[00:05:15.840]represent the standard deviation
[00:05:17.330]between each individual CubeSats.
[00:05:19.830]Each CubeSat has a unique spatial response
[00:05:23.020]and you need to account for that difference
[00:05:25.620]when you wanna create a continuous
[00:05:28.480]or consistent time series of data.
[00:05:31.220]As you can see, it's quite different
[00:05:32.870]from the spatial response you get from LandSat,
[00:05:34.663]that has much narrow bands,
[00:05:37.210]particularly in the visible domain.
[00:05:39.490]And this will result in inconsistencies
[00:05:41.380]or if you compare a common spatial metric,
[00:05:43.860]like the NDVI, it will not be consistent
[00:05:46.160]between the CubeSat and a LandSat observation.
[00:05:49.670]The CubeSat data would be much smaller
[00:05:51.570]in terms of NDVI and the dynamic range
[00:05:55.160]will be significantly smaller as well.
[00:05:57.350]So, you need to correct for that
[00:05:59.908]in order to make the data really useful,
[00:06:02.710]and for that purpose, I've developed a CubeSat-enabled
[00:06:05.510]spatiotemporal enhancement method called CESTEM,
[00:06:10.190]kinda explores these synergies and cross paths
[00:06:12.680]from observation by integrating MODIS, LandSat,
[00:06:16.250]and Sentinel Two as a reference for calibrating
[00:06:19.180]your CubeSat data.
[00:06:21.200]It serves as a multi-purpose tool
[00:06:24.690]for when you're measuring normalization.
[00:06:27.430]Phenology reconstruction, via NDVI for instance,
[00:06:30.900]as well as spatiotemporal enhancement
[00:06:33.210]of biophysical properties, including Leaf Area Index.
[00:06:42.197]So, this shows you a kind of a conceptual diagram
[00:06:44.120]of what CESTEM does.
[00:06:45.310]One one side, you have these satellites,
[00:06:49.842]the big satellites that cost quite a fortune to build
[00:06:55.430]and takes quite a bit of time to put into space,
[00:06:58.360]but they're rigorously calibrated
[00:07:00.080]and can be regarded as the gold standard,
[00:07:02.240]but then they're affected by the classical trade off
[00:07:05.190]between spatiotemporal resolution.
[00:07:07.370]This simply cannot satisfy both of these conditions.
[00:07:11.010]So, there's either gonna be high spatial
[00:07:12.610]or high temporal, rarely both.
[00:07:15.290]On the other side, you have SmallSats
[00:07:16.960]that are low cost, high volume system
[00:07:19.970]that are launched in flocks,
[00:07:21.970]which gives you a superior spatiotemporal resolution,
[00:07:24.132]but then again, they were basically built in a garage
[00:07:26.990]and don't have that same degree of calibration
[00:07:29.450]as you get from the conventional systems.
[00:07:32.000]So, these contrasting data sources or data stream
[00:07:34.720]feed into system, that then applies a number
[00:07:40.130]of data fusion and machine-learned techniques
[00:07:42.890]to kind of fix issues in the original data sets
[00:07:45.880]in order to produce and enhance the high quality product
[00:07:50.320]that is suitable for doing scientific applications
[00:07:53.160]and for delivering real insights of real value
[00:07:56.910]to the end-users.
[00:08:02.740]All right.
[00:08:03.720]So, this exemplifies the approach system framework
[00:08:07.550]for the MODIS, LandSat, and CubeSat.
[00:08:13.490]The general idea is to use, in this case,
[00:08:15.717]MODIS and LandSat surface reflectances as a reference
[00:08:19.490]to calibrate your raw CubeSat data,
[00:08:22.540]being digital numbers or top of atmosphere reflectances
[00:08:25.610]into a LandSat-consistent surface reflectances.
[00:08:30.470]CESTEM has the ability to draw the reference data
[00:08:32.890]from LandSat data from any scene in the past
[00:08:36.670]and then it uses down-scaled MODIS data
[00:08:40.060]to kind of assess or determine the change
[00:08:44.260]in cover conditions over the various acquisition time spans.
[00:08:48.820]So, by drawing LandSat data from past images,
[00:08:52.000]it establishes a map for each individual CubeSat scene
[00:08:55.770]and then uses that as a training for creating a model
[00:08:59.070]to predict LandSat-consistent surface reflectances
[00:09:02.240]at the spatiotemporal resolution of the CubeSat data.
[00:09:05.823]So, it basically brings your CubeSat data in line
[00:09:07.870]with your LandSat surface reflectance data,
[00:09:11.300]then you can then produce a time series
[00:09:13.770]of a spectral metric, like NDVI.
[00:09:17.375]And this is kind of a sensor agnostic approach,
[00:09:19.840]so you can also substitute LandSata eight with Sentinel Two
[00:09:23.510]and you can even use the system to calibrate
[00:09:25.660]other sensors that might need calibration,
[00:09:28.050]be it other CubeSat sensors or be it a UAV sensors.
[00:09:32.305]There's no requirement in terms of the input
[00:09:35.200]of the data to be calibrators and can be raw data.
[00:09:39.310]So, that's a powerful mechanism,
[00:09:40.700]a powerful element of this approach.
[00:09:44.017]Let's see.
[00:09:49.110]All right.
[00:09:49.943]So, CESTEM has also been extended to map
[00:09:53.260]biophysical plan buffers group properties
[00:09:55.690]at a CubeSat resolution and this approach can use
[00:10:00.120]either LandSat or Sentinel Two.
[00:10:02.387]It kind of exploits the enhanced spectral resolution
[00:10:05.460]of those sensors to better retrieve or better constrain
[00:10:09.030]your inversion process.
[00:10:11.690]I use a sale prospect model immersion approach
[00:10:17.160]and kind of combine that spatially enhanced data
[00:10:19.520]with your enhanced spatial information
[00:10:23.510]from the CubeSat data.
[00:10:26.360]And I will show you that a bit later.
[00:10:29.210]So first, I'll demonstrate the utility of system
[00:10:33.370]for radiometric normalization
[00:10:35.410]and this, I will base this on studies in Saudi Arabia,
[00:10:40.500]an agricultural farming area,
[00:10:43.820]consistent of an agricultural pivot
[00:10:46.070]about 800 meters in diameters.
[00:10:48.160]It's a combination of corn and alfalfa, primarily.
[00:10:55.760]So, this compares the NDVI from CubeSat and LandSat eight,
[00:11:01.282]the normalized frequency distribution shows NDVI
[00:11:05.151]from CubeSat data based on both
[00:11:07.430]top of atmosphere reflectances and Six S
[00:11:10.340]corrective surface reflectances
[00:11:11.760]compares with LandSat eight,
[00:11:13.470]and you do see a significant shift
[00:11:15.940]that's definitely not consistent.
[00:11:18.258]You can see the same tendency in the tendency scatter plots
[00:11:22.400]as significant bias between your LandSat and CubeSat NDVI,
[00:11:28.000]even after you correct for the atmospheric effects,
[00:11:31.664]there's significant bias and a mean absolute deviation
[00:11:34.190]of around 25% on average for this given day,
[00:11:40.090]but then when you apply the system algorithm,
[00:11:42.260]you're really able to kinda bring
[00:11:43.690]these different data sets together.
[00:11:46.830]They now align quite nicely and you would use
[00:11:48.540]significantly reduced bias to basically eliminate the bias
[00:11:53.090]and end up with a mean absolute deviation
[00:11:54.580]of around two percent for this particular day.
[00:12:02.130]And you can also apply this to the individual spatial bands
[00:12:05.660]and this shows an independent validation
[00:12:06.890]of the system derived reflectances
[00:12:10.450]against day coincident LandSat eight reflectances
[00:12:13.880]and then again, you see you're able to quite accurately
[00:12:16.900]reproduce the LandSat eight data using this approach.
[00:12:29.892]So now, I'm looking into how to use this
[00:12:32.990]for reconstruction phenology via NDVI
[00:12:36.490]and this shows a time series of an alfalfa field
[00:12:40.387]in Saudi Arabia.
[00:12:41.670]It's a six month period.
[00:12:44.006]We have different time series here.
[00:12:45.699]The blue one...
[00:12:46.877]Sorry.
[00:12:48.570]Yeah, the blue one is the system correct time series
[00:12:51.240]and the red's kind of a square of the LandSat eight data,
[00:12:56.240]and then I'm mostly showing the original,
[00:12:58.820]uncorrected CubeSat data.
[00:13:00.960]You have the top of atmosphere reflectances,
[00:13:02.413]as well as the surface reflectances
[00:13:04.340]corrected via Six S, and you can clearly see
[00:13:06.990]how the original data is really quite noisy
[00:13:09.360]over this site, but when you apply the CESTEM algorithm,
[00:13:13.160]you are able to kind of reproduce a sensible
[00:13:15.790]and meaningful time series that tracks what you see
[00:13:19.127]in LandSat quite nicely.
[00:13:29.090]And below is just an animation,
[00:13:30.680]kind of can see what's going on in the spatial domain
[00:13:34.500]for this specific pivot.
[00:13:36.550]Again, you can see there's quite a bit of fluctuation
[00:13:38.840]in the top of atmosphere data
[00:13:41.150]whereas the CESTEM corrected data is more stable in time,
[00:13:44.420]it's more consistent, what you would expect
[00:13:47.592]to observe in real time, in a real kind of setting.
[00:13:53.156]Let me just see.
[00:13:55.890]And this shows a fast forward animation of that,
[00:13:58.552]showing the CubeSat data before and after applying CESTEM
[00:14:04.325]and, in this case, you can also get a good appreciation
[00:14:06.680]of the noise in the original data and the capacity
[00:14:09.820]of CESTEM to correct for that and produce
[00:14:12.599]a meaningful NDVI time series.
[00:14:27.825]This one.
[00:14:31.710]Yeah, might work.
[00:14:32.543]All right.
[00:14:39.180]So, we also done some studies of alfalfa
[00:14:43.000]in Southern California that I will show next.
[00:14:46.846]So, this is another alfalfa site, multicolored alfalfa,
[00:14:50.420]a 10 month period, and again showing the raw CubeSat data
[00:14:55.530]in black and then the corrected CESTEM corrected data
[00:14:58.770]in while circles, and the Sentinel Two data,
[00:15:02.570]so I used Sentinel Two data in this case
[00:15:04.586]as a reference for calibrating the CubeSat data
[00:15:07.970]and you can see that the corrected data
[00:15:11.050]quite nicely matches the Sentinel Two data
[00:15:14.400]and produce a more realistic time series.
[00:15:18.050]It kind of corrects some of the dips and peaks
[00:15:20.240]you see in the raw data.
[00:15:24.063]And you also get an appreciation of, even with Sentinel Two,
[00:15:26.850]over this clear sky site, you do have quite a few
[00:15:29.440]observations, but you're still gonna miss critical periods
[00:15:33.330]of the growing seasons and the growing season,
[00:15:36.002]you have 10 growing season over a 10 month period,
[00:15:39.299]it's a really kind of rapid kind of finale
[00:15:41.690]that you need really frequent observations
[00:15:44.640]to capture sufficiently.
[00:15:48.930]This just highlights one of these growing seasons
[00:15:51.220]and you can get a feel of the spatial and temporal
[00:15:54.669]advantages of the CubeSat versus the, in this case,
[00:15:58.560]20 meter Sentinel Two data.
[00:16:10.540]And in this case, I've run it with one system
[00:16:14.210]with both LandSat and Sentinel Two independently,
[00:16:17.200]so I'm just seeing how they compare.
[00:16:19.860]If I do that and you get a pretty good agreement,
[00:16:22.250]fairly consistent results in both cases,
[00:16:26.020]I would expect some disagreement because LandSat
[00:16:28.687]and Sentinel Two themselves are not consistent
[00:16:31.077]in terms of the spectral configuration.
[00:16:33.360]You also have BIDF effects.
[00:16:35.630]Ideally, I would probably run CESTEM
[00:16:38.337]kinda using a combined harmonized Sentinel Two
[00:16:40.570]LandSat eight data set as input,
[00:16:42.490]as that will give me more training data
[00:16:44.590]and that will be a more consistent data set
[00:16:47.080]for doing the reference sampling,
[00:16:49.590]but that's the next step.
[00:16:52.660]So now, turning to Nebraska,
[00:16:56.920]looking at corn and soybean sites.
[00:17:00.690]So, this is a corn site, this is the NE one site
[00:17:04.490]in Mead in 2017.
[00:17:07.640]Again, you see the raw CubeSat data
[00:17:12.370]compared to the CESTEM corrected data
[00:17:14.640]and the Sentinel Two data as well
[00:17:17.000]and then you have a good consistency between the CESTEM
[00:17:18.917]and the Sentinel Two data
[00:17:20.620]and clearly your CubeSat data adds a lot of information
[00:17:25.340]between your Sentinel Two acquisitions
[00:17:28.920]that could be quite useful, even thought not a whole thing
[00:17:32.400]is happening in this case, but if you had to experience
[00:17:35.740]some kind of stress, this model kind of disturbance,
[00:17:39.542]the CubeSat data would be really helpful
[00:17:41.635]determining the exact timing of that even.
[00:17:47.446]And yes, you do have significant gaps
[00:17:49.800]in your Sentinel Two record,
[00:17:51.060]such as during the green up and during senescence
[00:17:54.930]and these are periods where a lot of things are going on,
[00:17:58.020]where we really need information to tell what is going on,
[00:18:01.589]so this is where data from CubeSat can be really helpful.
[00:18:08.190]With CubeSats, they have issues with clouds too
[00:18:11.188]and so I'm always been looking into technique
[00:18:13.806]to correct for that.
[00:18:16.000]I've been using some automated image feature recognition,
[00:18:20.810]optic-based feature recognition techniques for doing that
[00:18:26.402]and while it does a pretty good job
[00:18:29.000]in identifying the main structures,
[00:18:30.850]you still have areas where we don't really
[00:18:34.930]capture the contamination, particularly from the shadows.
[00:18:38.340]It can be quite difficult.
[00:18:41.410]So in those cases, we're gonna get a contaminated signal
[00:18:44.747]that's not really that useful,
[00:18:46.800]so I also been trying to correct for that using some...
[00:18:52.090]Let me see.
[00:18:54.568]There we go.
[00:18:55.401]Using some...
[00:18:56.234]Okay.
[00:18:58.619]All right.
[00:19:00.090]Using some gap-filling techniques.
[00:19:01.770]So, the image on your left is the gap-filled one
[00:19:10.073]and you can see, compared to the previous one,
[00:19:12.480]in this case you get rid of the cloud shadows
[00:19:16.220]and you get a realistic-looking image
[00:19:21.090]compared to the one that has not been cloud-filled
[00:19:25.230]or corrected for these cloud shadows.
[00:19:28.590]So, it seems to be working pretty well for this case
[00:19:33.174]and this is just another case,
[00:19:34.007]this is the NE one and NE two fields
[00:19:36.402]where you can see the shadows in one image
[00:19:39.014]is removed than the other image,
[00:19:42.910]in addition to the cloud-detected areas
[00:19:45.470]being restored with real values.
[00:19:51.120]So, this is the time series that include
[00:19:55.280]the gap-filled data points.
[00:19:57.060]So, I think I've added about 10 scenes.
[00:19:59.870]In principle, I could apply these techniques
[00:20:01.780]to add daily imagery, but at the moment
[00:20:04.477]I'm just kinda filling the scenes that I have,
[00:20:07.940]but I could apply to get a daily time series
[00:20:11.507]of CubeSat-scale data if I wanted to,
[00:20:15.370]but you can see here how it kind of restores the signal.
[00:20:18.465]You have the regional top of atmosphere reflectance data
[00:20:21.619]on one side and you have the CESTEM-corrected
[00:20:24.209]surface reflectance data has been corrected
[00:20:26.611]and you can see how it is able to remove the shadow
[00:20:29.270]in this case in NE two.
[00:20:32.566]In this case, you don't have any data
[00:20:34.740]in the original data, but you're able to restore it
[00:20:36.785]using my gap-filling approach.
[00:20:41.020]Same in this case.
[00:20:42.310]A lot of contamination going on in the original data,
[00:20:45.270]but then are able to get a pretty
[00:20:46.790]realistic-looking image like that.
[00:21:00.790]Right, so I've also been looking at validating
[00:21:03.410]these time series from Planet
[00:21:06.900]using data that Trenton has provided,
[00:21:12.210]as been mentioned by the Arable Mark's penetrometer.
[00:21:16.410]Obviously, the spatial conflation is not identical
[00:21:20.021]between Sentinel Two and the Arable Mark,
[00:21:25.310]but still fairly comparable,
[00:21:27.360]at least in giving idea about the phenology.
[00:21:29.740]Is it realistic?
[00:21:30.946]Is it realistic what we see from Planet?
[00:21:32.680]From the CESTEM-corrected Planet data.
[00:21:35.650]It's also used as the metric for evaluating
[00:21:37.782]the raw data from Planet.
[00:21:41.360]So, this is a validation of a soybean in Mead, Nebraska,
[00:21:47.910]comparing the CESTEM-corrected surface reflectances
[00:21:50.700]against the institute observations in 2018
[00:21:55.007]and you get a pretty good fit.
[00:21:56.680]Some bias is going on.
[00:21:58.080]It could be due to different spacial conflation.
[00:22:00.290]It could also be a scale issue,
[00:22:01.830]it could be an issue with the observation time, as well,
[00:22:08.060]but that apart, you get a pretty good correlation.
[00:22:10.380]Point nine eight seven for this site.
[00:22:13.420]And when I compare it to the...
[00:22:15.766]So, Planet also produces a surface reflective product
[00:22:21.530]that I'll show you here, in blue,
[00:22:24.800]and there's obviously quite a significant amount
[00:22:26.890]of scatter in that product.
[00:22:29.070]It does kinda reduce the bias a little bit,
[00:22:31.860]which might be something to do with the more
[00:22:34.190]similar spectral configuration.
[00:22:36.211]It has broader bands, more comparable to the Arable Mark,
[00:22:39.330]but then you have some significantly pronounced scatter.
[00:22:45.830]This is another soybean site,
[00:22:48.380]comparing the institute observations
[00:22:50.370]against the CESTEM-corrected surface reflectances
[00:22:52.710]during 2018 and, in this case,
[00:22:55.160]the agreement is quite exceptional.
[00:22:56.850]It's point nine nine.
[00:23:00.480]So, yeah.
[00:23:01.840]So, this might indicate that...
[00:23:04.370]I think it's difficult maybe.
[00:23:05.550]The bias can be a bias due to different kind of instruments.
[00:23:08.880]The Arable Mark might not be completely consistent
[00:23:11.890]between them individually,
[00:23:13.580]but in this case, they match pretty well
[00:23:17.210]and when I compare it against the Planet
[00:23:19.570]surface reflectance product,
[00:23:21.400]you can see how the diverge is quite significant from there.
[00:23:25.470]That's significant scatter and has some serious issues,
[00:23:29.660]so it also indicates and shows you
[00:23:31.874]or demonstrates that you really need to correct
[00:23:34.001]the original Planet data if you wanna make it useful
[00:23:39.370]for scientific analysis.
[00:23:42.050]This is just a side by side comparison
[00:23:43.810]between the CESTEM surface reflectance product
[00:23:45.577]and the Planet surface reflectance product.
[00:23:48.460]Again, kinda demonstrating the different performances
[00:23:52.150]of these two products.
[00:23:55.980]So, this highly calibrated.
[00:23:58.350]So, once you have a highly calibrate NDVI
[00:24:00.420]data set like that, You can really start looking
[00:24:02.370]at intra field dynamics,
[00:24:04.520]like I'm doing here.
[00:24:05.570]This is the NE three rain-fed soybean field in 2018
[00:24:11.910]and we really see in the beginning
[00:24:13.270]you have a quite homogenous field conditions,
[00:24:16.140]but then beyond day 180, the curves start to diverge,
[00:24:20.780]maybe due to some intra field conditions;
[00:24:26.050]and looking at time series like this,
[00:24:29.200]having that time of data and a daily kinda frequency,
[00:24:31.980]it's really useful and powerful for identifying
[00:24:34.910]those kinda suboptimal conditions
[00:24:37.970]and taking the necessary preventative measures
[00:24:40.730]to kind of avoid that progressing any further.
[00:24:47.050]So, I've mostly been looking at translating
[00:24:50.370]this spectral information into more quantitative insights
[00:24:56.120]by using such as Leaf Area Index
[00:25:00.560]and this really uses...
[00:25:02.820]The idea is to take advantage of the spectrally
[00:25:04.840]enhanced data from sensors inside LandSat eight
[00:25:08.140]and particularly Sentinel Two
[00:25:10.100]and combine that with the superior spatial
[00:25:12.650]and temporal resolution of the CubeSat data,
[00:25:14.930]so really exploiting the synergies.
[00:25:18.700]In this first example I used, based on the fusion
[00:25:22.420]of LandSat and CubeSat data with our site in Saudi Arabia.
[00:25:29.360]So, you can see that the CubeSat data,
[00:25:31.150]particularly when you progress this,
[00:25:32.920]you really start to provide close to daily information
[00:25:37.180]on the layout dynamics and again matches
[00:25:39.983]the LandSat LAI quite nicely.
[00:25:44.570]This high-resolution information is really quite beneficial,
[00:25:48.490]particularly during the green up,
[00:25:50.210]as well as harvesting.
[00:25:51.790]You can really get an idea when the actual harvest started.
[00:25:59.370]When you look at a green up cycle,
[00:26:00.950]as in the bottom figure, this is an eight day sequence
[00:26:05.750]of Planet's Scope CubeSat data over an eight day period
[00:26:10.830]during the green up, and then it's bound by LandSat data
[00:26:14.310]on day 206 and 230.
[00:26:17.590]So, you can really get an appreciation
[00:26:19.633]of what you're gonna miss if you only have LandSat data
[00:26:22.560]on an eight day visit.
[00:26:25.092]I should probably mention that in this site,
[00:26:27.130]you have LandSat data every eight day,
[00:26:29.180]or this specific farm, due to overlapping swaths.
[00:26:32.160]But in most cases, you would have a 16 day interval,
[00:26:37.580]so you're gonna miss a lot of things
[00:26:39.690]and you can use the CubeSat data to feel out
[00:26:41.720]that amazing information, in addition to providing
[00:26:45.920]addition information on the infield stuff and intra field
[00:26:49.540]dynamics with the enhanced spacial resolution.
[00:26:55.220]And again, this just highlights the spacial and temporal
[00:26:59.870]advantages of the daily CubeSat LAI data
[00:27:04.730]compared to LandSat, so it just shows the timing
[00:27:08.730]and progression of a harvesting event
[00:27:11.420]that has commenced on day 181 and terminated on day 183,
[00:27:16.850]but when you only have LandSat data on one day,
[00:27:19.290]you're really gonna miss a lot of critical
[00:27:21.470]day to day variations in this specific case.
[00:27:29.440]And so, I've also been looking at the constructing
[00:27:32.330]and producing LAI products while in Mead, Nebraska.
[00:27:38.880]So, this uses the Sentinel Two data.
[00:27:43.170]The previous work on LandSat that I used
[00:27:44.970]is a bit more empirical LAI retrieval approach.
[00:27:49.130]In this work, I've been using the Pro-Cell model
[00:27:51.820]so that combines the prospect,
[00:27:54.090]Leaf Optical Probability model,
[00:27:55.853]with a sale type of canopy bi-directional reflectance model
[00:28:00.430]and it uses the Sentinel data to invert that model,
[00:28:03.570]giving the additional spectral bands in the red edge
[00:28:06.450]and the shortwave infrared, you can better constrain
[00:28:10.040]the inversion process.
[00:28:11.210]It wouldn't be feasible to invert Pro-Cell
[00:28:13.180]using just the Planet scope, just the CubeSat data,
[00:28:16.300]given the limited spectral information,
[00:28:18.650]but I can use that information that I get from Sentinel Two
[00:28:22.610]as a reference to kind of enhancing the LAI information
[00:28:26.430]in space and time.
[00:28:27.920]And this is shown here,
[00:28:31.040]showing my Sentinel Two retrievals
[00:28:34.090]and the added spatial temporal information
[00:28:36.360]of the CubeSat data with this corn field in 2017.
[00:28:48.370]And then, I've also been looking at validating
[00:28:50.700]those kind of retrievals using the destructive
[00:28:55.520]institute measurements from this facility.
[00:28:59.149]In this plot, I'm just kind of showcasing
[00:29:01.930]a few time series from any three different points
[00:29:08.110]in any three, showing both the CESTEM-corrected,
[00:29:13.710]CESTEM-based LAI, the Sentinel Two LAI,
[00:29:16.860]and then your field measured LAI as the blue triangles,
[00:29:23.650]and in general, you get a pretty good agreement
[00:29:26.930]in this field.
[00:29:29.420]The Planet data is able to kind of track what you see
[00:29:32.750]quite well, what you see in the institute data.
[00:29:38.200]So, I think this is pretty encouraging.
[00:29:41.870]And when I look at a validation,
[00:29:43.910]this is the field-level validation,
[00:29:45.610]so I'm kind of being averaged over each of these fields
[00:29:48.670]over NE one, Ne two, NE three,
[00:29:50.860]and compared my CESTEM LAI with the measured LAI,
[00:29:55.870]we get a pretty good agreement for this specific data set.
[00:30:00.200]So, this is based on data from 2017.
[00:30:06.150]So, I think one of the next steps is always
[00:30:07.830]to extend this to pigments.
[00:30:10.780]I think there's the potential to also retrieve pigments
[00:30:14.640]like Leaf chlorophyll using this technique,
[00:30:17.570]particularly from Sentinel Two data
[00:30:18.990]and then use the CubeSat to enhance that in space and time.
[00:30:21.570]So, I think that's a critical next step of this work.
[00:30:26.207]So, a bit of a summary.
[00:30:30.085]CESTEM is really a cross-platform calibration approach
[00:30:34.100]that can be used for radiometric normalization,
[00:30:37.040]for phenology reconstruction,
[00:30:39.180]as well as higher level biophysical product generation.
[00:30:42.830]So, it's really a kind of technique
[00:30:44.170]to enhance data into probability
[00:30:48.715]and creating a harmonized analysis-ready data set
[00:30:52.473]that is consistent.
[00:30:55.433]A key feature is that it uses observations
[00:30:57.650]from gold standard satellites as a reference,
[00:31:01.320]so it cannot work without that,
[00:31:03.890]so it's critical to have the conventional systems
[00:31:06.680]to calibrate these CubeSat systems.
[00:31:11.470]A powerful element of CESTEM is that it is quite
[00:31:14.960]insensitive to the noise level of the input data,
[00:31:17.900]so it doesn't really matter if you put in digital counts,
[00:31:20.680]top atmosphere reflectances, or if it's really noisy,
[00:31:25.010]if there's a lot of noise in the original data,
[00:31:27.110]it should be able to correct for that,
[00:31:30.076]and you can extend it to other sensors, like UAVs.
[00:31:33.950]I see that as a particularly interesting follow-on study,
[00:31:36.770]kind of apply this technique in an automative matter
[00:31:40.770]to kind of correct data from collected intra field UAVs
[00:31:43.580]to surface reflectances without the need
[00:31:46.934]to do any ground-based calibration work.
[00:31:50.681]And then, it can also be used to...
[00:31:55.232]It goes beyond the immediate product of spectral data,
[00:31:59.220]like the NDVI, you can also kind of extend it
[00:32:01.720]to produce added-value higher-level products,
[00:32:04.650]like Leaf Area Index and maybe Leaf chlorophyll,
[00:32:07.760]that is maybe more useful for a lot of uses,
[00:32:14.834]particularly as LAI is more directly related
[00:32:16.960]to crop growth and crop condition.
[00:32:20.000]And just to finish off, I think the spatiotemporal insight
[00:32:24.390]that is provided by this type of resource
[00:32:27.140]really represent a game changer
[00:32:30.515]in space point monitoring with the precision Ag
[00:32:34.568]kinda being the obvious application area,
[00:32:37.380]but you can envision particularly other areas
[00:32:41.360]that could have potential use of this type of information.
[00:32:47.440]So, just wanna finish off with just this overview
[00:32:50.050]of the strengths and weaknesses
[00:32:52.790]of LargeSats and SmallSats.
[00:32:55.630]So, I think there's a lot of benefit in both
[00:32:58.690]of these systems.
[00:33:00.444]LargeSats, they are rigorously calibrated,
[00:33:03.682]high radiometric quality, and they operate for a long time,
[00:33:07.630]so there's a lot of continuity in terms,
[00:33:10.494]the continuity if very useful,
[00:33:12.600]but then again, you have also weaknesses.
[00:33:15.060]Typically, there is a trade off between spatial
[00:33:17.000]and temporal resolution, the systems are very expensive,
[00:33:20.970]they take years to produce,
[00:33:22.740]and once you get them up there,
[00:33:24.410]the technology is quite old.
[00:33:27.270]A lot of things happen in five or 10 years,
[00:33:30.248]particularly in these times.
[00:33:31.660]So, the SmallSats, they have superior
[00:33:34.280]spatiotemporal resolution, it's very cheap to produce,
[00:33:38.060]and you can implement new technology very rapidly.
[00:33:41.580]As I mentioned before, Planet can kind of turn around
[00:33:44.460]40 satellites in just one week,
[00:33:46.650]so they can really very quickly implement new technology,
[00:33:50.180]that makes this kind of a scheme very powerful,
[00:33:53.800]but then the calibration accuracy is less,
[00:33:55.880]is not consistent, as good as what you get
[00:33:59.324]from the conventional satellite systems,
[00:34:00.394]and having multiple satellites,
[00:34:03.170]hundreds of satellites to operate,
[00:34:05.060]you're gonna run into cross sensor inconsistencies.
[00:34:09.290]At the moment, the spectral resolution is low.
[00:34:11.351]That might change in the future.
[00:34:13.810]There has been a lot of talk about launching
[00:34:16.640]Sentinel Two-like CubeSats and even hyper-spectral CubeSats,
[00:34:21.230]but then again, you're gonna continue
[00:34:23.340]to have shorter lifespan and the continuity
[00:34:27.130]might be another issue, and the fact that it is
[00:34:29.990]a commercial company brings up some uncertainties.
[00:34:34.497]But I think a key issue here is to use these datas
[00:34:38.850]as in synergy.
[00:34:40.100]I think that's kind of the key message of this talk.
[00:34:43.880]The SmallSats should be used to augment
[00:34:45.820]and complement LargeSat systems.
[00:34:47.410]They shouldn't be used separately.
[00:34:49.120]You need to use these together.
[00:34:51.208]And with that, I can take questions.
[00:34:56.540](audience applauding)
[00:34:58.710]Hi, nice talk.
[00:35:00.550]I had a question about the case of clouds or shadows
[00:35:06.210]where you're creating data,
[00:35:09.200]where there appeared to be no data.
[00:35:11.570]So, I wanted to know just a bit more
[00:35:13.570]about how you do that, 'cause what you're showing
[00:35:15.570]is spatial detail that seemed to be completely lacking
[00:35:21.860]in the original condition,
[00:35:23.530]so I'm missing something in how
[00:35:26.600]Yeah, I didn't really talk about that.
[00:35:28.785]So, it basically uses CubeSat data from scenes
[00:35:36.150]in the past and around the day in question,
[00:35:41.230]so when given the high frequency of the CubeSat data,
[00:35:43.967]you're quite likely to get a scene quite close
[00:35:47.330]to the data.
[00:35:48.163]At the moment, I'm using five scenes
[00:35:49.640]surrounding the date in question,
[00:35:51.800]and then I'm using that in symphony with daily MODIS data.
[00:35:56.892]So, both incorporating the MODIS data into the interpolation
[00:36:02.530]and regression, and then also the surrounding CubeSat data.
[00:36:06.840]It's a bit similar to the Stair Algorithm in some ways
[00:36:10.570]in how it does it, if you're familiar with that,
[00:36:13.899]but it does create a land cover
[00:36:16.050]kind of a cover-specific progression
[00:36:17.849]to kind of do the interpolation in space and time
[00:36:23.750]and then it uses some other techniques to kinda
[00:36:25.830]make it spatially homogenous,
[00:36:27.810]so sometimes you can have unrealistic gradients
[00:36:30.710]when you do the gap-filling and I kinda worked
[00:36:33.900]on refining those techniques to avoid that.
[00:36:37.541]This is something that I did the last two weeks,
[00:36:40.070]so just built very new stuff.
[00:36:41.620]So, probably still need a bit of refining,
[00:36:44.450]but I think it's really useful to having
[00:36:46.070]this kind of daily information
[00:36:48.350]and I need to do some validation of it, of course.
[00:37:00.660]Thank you.
[00:37:02.610]That was a nice talk.
[00:37:05.310]So, you said that the lifespan of a satellite
[00:37:08.040]is from two to three years
[00:37:10.440]and you totally depend on LandSat or Sentinel
[00:37:13.840]or MODIS to calibrate.
[00:37:17.480]What would be the strategy if both
[00:37:21.200]the gold standard satellites fail?
[00:37:23.330]Is it possible to cross-calibrate against
[00:37:25.620]the own CubeSat constellation
[00:37:28.917]given the short lifespan of them?
[00:37:32.525]Is that any plan about that
[00:37:34.010]or does the company consider to cross-calibrate
[00:37:39.200]against UAV data?
[00:37:41.090]Is that a possibility?
[00:37:42.290]Or you're totally dependent on the LargeSat?
[00:37:47.710]Yeah, for the initial calibration,
[00:37:49.070]you will need the LargeSat data,
[00:37:52.510]but then once you have it calibrated,
[00:37:54.210]you can then use the CubesSat data to,
[00:37:56.580]for instance, calibrate the UAV data,
[00:38:00.080]but the first step, yeah, you're gonna need some kind
[00:38:03.454]of a highly calibrated system,
[00:38:05.380]like a conventional satellite, like LandSat and Sentinel Two
[00:38:09.884]and it shouldn't be an issue,
[00:38:10.717]because it's gonna be up there for a long time.
[00:38:13.440]Yeah.
[00:38:15.430]But maybe at some time they don't need it.
[00:38:17.090]I don't know.
[00:38:17.923]Who knows?
[00:38:21.161]Other questions?
[00:38:23.620]Making me work over here.
[00:38:33.150]Hi, great talk.
[00:38:35.090]So, another variable which is very relevant
[00:38:38.350]to agriculture production in parallel to LAI
[00:38:41.300]is fractional PAR.
[00:38:44.730]Do you think this approach can be extended to that?
[00:38:47.970]Especially given that it has a higher demand
[00:38:50.470]for crop modeling and applications such as those?
[00:38:54.928]FA Par? FA Par, yeah.
[00:38:56.949]Yeah, I guess.
[00:39:00.260]I think it's interesting to look into it.
[00:39:01.930]I've also been considering doing ET with this approach.
[00:39:07.930]It's a bit more tricky
[00:39:10.140]because in terms of counting for the change
[00:39:14.120]between your conventional satellite acquisition
[00:39:18.290]and your CubeSat acquisition,
[00:39:20.970]it's easier with spectral data, I think,
[00:39:22.890]but I think it might be feasible,
[00:39:26.330]but we'll need a bit of refinement, I think.
[00:39:32.339]Other questions?
[00:39:33.958]Oh.
[00:39:40.820]Good great talk.
[00:39:42.210]In one of the slides, you showed where you be able
[00:39:46.860]to invert the prospect in the cell
[00:39:50.890]to be able to extract a few vegetation traits,
[00:39:55.735]but maybe I missed some of that talk.
[00:39:59.850]You didn't mention whether you used the ground truth data
[00:40:04.220]to validate or train your data set.
[00:40:06.691]Could you just elaborate that?
[00:40:09.780]Yeah, I didn't explain that a bit.
[00:40:11.427]Yeah.
[00:40:12.860]Sorry.
[00:40:14.710]No, so I use a random floral forest implementation
[00:40:16.870]of Pro-Cell, so I use a synthetic training data sets
[00:40:21.660]that are basically just run the Pro-Cell
[00:40:23.550]in forward mode to establish the training data
[00:40:26.580]and then I use a random forest to kind of build
[00:40:28.930]the model based on that, and then I apply the model,
[00:40:31.600]so I don't use any institute data.
[00:40:34.700]I do fix a few parameters,
[00:40:36.590]like the Leaf mesophyll structure,
[00:40:39.100]I fix that to be appropriate for maize, in this case,
[00:40:45.540]but I keep the Leaf angle and chlorophyll
[00:40:48.750]and Leaf Area Index and most of the other parameters fixed.
[00:40:54.060]Yeah, but it's a random forest kind of implementation,
[00:40:56.780]like a hybrid training.
[00:41:09.720]Looking forward, I mean, this is a really nice
[00:41:11.610]framework for filling in gaps and getting a lot
[00:41:15.050]more information than we could possibly get
[00:41:16.880]from any one sensor.
[00:41:19.420]You mentioned hyper-spectral.
[00:41:21.860]Looking forward to a world of hyper-spectral satellites,
[00:41:27.070]we may not have the CubeSat version of it
[00:41:29.610]or we may have it in some limited format
[00:41:31.838]without the degree of coverage that you have,
[00:41:35.170]but there would be a different issue, of course.
[00:41:36.910]We'd be spectrally rich.
[00:41:39.600]So, do you have any thoughts on a future world
[00:41:44.160]where these kinds of algorithms might take us
[00:41:46.660]with a slightly different configuration
[00:41:48.570]or what objectives me might be able to achieve
[00:41:53.200]within that future world where hyper-spectral
[00:41:57.269]starts to come in, but maybe in a different configuration
[00:42:00.120]than what you're currently using?
[00:42:03.370]Yeah, well I think it's probably not gonna be daily.
[00:42:06.070]I don't think so.
[00:42:07.160]But maybe at some point.
[00:42:09.740]But I think still you need to do kind of the same thing.
[00:42:12.300]You're probably still not gonna need
[00:42:14.339]a conventional hyper-spectral emission
[00:42:16.790]to do the calibration, unless...
[00:42:19.360]So, the CubeSat is gonna be at the same issues
[00:42:21.390]in terms of the absolute calibration accuracy, I think.
[00:42:26.630]So, I think we need to have a space agency,
[00:42:33.240]like a real hyper-spectral emission up there
[00:42:35.380]to kind of provide the necessary reference
[00:42:38.580]for calibrating, and that might happen and it might not,
[00:42:43.610]but it hopefully will.
[00:42:44.780]I think some...
[00:42:46.280]Or maybe the commercial sector needs to take over,
[00:42:48.280]I don't know, providing that kind of good reference
[00:42:51.000]in terms of the hyper-spectral domain.
[00:42:55.120]But I think it's needed 'cause you can only get
[00:42:57.290]so much information out of a broad
[00:43:01.220]multi-spectral bands like this.
[00:43:05.670]But I think there's so many opportunities
[00:43:07.207]in the commercial sector at the moment
[00:43:10.550]in terms of technology and advances
[00:43:12.650]and you might also...
[00:43:14.460]Firmware could also be on the table.
[00:43:17.100]Maybe if they find a business model,
[00:43:19.213]they might put firmware on a CubeSat in the near future.
[00:43:22.360]I think it's technologically feasible,
[00:43:25.530]but if there's an economic benefit of it,
[00:43:27.470]I think that might be the question at the moment,
[00:43:29.850]but I think there's many opportunities.
[00:43:35.229]Speaking of the commercial side of things,
[00:43:37.660]what's the latency of getting the Planet data
[00:43:41.250]running it through your algorithm
[00:43:43.314]and what type of a cost per acre is a producer looking at?
[00:43:47.380]Any ideas on that?
[00:43:50.210]The latency, it's about one day.
[00:43:53.106]I think they make it available within one day,
[00:43:55.040]but as a researcher, you can get it either free
[00:43:59.130]or at a very reduced cost,
[00:44:01.840]but it is a commercial company,
[00:44:06.373]but they don't kind of make the cost available to you.
[00:44:10.303]Okay.
[00:44:11.350]But it's gonna be relatively cheap, I think,
[00:44:16.390]and maybe it's gonna be...
[00:44:17.410]In the future, it might be delivered free in some form,
[00:44:20.600]if the government or some other see the benefit
[00:44:23.910]in providing that, maybe purchasing that data
[00:44:26.207]and providing it free to the public or researchers.
[00:44:31.697]Other questions?
[00:44:40.710]So on the commercial side,
[00:44:45.220]from the perspective of the end-user,
[00:44:48.220]it's a lot of data, considering the time series
[00:44:51.676]and the pixel size, and I imagine that Planet
[00:44:56.520]or whatever company owns the constellation
[00:45:01.261]has some framework to provide this data totally processed,
[00:45:08.370]so the end-user just gonna use it.
[00:45:11.154]It's not gonna develop nothing on top of that.
[00:45:15.500]Is there any strategy to allow scientists
[00:45:21.720]to work on a broader scale, larger scale?
[00:45:29.200]Because what I see in remote sensing lately
[00:45:31.990]is just an increase in the constellation,
[00:45:35.620]the amount of data, but we still struggling
[00:45:37.980]to process all that is out there
[00:45:41.190]and so, with commercial companies taking over
[00:45:45.400]and closing pretty much a black box on processing data,
[00:45:51.540]how the scientific community will rely on this data
[00:45:56.800]to produce science other than just commercial applications?
[00:46:01.460]So, what is the potential of applying to science
[00:46:05.030]and open the box?
[00:46:06.860]Well, I think Planet might be a bit different.
[00:46:09.930]I think they want that to happen.
[00:46:11.950]I think they wanna make the product as available
[00:46:14.490]to as many people as possible
[00:46:17.242]and even talking about making free code available.
[00:46:20.510]I think the idea is to make an analysis-ready product
[00:46:23.493]that can be readily used and maybe having
[00:46:27.505]some add-on software that can be applied for free,
[00:46:30.930]open-source software that can be applied for free
[00:46:33.040]by the user if they need it.
[00:46:35.790]So, I think that's the direction they're moving,
[00:46:38.533]if they can make a business model out of it.
[00:46:40.760]So, it's a commercial company,
[00:46:42.800]so there needs to be some kind of revenue out of it,
[00:46:45.940]but I think that's what they want.
[00:46:51.133]They wanted to make it available
[00:46:52.510]to as many people as possible.
[00:46:57.008]Yeah.
[00:46:58.560]Any last questions for Rasmus?
[00:47:05.040]Alrighty, well let's thank him one more time.
[00:47:06.820]Thank you, Rasmus. (audience applauding)

The screen size you are trying to search captions on is too small!

You can always jump over to MediaHub and check it out there.

Comments

0 Comments

Earth observation by constellations of CubeSats: New opportunities and challenges

Description

Searchable Transcript

Comments icon comment

Related Channels

Comments