BRDF modelling of MODIS land surface data
This post introduces simple BRDF modelling (i.e. how can we account for the variation in surface reflectance due to acquisition geometry?) using some MODIS data. You can find the whole post in wakari. If you accept to not escape the HTML & JS, you will find a link to the data file there too.
A particularly simple way of dealing with BRDF effects is by using linear kernel models. These models are reduced physics versions of more complex radiative tranfer models with some empirical considerations to further simplify things. Since we are only interested in the shape of the BRDF, we can lump the parameters into some constants of limited physical explanatory value, and do a few simplifications so that we can write the the observed reflectance at band λi as:
where we have Nkern kernels, typically 3, and the acquisition geometry is given by (Ω, Omega’) . The three kernels that are typically used are those that have been popularised by the MODIS albedo product (MCD43):
- An isotropic kernel, (Kiso = 1 ), accounting for isotropic BRDF effects
- A volumetric scattering kernel (Kvol ), a simplified version of volume scattering as seen in Ross (1981)
- A geometric scattering that aims to mimick surface scattering and shadowing.
While different families of these kernels are used, we use the Ross-ThickLiSparse-Reciprocal kernels (these have been evaluated and found to be quite good).
The problem stated above cannot be solved on a daily basis (or whenever an observation becomes available): there are 3 parameters per band, and only one observation, so the system is underdetermined. If we assume that the parameters can be related to land surface parameters, and we assume that change is slow, a way around this is temporal compositing, where all the observations within a given period are inverted together. It is important to take uncertainty in the observations into account, as we don't measure ρλi , but rather ρ̂λi = ρλi + ϵλi , where ϵλi is additive noise, typically modelled as zero mean and with a known variance (see Roy et al for a table with typical values for MODIS).
Ill-posedness
The problem of inverting RT models (be they linear or not) is typically ill-posed. By contrast, a well-posed problem (in the Hadamard sense) is one where:
- A solution exists
- The solution is unique
- The solution's behavior hardly changes when there's a slight change in the initial condition
In this case, due to the uncertainty introduced by the noise (and limited angular sampling), point 2 is not typically met, resulting in many configurations of kernel weights that provide an adequate representation of the observations. Compositing is a good way of increasing the angular sampling and reducing the noise, but it can only be taken that far: too long a compositing window, and the assumption of static land surface properties weakens.
Reading the data
Let's look at some data. The data is stored in a text file, and belongs to a pixel somewhere in Angola. We only have band 2 data here. The first ten lines of the file show the format. The first line is a header, telling us that there are 92 observations, with 7 bands, as well as their centre wavelengths. After the first line, we have individual records with the following format:
- day of year (DoY)
- QA flag (1=use, 2=don't use)
- vza, the view zenith angle (in degrees)
- vaa, the view azimuth angle (in degrees)
- sza, the solar/illumiation zenith angle (in degrees)
- saa, the solar/illumiation azimuth angle (in degrees)
- the reflectance
Typically, we will use sza, vza and raa. raa``=``vaa - saa.
First, we need to get the data. The next code snippet will provide you with a link to the datafile that you can download (hopefully) from wakari
!head -n 10 data.r2023.c87.dat
BRDF 92 7 648 858 470 555 1240 1640 2130 181 1 65.419998 -84.470001 44.130001 20.090000 0.114600 0.243200 0.052800 0.087100 0.328300 0.302300 0.213400 182 1 23.410000 98.290001 50.220001 35.310001 0.113900 0.218100 0.051100 0.084500 0.315900 0.339000 0.205500 184 1 44.049999 100.730003 51.910000 38.360001 0.142900 0.269100 0.063400 0.107500 0.359300 0.375000 0.246500 185 1 40.400002 -82.199997 46.310001 27.700001 0.107000 0.212100 0.051600 0.079900 0.310400 0.310400 0.214100 186 1 57.720001 101.300003 53.700001 41.259998 0.152700 0.291400 0.064600 0.114800 0.395000 0.382700 0.258700 187 1 17.770000 -81.010002 47.630001 31.260000 0.111000 0.217700 0.050500 0.081600 0.307600 0.318500 0.205500 188 0 0.000000 0.000000 0.000000 0.000000 0.000000 0.000000 0.000000 0.000000 0.000000 0.000000 0.000000 189 1 10.470000 96.879997 49.090000 34.650002 0.117500 0.225000 0.054200 0.087700 0.334600 0.334300 0.231400 190 1 60.889999 -83.790001 44.070000 22.910000 0.100200 0.212100 0.045900 0.074100 0.296100 0.302100 0.188200
Let's read the data in and do some plots!
data = np.loadtxt ( "data.r2023.c87.dat", skiprows=1) # Read data in, but skip first line
doys = data[ :, 0]
qa = data [ :, 1 ]
vza = data [ :, 2 ]
sza = data [ :, 4 ]
raa = data [ :, 3 ] - data [ :, 5 ]
rho = data[:,7]
plt.plot ( doys[qa==1], rho[qa==1], 'o-')
plt.xlabel("DoY")
plt.ylabel(r'$\hat{\rho}^{NIR}$')
The high frequency variation is quite high, and over a fairly short period, where we only expect a smooth variability of reflectance, we interpret it as BRDF effects. There's a big dip on DoY 228, due to a fire (there's a hotspot there too), and then we see recovery due to char & ash dissipation processes. Let's look at the angular signal up to DoY 228:
passer = np.logical_and ( qa==1, doys < 228 ) # Select obs before the fire
plt.plot ( vza[passer], rho[passer], 'o' )
plt.xlabel("VZA [deg]")
plt.ylabel(r'$\hat{\rho}^{NIR}$')
We see a large scatter in reflectance for larger values of VZA. Now, let's try to model this. We need to calculate the kernel weights for the Ross and Li-Thick kernels for each individual observation. We have a Python class that does that, kernels.py, which can be imported and used for this:
from kernels import Kernels # Import ther kernels
# Generate the kernels, only bother for obs where the QA is OK
K_obs = Kernels( vza[qa==1], sza[qa==1], raa[qa==1], \
LiType='Sparse', doIntegrals=False, \
normalise=1, RecipFlag=True, RossHS=False, MODISSPARSE=True, \
RossType='Thick' )
kern = np.ones (( np.sum(qa==1), 3 )) # Store the kernels in an array
kern[ :, 1 ] = K_obs.Ross
kern[ :, 2 ] = K_obs.Li
Now we can just solve our system. First with brute force: let's take 8 days worth of data, and invert the kernel weights to see what we get. This is a straightforward least linear squares problem, and we can use lstsq to solve it.
passer = np.logical_and ( qa==1, np.logical_and( doys > 200, doys <=209))
obs = rho[passer]
K = kern[ passer, :]
(f, rmse, rank, svals ) = np.linalg.lstsq( K, obs )
print "%-20s\t%20s" % ( "Kernel", "Value")
for i, k in enumerate( ["Isotropic", "Ross", "LiThick"] ):
print "%-20s\t%20f" % ( k, f[i] )
Kernel Value Isotropic 0.242057 Ross -0.065568 LiThick 0.075140
So we have a solution. How good is it? Let's try to forward model the observations:
fwd = K.dot(f.T)
print "R2: %g" % np.corrcoef ( obs, fwd)[1,0]
plt.plot ( obs, fwd, 'sk')
plt.plot ( [0.2, 0.25], [0.2, 0.25], 'k--', lw=0.5)
plt.ylim(0.2, 0.25)
plt.xlim(0.2, 0.25)
R2: 0.445343
So, not a great fit! Maybe if we use a larger window? Let's create a function to the the above conveniently:
def invert_composite ( qa, doys, rho, kern, doy_min, doy_max ):
"""A function to invert reflectance assuming the kernels in `kern`,
`qa` holds the QA data, `doys` the dates, and `rho` the actual
observations. Will composite all observations between `doy_min` and
`doy_max`."""
passer = np.logical_and ( qa==1, np.logical_and( \
doys > doy_min, doys <= doy_max))
if np.sum(passer) < 3:
raise np.linalg.LinAlgError
obs = rho[passer]
K = kern[ passer, :]
(f, rmse, rank, svals ) = np.linalg.lstsq( K, obs )
print "N. Obs: %d" % np.sum ( passer )
print "%-20s\t%20s" % ( "Kernel", "Value")
for i, k in enumerate( ["Isotropic", "Ross", "LiThick"] ):
print "%-20s\t%20f" % ( k, f[i] )
fwd = K.dot(f.T)
print "R2: %g" % np.corrcoef ( obs, fwd)[1,0]
plt.plot ( obs, fwd, 'sk')
plt.plot ( [0.18, 0.3], [0.18, 0.3], 'k--', lw=0.5)
plt.ylim(0.18, 0.3)
plt.xlim(0.18, 0.3)
return obs, fwd, passer, f, rmse
retval = invert_composite ( qa, doys, rho, kern, 200, 210 )
plt.title("Compositing between DoYs 200-210")
N. Obs: 9 Kernel Value Isotropic 0.235238 Ross -0.033419 LiThick 0.054909 R2: 0.386086
retval = invert_composite ( qa, doys, rho, kern, 180, 227 )
plt.title("Compositing between DoYs 200-227")
N. Obs: 41 Kernel Value Isotropic 0.231309 Ross 0.009338 LiThick 0.000748 R2: 0.05195
Clearly, this window is too big: changes are happening. Also note the difference of adding just 1 observation to the 10 day period: the parameters change perceivably, the sign of ill-posedness. Let's see how the compositing period affects the estimate: we'll try to invert the pre-fire observations using a 16 and an 8 day compositing period:
retvals = []
for period in xrange ( (227-180)/16 + 1 ):
doy_min = 180 + 16*period
doy_max = 180 + 16*(period+1)
if doy_max > 227:
doy_max = 227
plt.subplot ( 1, 3, period + 1 )
print "=============================================================="
print "DoY Min: %d" % doy_min
print "DoY Max: %d" % doy_max
try:
retvals.append ( invert_composite ( qa, doys, rho, kern, \
doy_min, doy_max) )
except np.linalg.LinAlgError:
print "Not enough obs!"
plt.title("Compositing %d-%d" % \
( doy_min, doy_max) )
============================================================== DoY Min: 180 DoY Max: 196 N. Obs: 14 Kernel Value Isotropic 0.238851 Ross -0.006477 LiThick 0.009853 R2: 0.100897 ============================================================== DoY Min: 196 DoY Max: 212 N. Obs: 15 Kernel Value Isotropic 0.216249 Ross 0.066691 LiThick -0.059062 R2: 0.521575 ============================================================== DoY Min: 212 DoY Max: 227 N. Obs: 12 Kernel Value Isotropic 0.243611 Ross -0.062851 LiThick 0.063866 R2: 0.646983
retvals = []
for period in xrange ( (227-180)/8 + 1 ):
doy_min = 180 + 8*period
doy_max = 180 + 8*(period+1)
if doy_max > 227:
doy_max = 227
plt.subplot ( 2, 3, period + 1 )
print "=============================================================="
print "DoY Min: %d" % doy_min
print "DoY Max: %d" % doy_max
try:
retvals.append ( invert_composite ( qa, doys, rho, kern, \
doy_min, doy_max) )
except np.linalg.LinAlgError:
print "Not enough obs!"
plt.title("Compositing %d-%d" % \
( doy_min, doy_max) )
============================================================== DoY Min: 180 DoY Max: 188 N. Obs: 6 Kernel Value Isotropic 0.215737 Ross 0.217461 LiThick 0.004699 R2: 0.960286 ============================================================== DoY Min: 188 DoY Max: 196 N. Obs: 8 Kernel Value Isotropic 0.256502 Ross -0.148200 LiThick 0.022454 R2: 0.641367 ============================================================== DoY Min: 196 DoY Max: 204 N. Obs: 7 Kernel Value Isotropic 0.222352 Ross -0.073762 LiThick -0.074442 R2: 0.825704 ============================================================== DoY Min: 204 DoY Max: 212 N. Obs: 8 Kernel Value Isotropic 0.211300 Ross 0.154030 LiThick -0.051295 R2: 0.833228 ============================================================== DoY Min: 212 DoY Max: 220 N. Obs: 7 Kernel Value Isotropic 0.241598 Ross 0.021193 LiThick 0.066551 R2: 0.838989 ============================================================== DoY Min: 220 DoY Max: 227 N. Obs: 5 Kernel Value Isotropic 0.247820 Ross -0.146072 LiThick 0.076207 R2: 0.74752
We see that the kernel weights change substantially from one period to the next. The 16 day window seems quite poor, whereas the 8 day one is remarkably better. This means that fast changes are occurring in the ground, and the 8 day window is more appropriate.We can see how this works in time
for period in xrange ( (227-180)/8 + 1 ):
plt.plot( doys[retvals[period][2]], retvals[period][0], 'sk')
plt.plot( doys[retvals[period][2]], retvals[period][1], '-', lw=1.8)
plt.ylim(0.18, 0.3)
Stuff missing
- Uncertainty
Refs
Schaaf, C. B., F. Gao, A. H. Strahler, W. Lucht, X. Li, T. Tsang, N. C. Strugnell, X. Zhang, Y. Jin, J.-P. Muller, P. Lewis, M. Barnsley, P. Hobson, M. Disney, G. Roberts, M. Dunderdale, C. Doll, R. d Entremont, B. Hu, S. Liang, and J. L. Privette, First Operational BRDF, Albedo and Nadir Reflectance Products from MODIS, Remote Sens. Environ., 83, 135-148, 2002.
Lucht, W., C.B. Schaaf, and A.H. Strahler. An Algorithm for the retrieval of albedo from space using semiempirical BRDF models, IEEE Trans. Geosci. Remote Sens., 38, 977-998, 2000.
Wanner, W., A.H. Strahler, B. Hu, P. Lewis, J.-P Muller, X. Li, C. Schaaf, and M.J. Barnsley, Global retrieval of bidirectional reflectance and albedo over land from EOS MODIS and MISR data: Theory and algorithm, J. Geophys. Res., 102, 17143-17161, 1997.