Processing and Analysis of ALOS/Palsar Imagery Yrjö Rauste, Anne Lönnqvist, and Heikki Ahola Kaukokartoituspäivät 6.11.2006 NewSAR Project The newest generation of space borne SAR sensors have polarimetric capacity: Japanese ALOS, launched in January 2006 German TerraSAR X, launched in June 2007 Canadian Radarsat 2, to be launched in December 2007 Objectives: to develop new forms of utilising polarimetric and nonpolarimetric SAR data from the new SAR sensors Participants: VTT TKK (laboratory of Space Technology, Laboratory of Computer and Information Science) Finnish Geodetic Institute 2 1
Geometry Information in Rectification of Palsar Data The products of JAXA (in the ESA sponsored Aden AO projects) do not give convenient geo location data All products include state vectors (platform position and velocity vectors for 40 points with 60 second interval) Level 1.5 products have geo location polynomials in MapProjection record, and FacilityRecord11; level 1.5 products lack the time code in connection of each image line; polarimetric level 1.5 products only includes 4 amplitudes application of most polarimetric techniques impossible Level 1.1 products lack MapProjection record, and geolocation data for each line (data in SignalData format, not ProcessedData, where geo location data could be included) The only way to get image geo located is by using state vectors and time codes 3 Computation of Palsar Geo Locations Starting with state vectors (positionvelocity data) Interpolate state vector to the exact time of image line Intersect: Earth, range sphere, Doppler cone 4 2
Equations by Curlander (1982): Palsar Geo coding Algorithm Classical least squares formulation: Solution from: Curlander, J. 1982. Location of spaceborne SAR imagery, IEEE Transactions on Geoscience and Remote Sensing, Vol. GE 20, No. 3, July 1982, p. 359 364. 5 Ortho Rectification Software Framework for Experimentation (Simplified) yrpalsarpol2s2 yrpalsarpolcal2s2 JAXA pol. prod. L1.1 S or Dbl pol. L1.1 yrpalsarsinglepolextract S2 S2 files S2 S2 files yrs2tocohavg T3 T3 T3 files T3 files T3 T3 T3 files T3 files T3 files yrpalsargrid.gdc file DEM S2 file yrdetectfloat2i16 16 bit file yrortho5_grd Rectified T3 T3 T3 T3 T3 files files T3 files T3 files T3 T3 files yrt3toampl yrortho4_grd Rectified 16 bit file 6 3
Radiometric Correction in Stokes Domain Gamma correction: P corrected = (P in P noise )A n /A projected + P noise A n = nominal projected pixel area, A projected = actual, DEMcomputed projected pixel area Applied to Stokes matrix data F (16 elements, 9 independent elements): F corrected = (F in F noise )A n /A projected + F noise A projected computed by a "pixel counting" algorithm Experimented in an older program environment (yrortho4_grd, modified to yrortho5_grd to accommodate Stokes data instead of a single amplitude image) 7 Pixel Counting Algorithm in Radiometric Correction The centre of output pixel centred with a DEM element Going towards the sub satellite point in DEM, find the first DEM element that is outside the slant range resolution Find the intersection between DEM surface and range resolution Repeat going away from the sub satellite point Compute the vertical dimension x of the projected pixel area 8 4
Stokes Matrix Definition Zebker, H. and Lou, Y. 1990. Phase calibration of imaging radar polarimeter Stokes matrices, IEEE Transactions on Geoscience and Remote Sensing, Vol. 28, No. 2, p. 246 252. 9 Palsar Ortho rectification/noise component Assumptions: SHH, SVV, and SHV amplitude = 29 db (from specs) Phase in all components is random Test with simulation (left figure): 10000000000 samples with random phase summed in Stokes domain Form of the noise Stokes matrix (right figure): 1 0 0 0 0 ½ 0 0 0 ½ Eigenvalue decomposition of coherency matrix (e.g. Cloude, S. 1996. A review of target decomposition theorems in radar polarimetry, IEEE Transactions on Geoscience and Remote Sensing, Vol. 34, No. 2, March 1996, p. 10498 518.) 5
Sample Polarimetric Rectification Kuortane site, 14.5.2007, original scene left Red = HH VV, green = HV, blue = HH+VV ALOS/Palsar data JAXA, METI 2007 Ortho rectified sub scene in 3 D view: 11 ALOS/Palsar Data and Forest Inventory Data in Heinavesi Site Dual pol scene 12.6.2007 shown (R=B=HH, G=HV) in 3 D perspective with ground inventory stands (from UPM) overlaid A subset of variables in ground data: C Name 1 KUVIO 2 KEHITYSLUOKKA 3 VALTAPITUUS 4 POHJAPINTA_ALA_YHT 5 RUNKOLUKU_YHT 6 KESKIPITUUS_YHT 7 KESKILAPIMITTA_YHT 8 KUUSI_PROS 9 MANTY_PROS 10 LEHTI_PROS 11 KESKI_IKA 12 TILAVUUSM3_HA_YHT 12 6
Dual Pol/Heinavesi 12.6.2007: Stem Volume Correlation/HH Whole dataset: r = 0.82 0 150 m 3 /ha: r = 0.82 Saturation around 150 m 3 /ha One obvious clear cut stand screened out 13 Dual Pol 12.6.2007/Heinavesi: Stem Volume Correlation/HV Whole dataset: r = 0.79 0 150 m 3 /ha: r = 0.93 Saturation around 150 m 3 /ha One obvious clear cut stand (harvested March 2007, ground March 2007) 14 7
Forest with "saturated" return Biomass amplitude relation in L band saturates around 150 m 3 /ha Stand 144, 155 m 3 /ha, latest thinning in November 2006 Palsar dual pol scene acquired on 12 6 2007 15 RMSE = 50 m3/ha Stem volume (m 3 /ha) 140 120 100 Proba estimate 80 60 40 20 0 0 20 40 60 80 100 120 140 160 180 200 220 240 260 280 300 Ground data Proba chain estimate for stem volume averaged over the plots 2 ha, Kuortane study site, ALOS/Palsar dual pol data 2007 06 27 16 8
RMSE = 6 cm Stem diameter (cm) 20 15 Proba estim ate 10 5 0 0 5 10 15 20 25 30 Ground data Proba chain estimate for stem diameter averaged over the plots 2 ha, Kuortane study site, ALOS/Palsar dual pol data 2007 06 27 17 Classification Method Comparison in Kuortane Four land cover classification methods: AutoChange, ERMapper maximum likelihood, unsupervised and supervised classification methods (based on Wishart distribution) implemented in the PolSARpro software package Corine land cover data (SYKE) as reference data Supervised Wishart classification gave marginally better results than other methods A sample confusion matrix: Corine ground data User's Water Open Forest Total accuracy Palsar Water 2.8 0.0 0.0 2.9 96.7 class Open 0.6 16.5 3.6 20.7 79.9 Forest 0.2 8.6 67.6 76.4 88.5 Total 3.6 25.2 71.3 100.0 Procucer's 77.4 65.7 94.9 87.0 accuracy 18 9
Example of classification result May: PolSARpro supervised 19 7000 AutoChange with 100 clusters HH+VV 6000 5000 33 37 marsh built up areas 4000 67 31 23 61 17 44 3000 42 62 15 35 60 92 27 38 69 48 50 83 14 25 36 45 63 66 8 1 2 3 6 7 20 29 40 43 21 34 22 39 51 525354 5759 55 70 71 13 24 49 72 74 75 76 88 89 77 8084 8586 30 9 1619 28 46 58 47 646568 78 79 90 93 94 95 98 99 2000 81 87 96 97 100 4 10 5 11 12 26 41 56 73 82 91 1000 18 32 sparse dense w ater field forest forest 0 200 400 600 800 1000 1200 HV 20 10
Conclusions Polarimetric ortho rectification facilitates the use of polarimetric SAR data and comparison with ground data Radiometric effects of topography are strongly reduced in radiometric correction Radiometric correction does not introduce artefacts to polarimetric signatures Cross polarised (HV) L band SAR data has higher correlation with forest biomass than HH polarised, saturation still around 150 m 3 /ha Fully polarimetric supervised Wishart classification produced the best performance in land cover classification 21 Acknowledgements JAXA and ESA for providing ALOS/Palsar data in the ESA/ADEN project ALO.AO 1364 UPM for providing forest inventory ground data of the Heinavesi study site and assistance in using the data Etelä Pohjanmaan Metsäkeskus for providing forest inventory ground data of the Kuortane study site SYKE for providing CORINE land cover data TEKES for funding project NewSAR 22 11