Estimation of surface variables at the sub-pixel level for use as input to climate and hydrological models Jean-Pierre Fortin Monique Bernier Danielle DeSève Stéphane Lapointe Institut national de la recherche scientifique, INRS-Eau 2800 rue Einstein, Case postale 7500 Sainte-Foy (Québec) G1V 4C7 November 1996
OBJECTIVES OF THE INVESTIGATION The general objectives of the investigation are: Estimation of surface variables using data from a medium spatial resolution and high frequency remote sensing sensor in orbit. Increased accuracy of spatial positionning for multitemporal analysis of data.
OBJECTIVES OF THE INVESTIGATION (cont.) Derived from those are the following specific objectives: Estimation of the percentage of various land covers within each pixel. Estimation at the sub-pixel level of the spatial distribution of snow cover on the ground and other physical variables of the surface, corresponding to each land cover within the pixel, the albedo for example. As accurate as possible positionning of the images for multitemporal input into an spatially distributed hydrological model using geocoded data.
Labrador Québec Ontario Québec Montréal USA TM image 0 12.5 km Geographical location of Landsat-TM image
OBJECTIVES OF THE INVESTIGATION (cont.) Derived from those are the following specific objectives: Estimation of the percentage of various land covers within each pixel. Estimation at the sub-pixel level of the spatial distribution of snow cover on the ground and other physical variables of the surface, corresponding to each land cover within the pixel, the albedo for example. As accurate as possible positionning of the images for multitemporal input into an spatially distributed hydrological model using geocoded data.
LANDSAT image (25 m) corrected for atmospheric effects Nicolet area, May 9th, 1993 2000 x 3000 pixels (50 km x 75 km) TM1 TM2 TM3 TM4 TM5 SPOT-4 HRVIR image (20 m) simulated 2500 x 3750 pixels Nearest neighbor Resampling method Resampling Arithmetic mean SPOT-4 VGT image (1150 m) simulated 43 x 65 pixels Supervised classification with training sites Pixel selection algorithm Land use map K groups of pixels r ij = mean reflectance of land use class j in spectral band i f= ij proportion of land use class j in pixel i Estimation of reflectances of each land use class Reflectance linear mixture model Least squares algorithm COMPARISON r ^ ij(k) = estimated reflectance of land use class j in spectral band i using the kth group of pixels r ^ ij = estimated reflectance of land use class j in spectral band i Mean on the K groups of pixels ET ij = Standard deviation of estimated reflectances in the K groups of pixels for land use class j in spectral band i
Relative spectral response INRS-Eau 1 0.8 0.5 0.4 0.2 400 600 800 1000 1200 1400 1600 1800 Wavelenght (nm) TM HRVIR VGT Relative spectral responses of TM, HRVIR and VGT bands
HRVIR1 HRVIR2 VGT1 VGT2 Simulated HRVIR and VGT bands 1 and 2 a) Petite rivière du Loup river, b) Yamachiche river, c) St-Maurice river, d) St-François river, e) Nicolet river, f) Bécancou r river, g) St-Lawrence river, h) lake St-Pierre, i) Trois-Rivières city, j) sandy soil, k) clayey soil
HRVIR3 HRVIR4 VGT3 VGT4 Simulated HRVIR and VGT bands 3 and 4 a) Petite rivière du Loup river, b) Yamachiche river, c) St-Maurice river, d) St-François river, e) Nicolet river, f) Bécancou r river, g) St-Lawrence river, h) lake St-Pierre, i) Trois-Rivières city, j) sandy soil, k) clayey soil
Simulated false color composite of HRVIR bands (HRVIR3, HRVIR4 and HRVIR2) Simulated false color composite of HRVIR and VGT bands a) Petite rivière du Loup river, b) Yamachiche river, c) St-Maurice river, d) St-François river, e) Nicolet river, f) Bécancou r river g) St-Lawrence river, h) lake St-Pierre, i) Trois-Rivières city, j) sandy soil, k) clayey soil Simulated false color composite of VGT bands (VGT3, VGT4 and VGT2)
Classified HRVIR image Water Dry bare soils Wet bare soils Wetlands Peat bogs with spruces Peat bogs with alders Pastures, alfalfa and hay Herbacious follow lands and bushes Forests Urban areas and roads
30,0 25,0 20,0 Relative error (%) for each VGT band, as a function of the number of pixel groups. Mean on land-use classes and ten trials. Relative error in % (mean on classes) 15,0 10,0 VGT1 VGT2 VGT3 VGT4 5,0 Mean 0,0 1 2 3 4 5 6 7 8 9 10 Number of groups of pixels
TABLE 6 Maximum relative error for a shift up to 300 m and relative error corresponding to perfect registration Landuse classes Water Dry bare soils Wet bare soils Pastures, alfalfa and hay Forested areas Wet lands Peat bogs with spruces Peat bogs with alders Herbaceous fallow lands and bushes Maximal relative error for shift up to 300 m 11,1% 9,6% 8,2% 4,4% 21,1% 36,5% 5,6% 22,1% 11,4% Urban areas and roads 15,7% Relative error no shift 8,1% 9,6% 4,8% 4,4% 21,1% 6,5% 5,2% 12,6% 9,7% 15,5% Difference 3,0% 0,0% 3,4% 0,0% 0,0% 30,1% 0,5% 9,5% 1,7% 0,2%
300 0.75 0.75 300 1 1.1 200 0.7 200 North-south shift (m) 100 0-100 0.6 0.65 0.6 0.75 North-south shift (m) 100 0-100 0.7 0.8 1.1-200 0.6 0.75 0.8 0.6-300 -300-200 -100 0 100 200 300 VGT-1 West-east shift (m) -200 1.3 0.9 1 1.4 1.2-300 -300-200 -100 0 100 200 300 VGT-2 West-east shift (m) North-south shift (m) 300 200 100 0-100 -200 1.8 1.6 1.2 1.2 2 1.4 1 1.4 1.8 2.2 2.4-300 -300-200 -100 0 100 200 300 VGT-3 West-east shift (m) 1.6 North-south shift (m) 300 200 100 0-100 -200 2 1.5 1 3 2.5-300 -300-200 -100 0 100 200 300 West-east shift (m) Standard deviation on reflectance estimations as a function of registration accuracy, using four groups of pixels. Mean on land-use classes and ten trials VGT-4 2 2.5 1.5
CONCLUSIONS HRVIR and VGT images have been simulated from a TM image. The main features of the HRVIR image are still recognizable on the VGT image. For the estimation of class reflectances, the optimal number of groups is 4. The increase in relative error for estimation of the reflectances of landuse classes at maximal x-y shift relative to perfect registration in generally less than 3%. It seems possible to locate the individual spectral bands within 100 to 150m from their true location. Verifications will have to be made on other images to confirm these results.