Compact hyperspectral imaging system (COSI) for RPAS system overview and first performance evaluation results A. SIMA, P. Baeck, D. Nuyts, S. Delalieux, S. Livens, J. Blommaert, B. Delauré - Flemish Institute for Technological Research VITO NV, Belgium M. BOONEN pcfruit vzv, Belgium
COMPACT HYPERSPECTRAL IMAGING SYSTEM COSI camera specification @40m [@150] Image size 2048 x 1088 Swath Hypercube ground sampling distance Spectral range Spectral resolution Data storage Mass Dimensions 40 m [150m] 4 cm [15cm] 600-900 nm 470-900 nm (in prep.) 5-10 nm (72 bands) On-board 500 g 6cm x 7cm x 12cm
LINEARLY VARIABLE FILTERS STEP LINE FILTER» Line based interference filter» Filter with varying thickness deposited directly on the detector» Every image row (8 rows) captures different spectral band and cover different spatial location» Scanning motion required 8 pix = 1 spectral band ISPRS Congress 2016 3
RAW IMAGE CAPTURED BY THE COSI CAMERA 600nm 700nm 800nm 900nm ISPRS Congress 2016 4 Data acquired in collaboration with pcfruit, St. Truiden (B)
DATA PROCESSING» Dedicated data processing chain developed in-house» Photogrammetry and computer vision raw image spectral image» GNSS data or GCPs not required but increase quality of data scaling and georeferencing» Radiometric and spectral corrections required to derive reflectance values EGU 2016 5» Need of calibrated spectral reference targets index map Action information map
RECONSTRUCTION OF SPECTRAL BANDS ISPRS Congress 2016 6
HIGH SPATIAL RESOLUTION HYPERSPECTRAL DATA 2 cm GSD 7 1 m Data acquired in collaboration with pcfruit, St. Truiden (B)
SPECTRAL MEASUREMENTS 2m First results on spectral COSI camera performance: strawberries Reference targets grass ISPRS Congress 2016 8 dashboard
DIGITAL SURFACE MODEL 5 m ISPRS Congress 2016 9 Data acquired in collaboration with pcfruit, St. Truiden (B)
CHLOROPHYL RELATED VEGETATION INDICES 10
CHLOROPHYL RELATED VEGETATION INDEX (ZONES) 11
Data acquisition campaigns Strawberry field Winter wheat Natural grassland, Pear orchard
EXPERIMENTAL STRAWBERRY FIELD NEAR SINT TRUIDEN, BELGIUM» Area of interest = 60m x 80m» Flight altitude: 30m - rotary wing» 14 400 images with 1.4cm GSD» 9 flight lines, 80% side lap» GSD of the hypercube = 2cm» 9 ground targets not used during data processing ISPRS Congress 2016 13 Data acquired in collaboration with pcfruit, St. Truiden (B)
BAND CO-REGISTRATION QUALITY ASSESSMENT VISUAL VERIFICATION 2m 736.8nm 604.0nm 736.8nm 604.0nm 736.8nm 860.4nm 736.8nm 860.4nm 14
ESTIMATION OF BAND CO-REGISTRATION QUALITY USING SIFT» SIFT with optimized parameters» > 124 000 points matched (ref band vs. 71 bands)» Reference band 31- intermediate vegetation reflectance values between the visible and the infrared part» 0 points matched in ALL bands!!! Band 33 745.7nm Band 70 888.5nm ISPRS Congress 2016 15
ESTIMATION OF BAND CO-REGISTRATION QUALITY USING SIFT» Equal samples ensured (299)» Average RMS error between all the bands and the reference band 31equal to 0.45pix (min = 0.36pix, max = 0.55pix» Result comparable or better than results reported in Tommaselli et al., 2015 for the Rikola camera:» RMSE of 0.48 pix to 0.90 pix for camera moving at 0.16 m/s over a scene with 1 m of depth variation Tommaselli, A.M.G., Oliveira, R.A., Nagai, L.Y., Imai, N.N., Gabriela, T., Honkavaara, E., Hakala, T., 2015. Assessment of bands coregistration of a light-weight spectral frame camera for UAV, in: Proceedings of the ISPRS Geospatial Week 2015. pp. 4.
TEMPLATE MATCHING WITH NORMALIZED CROSS-CORRELATION» Template (band 31: 736.8nm) = 50pix x 50 pix» Search window = 100pix x 100pix GT1 GT2» Matching failed for templates covering vegetation» Matching successful for templates covering wood shavings GT3 GT4 GT5 GT6 2m 17 GT7 GT8 GT9
TEMPLATE MATCHING WITH NORMALIZED CROSS-CORRELATION» Average RMS error between all the bands and the reference band 31equal to 0.25pix (min = 0.11pix, max = 0.35pix» Result comparable or better than results reported in Tommaselli et al., 2015 for the Rikola camera 18
TEMPLATE MATCHING WITH NORMALIZED CROSS-CORRELATION GSD = 2cm» Results for flat objects > optimistic» vectors of extreme 2D residuals between bands can get close to 1pix» Thus spectral measurements might be incorrect!» Solution: choose larger ground sampling distance for the final product, e.g. 2 x larger ISPRS Congress 2016 19
ORIGINAL VS REPROCESSED DATA GSD = 2cm GSD = 4cm GT1 GT2 GT3 GT4 GT5 vectors of extreme 2D residuals between bands < 0.5pix GT6 2m thus spectral measurement possible 1 pix residual 20 GT7 GT8 GT9
CONCLUSIONS» Compact hyperspectral imaging system has been prototyped, suited for small RPAS, providing very high spatial resolution data» Data processing modules developed to provide wide variety of image products:» Orthorectified hypercube» Digital surface model» Vegetation index maps» Action maps COSI, GSD = 2cm» Specific data capturing design requires in-depth product analysis and validation Nex-6, 16mm lens, GSD = 0.8cm ISPRS Congress 2016 21
CONCLUSIONS 2/2» Interband co-registration error assessment is not a trivial task due to different appearance of objects in different wavelength» Objects appearing similar in all the spectral camera range are recommended as reference targets» Ground sampling distance of the final hyperspectral product should be chosen carefully to allow correct spectral measurement» In the presented dataset GSD=4cm is a good choice, where vectors of extreme 2D residuals between bands stay below 0.5pix Nex-6, 16mm lens, GSD = 0.8cm ISPRS Congress 2016 22
COMMERCIALIZATION PROCESS OF THE COSI SYSTEM IS ONGOING 23