Haze Detection and Removal in Sentinel 3 OLCI Level 1B Imagery Using a New Multispectral Data Dehazing Method

Transcription

1 Haze Detection and Removal in Sentinel 3 OLCI Level 1B Imagery Using a New Multispectral Data Dehazing Method Xinxin Busch Li, Stephan Recher, Peter Scheidgen July 27 th, 2018

2 Outline Introduction» Why do we need to remove HAZE from Sentinel 3 OLCI data? Method» How can we remove HAZE from Sentinel 3 OLCI Level 1B data? Results» How is hazy Sentinel 3 OLCI image restored step by step? Evaluation» How can we evaluate the results of dehazing? Conclusion» What have we gained from dehazing processing? 2

3 Introduction Satellite images from NASA HAZINESS is originated from fractions of water vapor, ice, fog, sand, dust, smoke, or other small particles in the atmosphere. Haze can affect human s visibility and health. Haze degrades optical remote sensing images by its scattering. western Africa, dust haze eastern USA, morning haze Singapore, affected by severe smoke haze due to forest fires in the region periodically ( Reuters, 2013) Italy, air pollution haze China, burning and biomass haze 3

4 Introduction Haziness on optical image: Sentinel 3 OLCI Level 1B Time: 08/12/2016, UTC 7:19 Time: 04/12/2016, UTC 7:23 S3 OLCI RGB image (R:Oa10; G:Oa6; B:Oa3) Haze free area: land/ocean surface Haze area: semi transparent white overlay + land/ocean surface possible to restore land/ocean surface Cloud area : non-transparent white overlay not possible to restore land/ocean surface 4

5 Introduction Haziness on optical image: Sentinel 3 OLCI Level 1B Time: 08/12/2016, UTC 7:19 Time: 04/12/2016, UTC 7:23 geolocated possible view after restoration 5

6 Method A new multispectral data dehazing method (A. Makarau et al., 2014) Hazy image R TOA = R TOA0 + R haze Haze detection R haze = f(htm image, HTM bandi ) Haze removal R TOA0 = R TOA R haze Aerosol compensation R TOA = R TOA + R TOA0 R TOA R TOA, Top of Atmosphere (TOA) radiance of hazy image; R TOA0, TOA radiance of restored hazy image; R haze, radiation contribution from haze; HTM image, Haze Thickness Map of the hazy image, and HTM bandi, HTM per band; R TOA0 averaged TOA radiation of haze-free pixels in restored hazy image; R TOA averaged TOA radiation of haze-free pixels in hazy image R TOA dehazed image with aerosol compensation 6

7 Method A new multispectral data dehazing method (A. Makarau et al., 2014) Hazy image R TOA = R TOA0 + R haze Haze detection R haze = f(htm image, HTM bandi ) Haze removal R TOA0 = R TOA R haze Aerosol compensation R TOA = R TOA + R TOA0 R TOA R TOA, TOA radiance of hazy image; R TOA0, TOA radiance of restored hazy image; R haze, radiation contribution from haze; HTM image, Haze Thickness Map of the hazy image, and HTM bandi, HTM per band; R TOA0 averaged TOA radiation of haze-free pixels in restored hazy image; R TOA averaged TOA radiation of haze-free pixels in hazy image R TOA dehazed image with aerosol compensation 7

8 Method Haze detection» Haze Optimized Transform (HOT) (Y. Zhang et al., 2003) limited by water bodies and man-made features Haze detection: one blue band and one red band (Nicholas Pringle et al., 2015): HOT test =Band 490nm 0.5 Band 560nm 0.08>0» Haze Thickness Map (HTM) (A. Makarau et al., 2014) precise detection for an inhomogeneous and structured haziness Haze detection: two blue bands Band extrapol = (Band 443nm + (Band 443nm 0.95 Band 490nm )) > 0 Haze removal» Dark-object Subtraction (DOS) (G. Dal Moro et al., 2007) subtraction of the haze thickness based on several pixels» Improved DOS(A. Makarau et al., 2014) subtraction of the haze thickness based on all pixels 8

9 Method Haze detection derive R haze 1.Detect Haziness Band extrapol Band extrapol = (Band443nm + (Band443nm 0.95 Band490nm)) > 0 9

10 Method Haze detection derive R haze 1.Detect Haziness Band extrapol 2.Generate HTM for the whole image HTM image HTM image = MIN Radiance (Band extrapol, window 3x3 ) 10

11 Method Haze detection derive R haze 1.Detect Haziness Band extrapol 2.Generate HTM for the whole image HTM image 3.Optimize HTM image HTM correct HTM correct =triangular_interpolation(htm image,*bright_objects) * can be derived from Sentinel3 OLCI L1B flag dataset bright pixels 11

12 Method Haze detection derive R haze 1.Detect Haziness Band extrapol 2.Generate HTM for the whole image HTM image 3.Optimize HTM image HTM correct 4.Separate haze pixels and non haze pixels HTM mask HTM mask = MIN Radiance (Band extrapol, window 21x21 ) 12

13 Method Haze detection derive R haze 1.Detect Haziness Band extrapol 2.Generate HTM for the whole image HTM image 3.Optimize HTM image HTM correct 4.Separate haze pixels and non haze pixels HTM mask 5. Generate HTM for each band HTM bandi HTM Bandi = MIN Radiance (Band i, window 3x3 ) 13

14 Method Haze detection derive R haze 1.Detect Haziness Band extrapol 2.Generate HTM for the whole image HTM image 3.Optimize HTM image HTM correct 4.Separate haze pixels and non haze pixels HTM mask 5. Generate HTM for each band HTM bandi 6. Calculate regression coefficient k k = regression_fit(htm image, HTM Bandi )! k is decreased with band wavelength 14

15 Method Haze detection derive R haze 1.Detect Haziness Band extrapol 2.Generate HTM for the whole image HTM image 3.Optimize HTM image HTM correct 4.Separate haze pixels and non haze pixels HTM mask 5. Generate HTM for each band HTM bandi 6. Calculate regression coefficient k 7. Calculate haze radiation contribution R haze R haze = f(htm image, HTM bandi ) = k HTM correct 15

16 Method Haze detection derive R haze 1.Detect Haziness Band extrapol 2.Generate HTM for the whole image HTM image 3.Optimize HTM image HTM correct 4.Separate haze pixels and non haze pixels HTM mask 5. Generate HTM for each band HTM bandi 6. Calculate regression coefficient k 7. Calculate haze radiation contribution R haze 1. Band extrapol = (Band443nm + (Band443nm 0.95 Band490nm)) > 0 2. HTM image = MIN Radiance (Band extrapol, window 3x3 ) 3. HTM correct =triangular_interpolation(htm image,*bright_objects) 4. HTM mask = MIN Radiance (Band extrapol, window 21x21 ) 5. HTM Bandi = MIN Radiance (Band i, window 3x3 ) 6. k = regression_fit(htm image, HTM Bandi )! k is decreased with band wavelength 7. R haze = f(htm image, HTM bandi ) = k HTM correct * can be derived from Sentinel3 OLCI L1B flag dataset: bright pixels 16

17 Results Data source: live Sentinel 3 OLCI Level 1B data from EUMETcast Data collection time period: Dec Hazy image is detected on , UTC 7:19 17

18 Results Haze detection: image HTM and band-dependent HTM Detect haze: Band extrapol Estimate haze thickness: HTM image Segment HTM: HTM mask Correct HTM: HTM correct Band 1 Estimate haze thickness at Band 1: HTM Band1 Band 9 Estimate haze thickness at Band 9: HTM Band9 18

19 Results Haze removal including aerosol compensation» Single band e.g. Oa2 Band 2 HTM correct Band 2_dehazed Band 2_compensated 19

20 Results Haze removal including aerosol compensation» RGB composite: R(Oa10), G(Oa6), B(Oa3) Legend Cloud mask Hazy RGB image Dehazed RGB image 20

21 Evaluation Compare hazy image with haze free image» Minimal difference of the image acquisition conditions: time and Sun/sensor geometry Data acquisition time: 08/12/2016 UTC 07:19 04/12/2016 UTC 07:23 21

22 Evaluation Compare hazy image with haze free image» Averaged TOA radiance of hazy/haze-free region on hazy image and haze-free reference image, covering the spectrum from 400 to 1020 nm Haze-free region: all the non-haze and cloud-free pixels of the hazy image Hazy region: all the hazy and cloudfree pixels of the hazy image 22

23 Conclusions The new haze detection and removal method succeeded in processing haziness of Sentinel 3 OLCI L1B image data. The algorithm can be easily integrated with other data processors in remotely sensed data processing chain. Based on this algorithm, a dehazing processor for multispectral meteorological satellite data will be further developed and can be applied on original data digital numbers (DNs) counts or calibrated radiance data. 23

24 Thank you for your attention