Geographical Weighted Regression Model for Improved Near-shore Water Depth Estimation from Multispectral Imagery Vinayaraj Poliyapram 1 , Venkatesh Raghavan 1 and Masumoto Shinji 2 1 Graduate School for Creative Cities, Osaka City University 3-3-138 Sugimoto, Osaka 558-8585, Japan Email: vinay223333@gmail.com 2 Graduate School of Science, Osaka City University. 1
Presentation Outline : INTRODUCTION Optical Remote Sensing for near-shore depth Physical principles MATERIALS AND METHOD FOSS4G used (GRASS GIS, R and Qgis) Spectral radiance correction method Study area and data used Global regression model Motivation to geographically fitted models Geographical Weighted Regression (GWR) model RESULT AND DISCUSSION Evaluation of estimated depth at different scenarios Evaluation by near-shore cross profiles CONCLUSIONS 2
Remote Sensing of Water Depth Source: NOAA Multibeam echo sounding Source: Fugro LiDAR bathymetry Available Bathymetry data 1.DBDB5 2.ETOPO5 3.ETOPO2 4.ETOPO1 5.GEBCO 3
Bathymetric Mapping using Multi-spectral Imagery Advantages Wide availability of data: IKONOS, QuickBird, RapidEye, Landsat, etc Relatively low cost Large spatial coverage, high spatial resolution Derive water depth up to 20 m, depending on water turbidity and atmospheric conditions Disadvantages Relatively low accuracy Need water depth truth 4
Bathymetric Mapping using Multi-spectral Imagery Physical Principle Attenuation of light in water is a function of wavelength, at shorter wavelength region in electromagnetic spectrum attenuates slowly and consistently increases the attenuation speed when wavelength increases. Meanwhile, as depth increases light attenuates rapidly and separability of bottom type spectra declines. Schematic view of radiance components Behavior of attenuation of light at observed using the visible sensor over near-shore water 5 optically shallow water (Kanno et al., 2012)
Study Area 67 ° 18' 67 ° 00' 18 ° 5.4' 18 ° 5.4' Puerto Rico Study area 17 ° 49.2' 17 ° 49.2' 0 2 4 6 8 KM ' 67 ° 18' 67 ° 00' Puerto Rico Fig.4. Puerto Rico, Northeastern Caribbean Sea 6
Data sets Data Date of Resol Spectral bands Band used collection ution used for estimation for correction 0.43-0.45 µm (coastal Landsat 8 25 November, 30 M 1.57-1.65 µm (SWI) Aerosol) 2013 0.45-0.51 µm (blue) 0.53-0.59 µm (green) 0.64-0.67 µm (red) 0.85 - 0.88 µm (NIR) 0.44-0.51 µm (blue) RapidEye 01 May, 2010 5 M 0.76-0.85 µm (NIR) 0.52-0.59µm(green) 0.63-0.69µm (red) 0.69-0.73µm (Red-edge) LiDAR 12 November, 4 M 7 depth 2006
Correction method to remove unwanted components Lyzenga (1978; 1981; 1985) , Philpot (1989) and Kanno and Tanaka (2012) Transformed Area of interest Band ( X(λ) i ) (Shallow water) Landsat 8 Transformed band (X( λ) i) = log ((L λ i - α0 - α1 L( λ SWI))/ L λ i) SWI =1.57-1.65µm RapidEye Transformed band (X( λ) i)= log ((L λ i - α0 - α1 L( λ NIR))/ L λ i) NIR= 0.76-0.85µm Where, L λ i is the radiance value of selected region, α0 is the intercept and α1 is the slope of the regression. 8
Water Depth Retrieval Models Global Regression Model Only one regression equation is used for the entire image The conventional regression models are a common practice by several authors (Clark , (1987; 1988), Lyaenga, (2006) and Kanno and Tanaka, (2012). Multi-band method Depth= β 0 + β 1 X( λ) 1 + β 2 X( λ) 2 +…….+β n X( λ) i NIR band Red band β 0, β 1, β 2… β n + LiDAR Green band Blue band Coastal band Transformed bands (X( i ) Reality heterogeneous bottom types and varying water quality 9
Motivation to the Geographically Fitted Models Assessment of spatial propagation of heterogeneity Residual map=Depth estimated by global regression model-LiDAR depth LiDAR reference depth (m) R=0.88 Residual map of global regression model Solution Clear indication of different spatial clusters of residuals 10
Motivation to the geographically fitted models Assessment of spatial heterogeneity by classes 6 residual classes clustering 11
Motivation to the Geographically Fitted Models Depth class = β 0 + β 1 X( λ) 1 + β 2 X( λ) 2 +…….+β n X( λ) i Residual class based model R 2 Method R RMSE R=0.94 (m) Global model 0.88 0.78 2.63 Class based 0.94 0.90 1.70 model 12
Geographically Weighted Regression (GWR Models) (x, y): geographical coordinates of the centroid point for each local area i x,y d Depth= β 0(x,y) + β 1(x,y) X 1(x,y) + β 2(x,y) +X 2(x,y) + +……..+β n(x,y) + X i(x,y) Gaussian w = exp(-0.5 * (d / bw)2) i Kerne l Yellow points are LiDAR depth displayed on Landsat 8 13
Results and comparison of models at various scenarios 14
Results and comparison of models at various scenarios Scenario1 Randomly distributed LiDAR depth points used for calibration Landsat 8: 10,000 points RapidEye: 60,000 point Evaluation carried out with separate depth points Yellow points are LiDAR depth displayed on RGB (NIR,Red,Green) of Landsat 8 Satellite data Global model GWR model R R2 RMSE R R2 RMSE (m) (m) Landsat 8 0.88 0.78 2.63 0.97 0.93 1.41 RapidEye 0.88 0.78 2.48 0.97 0.93 1.34 15
Results and comparison of models at scenario1 LiDAR depth m R=0.88 R=0.97 LiDAR depth m Landsat 8 Estimated depth (GWR model) m Estimated depth (Global model) m LiDAR depth m R=0.88 R=0.97 LiDAR depth m RapidEye Estimated depth (Global model) m Estimated depth (GWR model) m LiDAR depth LiDAR depth RapidEye Landsat 8 16
Results and comparison of models at various scenarios Scenario2 LiDAR depth points are at equal interval of 300m Landsat 8: 1750 points RapidEye: 1750 point Evaluation carried out with separate depth points Yellow points are LiDAR depth displayed on RGB (NIR,Red,Green) of Landsat 8 Satellite data Global model GWR model R R2 RMSE R R2 RMSE (m) (m) Landsat 8 0.88 0.78 2.63 0.96 0.91 1.72 RapidEye 0.89 0.78 2.51 0.95 0.91 1.61 17
Results and comparison of models at scenario2 LiDAR depth m R=0.88 R=0.96 LiDAR depth m Landsat 8 Estimated depth (GWR model) m Estimated depth (Global model) m LiDAR depth m R=0.89 LiDAR depth m R=0.95 RapidEye Estimated depth (Global model) m Estimated depth (GWR model) m LiDAR depth LiDAR depth RapidEye Landsat 8 18
Results and comparison of models at various scenarios Scenario3 LiDAR depth points are at equal interval of 600m Landsat 8: 450 points RapidEye: 450 points Evaluation carried out with separate depth points Yellow points are LiDAR depth displayed on RGB (NIR,Red,Green) of Landsat 8 Satellite data Global model GWR model R R2 RMSE R R2 RMSE (m) (m) Landsat 8 0.88 0.78 2.62 0.94 0.88 1.91 RapidEye 0.88 0.78 2.48 0.93 0.87 1.87 19
Results and comparison of models at scenario3 LiDAR depth m R=0.94 LiDAR depth m R=0.88 Landsat 8 Estimated depth (GWR model) m Estimated depth (Global model) m R=0.93 LiDAR depth m R=0.88 LiDAR depth m RapidEye Estimated depth (Global model) m Estimated depth (GWR model) m LiDAR depth LiDAR depth RapidEye Landsat 8 20
Results and comparison of models at various scenarios Scenario4 Random LiDAR depth points are at small portion of study area Landsat 8: 3300 points RapidEye: 20,000 points Evaluation carried out with separate depth points Yellow points are LiDAR depth displayed on RGB (NIR,Red,Green) of Landsat 8 Satellite data Global model GWR model R R2 RMSE R R2 RMSE (m) (m) Landsat 8 0.83 0.68 3.49 0.87 0.75 2.86 RapidEye 0.88 0.78 2.60 0.90 0.81 2.26 21
Results and comparison of models at scenario4 LiDAR depth m R=0.83 R=0.88 LiDAR depth m Landsat 8 Estimated depth (GWR model) m Estimated depth (Global model) m LiDAR depth m R=0.88 R=0.90 LiDAR depth m RapidEye Estimated depth (Global model) m Estimated depth (GWR model) m LiDAR depth LiDAR depth RapidEye Landsat 8 22
Results and comparison of models by using near- shore cross profiles spaced 2.5km Global model Global model LiDAR depth LiDAR depth GWR model GWR model Global model Global model LiDAR depth LiDAR depth GWR model GWR model 23
Results and comparison of models by using near- shore cross profiles spaced 2.5km Global model Global model LiDAR depth LiDAR depth GWR model GWR model Global model Global model LiDAR depth LiDAR depth GWR model GWR model Global model LiDAR depth GWR model 24
Conclusions and future study 12-bit dynamic range was provided by Landsat 8 and RapidEye images were used to estimate good accuracy near- shore water depth. Global regression model was used to estimate depth was not fully able to address the spatial non-stationary introduced by various bottom types and water quality. Meanwhile, GWR model was able to address the spatial non- stationary introduced by various bottom types and water quality. Therefore produces high accuracy depth estimation. GWR models can give better and more consistent depth estimates than the conventional global regression model. 25
Thank you for your time 26
Recommend
More recommend
