MTRI Water Quality Remote Sensing Capabilities MTRI Remote
MTRI Water Quality Remote Sensing Capabilities MTRI Remote Sensing Water Quality Toolbox Robert Shuchman, Michael Sayers, Karl Bosse, Reid Sawtell, Amanda Grimm
NGA CRADA: Bathymetry and Bottom Type Mapping IR&D Objective: Demonstrate MTRI’s state of the art bathymetry and bottom type mapping algorithm on a 50 x 50 km test site mutually selected by MTRI and NGA personnel Algorithm uniqueness: – 2 nd/3 rd generation extension of Lyzenga and Stumpf work – Does not assume a singe Kd value for entire scene – Automated process examines a multi-temporal satellite image stack to evaluate a set of images to be used in the product generation – Algorithm is adaptable in respect to estimating required input parameters (depth, Kd, bottom type/reflectance) from multi-temporal observations 2
NGA CRADA: Bathymetry and Bottom Type Mapping IR&D (Continued) IR&D Program – 6 month duration – 50 x 50 km test site of coastal or inland waters – Test site should have ground truth so that algorithm performance can be evaluated IR&D Key Program Elements – – – Select test site (MTRI+NGA) Acquire imagery (MTRI+NGA) Generate bathymetry and bottom feature maps (MTRI) Evaluate products (MTRI) Brief NGA on results (MTRI+NGA) Provide report in power point format (MTRI) 3
NGA CRADA: Bathymetry and Bottom Type Mapping IR&D (Continued) Deliverables – – Bathymetry and bottom feature maps of 50 x 50 km test site Evaluation of algorithm performance Brief to NGA our IR&D results Report in power point format 4
Great Lakes Water Quality Assessment using Tuned Ocean Color to Lake Color Algorithms MTRI has developed a robust algorithm toolbox for water sensing Remote sensing inputs include satellite, airborne, UAV, ship-based, and handheld Motivation: Optically complex water such as coastal and inland system require specific tuned algorithms – Ocean Color standard algorithms do not perform adequately in these systems 5
MTRI Water Sensing Toolbox Generated Parameters – Bathymetry – Optically shallow bottom mapping – Water Visibility – Water Type Mapping – Chlorophyll – CDOM (estimate DOC) – Suspended sediments – Light extinction coefficient (Kd) – KPAR – Photic Zone Depth – Bulk absorption – Bulk scattering – Cyanobacteria Index (CI) – Harmful Algal Blooms (HABs) – Surface HABs scum – Primary productivity – Algal species composition – Particle size distribution – Data assimilation products 6
Light Interaction in Water Inherent Optical Properties (IOPs) are dependent on the constituents and their concentrations within the water column – Examples: absorption (a), scattering (b), beam attenuation coefficient – a. TOTAL = a. CDOM + a. PHY + a. NAP+ a. W – b. TOTAL = b. PHY + b. NAP + b. W IOPs provide information about the water constituents Apparent Optical Properties (AOPs) are dependent on solar geometry, surface roughness, lake bottom reflectance, and the IOPs – Examples: water leaving radiance (n. Lw), remote sensing reflectance (Rrs), diffuse attenuation coefficient (Kd) MTRI utilizes physics based approaches in our algorithms 7
MTRI/GLERL Color Producing Agents Algorithm (CPA-A) Satellite Observed Reflectance Specific IOPs Remote Sensing Reflectance (sr-1) 0. 025 MODIS Bands 0. 02 0. 015 0. 01 0. 005 0 400 450 500 550 600 650 700 750 Wavelength (nm) Shuchman, R. , Leshkevich, G. , Sayers, M. , Johengen, T. , Brooks, C. , & Pozdnyakov, D. (2013). An algorithm to retrieve chlorophyll, dissolved organic carbon, and suspended minerals from Great Lakes satellite data. Journal of Great Lakes Research, 39, 14 -33.
Radiative Transfer Modeling MTRI water quality toolbox algorithms are physics based Radiative transfer modeling using Hydrolight is gold standard for modeling light interaction in water Used by MTRI for physics based algorithm development Used to assess optical closure and providing insight into hyperspectral algorithm development 9
CPA-A Example Output Products This algorithm is now part of operational NOAA NESDIS production 10
CPA-A Evaluation using MODIS Comparison between MODISderived OC 3 and CPA-A-derived chlorophyll concentration and NOAA/MTRI in situ chlorophyll concentrations In situ data from 2010 -2012 Comparisons were made against satellite retrievals collected within 24 hours of in situ measurements CPA-A produces more accurate chlorophyll than OC 3 in both nearshore and offshore % Difference RMSE CPA-A 36. 6 0. 593 OC 3 75. 6 8. 086 Fahnenstiel, G. L. , Sayers, M. , Shuchman, R. A. , Yousef, F. , and Pothoven, S. , 2016. Lake-wide phytoplankton production and abundance in the Upper Great Lakes: 2010 -2013. Journal of Great Lakes Research, 42(3), pp. 619 -629.
Operational CPA-A Processing by NESDIS MTRI/GLERL-developed CPA-A VIIRS algorithm was successfully transitioned to NESDIS processing all Science Quality VIIRS images with CPA Algorithm on an approximately 8 day delay Posting derived products to Great Lakes Coast. Watch Products generated include chlorophyll, DOC, and SM concentrations, CDOM absorption, KD 490, KDPAR, and photic zone depth Derived products available to download in HDF, Geo. TIFF, Net. CDF, and PNG formats
Western Basin Lake Erie Harmful Algal Blooms (HABs) MTRI uses the CPA-A and a surface scum index to assess the average and maximum extent of algal blooms and surface scums each year in WBLE Algorithm uniquely addresses both surface and water column mixed HABs 13
Annual Extent of WBLE HABs Using the Sea. Wi. FS and MODIS sensors, the annual surface scum and HAB extent in WBLE has been studied going back to 1998 19 22 05 /1 6/ 90 24 5 /1 6/ 90 26 5 /1 6/ 90 28 5 /1 6/ 90 30 5 /1 9 7/ 05 2/ 19 7/ 05 4/ 19 7/ 05 6/ 19 7/ 05 8/ 19 05 0 6/ 6/ 1500 1000 500 0 Maximum Annual Surface Scum Extent (km 2) 300 2000 400 350 300 250 200 150 100 50 0 20 / 6/ 190 22 5 / 6/ 190 24 5 / 6/ 190 26 5 / 6/ 190 28 5 / 6/ 190 30 5 /1 7/ 905 2/ 1 7/ 905 4/ 1 7/ 905 6/ 1 7/ 905 8/ 19 05 600 2500 6/ 900 Maximum Annual HAB Extent (km 2) 1200 90 80 70 60 50 40 30 20 10 0 Average Annual Surface Scum Extent (km 2) 1500 20 / Average Annual HAB Extent (km 2) From 1998 to 2017 there was a significantly increasing trend in average and maximum submerged and surface algal blooms Sayers, M. , Grimm, A. , Shuchman, R. , Bosse, K. , Fahnenstiel, G. , Ruberg, S. , Leshkevich, G. , 2019. Satellite Monitoring of Harmful Algal Blooms in the Western Basin of Lake Erie: a 20 -Year Time-Series. Journal of Great Lakes Research. 14
Annual Spatial Extent of WBLE HABs Annual heat maps of HABs in WBLE show the spatial distribution of blooms and how it changes year-to-year Most impacted area is in Maumee Bay and along the Ohio coastline Very little HAB presence in the Detroit River outflow Time series heat maps can be temporally scaled to specific intelligence requirements 15
Remote Sensing Derived Primary Productivity in the Great Lakes Primary production is the process where algae utilizes carbon dioxide to generate sugar through photosynthesis Extended historical Lake Michigan Analysis Summary Annual Carbon Fixation for the Great Lakes 2011 –Superior (7. 4) –Michigan (5. 9) –Huron (5. 4) –Erie (10. 1) –Ontario (3. 8) – 33 Tg C/Year TOTAL 0. 07% of Worldwide Estimate of Oceanic Carbon Fixation 73% decrease from 1980 -2012 Satellite-derived Primary Productivity down majorly in the years since mussel invasion began Quagga Mussel Density 1994/95 2000 2005 2010 Shallow is significantly lower for all three lakes than the mid and deep zones Production is similar for all three lakes (Michigan, Huron, Superior) Lakes Michigan and Huron shallow water area are most influenced by mussels Lakewide production is driven by mid and deep depth zones contributions Remote sensing can be used to assess long term water security 16
Global Chlorophyll Distribution Observable lakes concentrated in northern hemisphere Chlorophyll variable among and within FEOW ecoregions Regional chlorophyll averages not particularly useful 17
Great Lakes Water Clarity Ocean color satellite data can also be used to study water clarity in the Great Lakes First Optical Depth Trends from 2003 -2013 show that Lake Huron has become more clear than Lakes Michigan and Superior in offshore waters (>30 m) The CPA-A is better able to estimate light attenuation (KD 490) in the Great Lakes than the band ratio algorithm used by MODIS (RMSE: 0. 25, 0. 77, respectively) 18
Bathymetry and Bottom Type Mapping MTRI has built upon the work of Lyzenga, Stumpf, and others MTRI algorithm utilizes multiple datasets as inputs to dynamically estimate light extinction on a pixel by pixel basis which is a major advancement We have developed novel procedures to obtain other required input parameters (i. e. we need two of three parameters (depth, bottom type, in water properties)) MTRI can map bathymetry and/or bottom type in denied areas 19
Updating Bathymetry with Multi-Source Remote Sensing – National Park Service Multiple incomplete datasets (Li. DAR, Sonar) existed. We used high resolution commercial and Landsat imagery to fill in the gaps. Created multi-source updated bathymetry map for NPS Can be applied elsewhere in US & worldwide 138 136 134 132 130 128 126 124 RMSE = 0. 42 Li. DAR Elevation World View 2 Elevation 0 37 73 110 146 183 220 256 293 329 366 403 439 476 513 549 586 622 659 696 732 769 805 842 879 915 NPS task to generate a seamless bathymetric map of the SBDNL. Elevation (meters) Transect Length (meters)
Automated Bathymetry Estimation in Remote Alaska
Automated Bathymetry Estimation in Remote Alaska
Seasonal variation in SAV Extent Increased looks gives us a “heat map” of SAV persistence – Created by summing 5 classified 2018 L 8 images from 4/21, 5/22, 7/25, 8/10, and 9/11 – Shows that core (darker) areas are persistent across the growing season, with larger peripheral areas (lighter) that are colonized during peak growing conditions
Sleeping Bear Dunes National Lakeshore - Lake Michigan Updated 2018 SAV Distribution 2010 -2018 Change 24
400 20 350 18 Mean Bottom Detection Depth Limit 16 300 14 250 12 200 10 150 8 6 100 4 50 Mean BDDL (m) Mapped Area (km 2) Lake Bottom Mapping Pixels outside standardized area classified as uncolonized substrate Standardized area classified as uncolonized substrate Pixels outside standardized area classified as SAV Standardized area classified as SAV 2 0 0 1974198119861990199620012006201020122018 25
Automated Algorithm to Generate Depth Invariant Bottom Reflectance from Multiple Remote Sensing Platforms – Great Lakes Example Problem: need to map submerged aquatic vegetation (SAV) extent for all optically shallow regions throughout all 5 great lakes – Nearly 10, 000 miles of shoreline, plus offshore shallow regions – SAV is time dependent, it grows and decays, spreads or recedes • Multiple images over multi-year period (2015 -Present), from multiple sensors (LS 8, Sentinel), and of varying quality (atmospheric contamination, water quality) Lake bottom classification requires correction of light attenuation as a function of water depth – Each class: sand, vegetation, or rocks, needs to look the same at 3 ft or 10 ft • Naturally all appear darker as water depth increases, following exponential decay profile – Common method: Lyzenga depth invariant bottom index • Can be successfully employed to distinguish SAV from other bottom types under certain conditions 26
Motivation for Bottom Type Mapping Automation Existing methods require too much manual interpretation, limits the number of images that can be processed in the time available – Or requires higher cost to pay for the labor – The MTRI approach incorporates automated atmospheric correction using ACOLITE software, then in this case applying NOAA bathymetry to automatically infer deep water vs shallow water inputs to deglinting and bottom invariant index algorithms • Still needs per-image tuning of Kd • Landsat 8 tiles are huge and the lake conditions can quickly change, algorithm assumptions (constant Kd) often don’t hold MTRI developed a system that automatically accounts for variation in Kd – Produces a single classification at the end – Incorporate quality checks to filter bad images (otherwise a human still has to perform that evaluation and make decisions on how to merge) • Facilitates processing of many images 27
Algorithm Challenges The larger the area of interest, the more likely it is that the core assumptions of depth invariance algorithms break: – Water or atmospheric conditions change sufficiently to create class ambiguity • River Plumes • Cloud Shadows • Reef Suspension River Plume Clear Water 28
Algorithm Challenges Addressed Solution: Split each image into smaller, overlapping segments – Need to do this intelligently so we have a good cross-section of shallow to deeper water, can’t just grid it • In this case use NOAA depth contours to inform selection – Forward (shallow to deep) and Reverse (deep to shallow) ensures dense coverage of all areas of interest 29
MTRI Automated Depth Invariant Algorithm Level 1 Ocean Color Data Atmospheric Correction Depth Deglinting Image Segmentation Attenuation Calculation Radiance Correction per Segment Depth Invariant Image Many Methods for Lake Bottom Mapping Lyzenga Method Supervised Classificati on Unsupervis ed Classificati on Neural Nets Deep Learning Radiative Transfer Other Bottom Feature Map Atmospheric correction & deglinting Attenuation calculation and radiance correction Classification 30
Automated Attenuation (Kd) Analysis For each sub-image, plot pixel radiance vs. depth – For a two-bottom system, this will result in two interposed curves plus noise • Spatial distribution of bottom types can complicate this • Need robust routine to fit exponential curve to observed signal – Most important quantity is the slope of the exponential (proxy for KD) Invert equation to ‘solve’ for bottom reflectance – Output quantity contains all our errors (depth error, fit error, sensor noise, etc. ) 31
Automated Bottom Type Classification After depth invariant radiance generation classification is performed Given each sub-image, run a statistical analysis to separate light/dark – Analysis based on histogram distribution of intensity • Single mode, most common, large sand areas with small SAV ‘spots’ – Mixed pixel and SAV density variation erase distribution boundaries » The SAV is the tail of the distribution, use well chosen threshold below the peak of the distribution as class boundary • Dual mode, SAV and sand are more equally present – SAV and sand appear as distinct but overlapping distributions – Choose the split point between distributions as class boundary
Single vs Dual Bottom Types Single and Dual Gaussian fits (seen right), are computed for the histogram distribution. – Quality of fit determines which is used (E% is fit error for Single and Dual respectively) Median used to determine thresholds in Single mode (seen above) Intersection of distributions is used in Dual mode (not shown)
Tuning Bottom Type Classification We now have results for all sub images, but some are full of error, how do we produce a consistent output? – Remember our sub-images overlap, we can weight them according to all the problems we recorded in the previous steps • Eg: clouds presence means really low, or zero, weight – Output the weighted classification, and the sum of the weights • Works because our classes are a binary mixing problem – Assign value 0 to SAV, . 5 to moderate SAV, 1 to sand – Multiply class value by weight – Divide sum of weighted class value by sum of weights • Low output weights indicate we don’t have confidence in the result – We now have a single full scene image that is analogous to one of our sub images: we have a classification and a measure of our confidence in the result • Can use the same procedure to combine multiple images
Preprocessing ACOLITE Atmospheric Correction Input: Landsat 8 Automated Deglint Correction Lyzenga Bottom Index Deep Water Values Interpolated to Shallow Regions Input: NOAA Depth User Defined Bottom Classification Input: Coastal LIDAR Per Fragment 6 m Contour Shoreline Extraction 30 m Contour Forward Window Generator Reverse Window Generator Depth vs. Radiance Curve Fit Depth Invariant Bottom Max Visible Depth Statistical Light vs. Dark Separation Flags Bottom Classification Weighted Window Integration Flags Classification Final Classification Multi-Image Weighted Integration Valid Mask
Bottom Type Mapping Algorithm Output and Flags Final classification includes non-shore water and land in addition to Sand, Moderate SAV, and Dense SAV Flags are output as an integer, each bit indicating a certain issue that can be weighted according to it’s severity
Bottom Type Map Comparison Atmospherically Corrected & Deglinted Our Automated Analysis Lyzenga With User Defined Thresholds
Summary of MTRI Bottom Type Mapping Algorithm New technique successfully used to process hundreds of scenes Very effective in generating 20 year time-series Invaluable for seasonal analysis Built-in flags allow for utilizing optimal data sets as well as assess image information robustness Side product of algorithm is near-shore water clarity assessment
Hyperspectral Remote Sensing of Water Quality Hyperspectral data provides more information about constituent composition ― ― ― Diagnostic pigment absorption Spectral scattering Inelastic scattering Satellite hyperspectral systems are limited (Hyperion, HICO) NASA Plankton, Aerosol, Cloud, ocean Ecosystem (PACE) mission will launch 2022 ― ― ― Hyperspectral 350 -885 nm Polarimeters (hyperspectral and hyperangular) Daily revisit 0. 025 Remote Sensing Reflectance (sr-1) Hyperspectral MODIS Bands 0. 02 0. 015 0. 01 0. 005 0 400 450 500 550 600 650 700 750 Wavelength (nm) 39
MTRI Hyperspectral Algorithm Specific IOP Input Parameters Bricaud et al. Sayers et al. Roesler et al. Observed Reflectance
MTRI Hyperspectral Algorithm Cyanobacteria bloom condition in Lake Erie Dominant pigments present: Chla, Phycocyanin (mg/m 3) = apc 620/apc*620 (Simis et al. 2005) = 0. 23/0. 01 = 23 mg/m 3 Lab extracted Phycocyanin = 33 mg/m 3 Can identify Phycocyanin fluorescence? 41
- Slides: 41