a rapid extraction of water body features from antarctic coastal oasis using very high resolution satellite remote sensing data

Peer-review under responsibility of organizing committee of ICWRCOE 2015 doi: 10.1016/j.aqpro.2015.02.018 INTERNATIONAL CONFERENCE ON WATER RESOURCES, COASTAL AND OCEAN ENGINEERING ICW

Aquatic Procedia 4 ( 2015 ) 125 – 132

2214-241X © 2015 Published by Elsevier B.V This is an open access article under the CC BY-NC-ND license

(http://creativecommons.org/licenses/by-nc-nd/4.0/).

Peer-review under responsibility of organizing committee of ICWRCOE 2015

doi: 10.1016/j.aqpro.2015.02.018

INTERNATIONAL CONFERENCE ON WATER RESOURCES, COASTAL AND OCEAN

ENGINEERING (ICWRCOE 2015)

A Rapid Extraction of Water Body Features From Antarctic Coastal Oasis Using Very High-Resolution Satellite Remote Sensing Data

S D Jawaka* and A J Luisa

a Polar Remote Sensing Department, Earth System Science Organization (ESSO), National Centre for Antarctic & Ocean Research (NCAOR),

Ministry of Earth Sciences, Govt of India, Headland Sada, Vasco-da-Gama, Goa - 403804, India

Abstract

Antarctic coastal oases are essential sources of spatially distributed fresh water bodies Mapping water bodies from remote places, such as polar regions, using traditional surveying method is a laborious and logistically expensive task A rapid method for extracting and monitoring water bodies in Antarctic coastal oases has a tremendous application in remote sensing This study discusses the design of a rapid and novel method to extract water body features in Antarctic coastal oasis environment from remotely-sensed images We devised semiautomatic approach for extracting water body features based on a novel set of normalized difference water index (NDWI) by incorporating high-resolution WorldView-2 (WV-2) panchromatic and multispectral image data This study highlights and compares the viability of state-of-the-art spectral processing water body extraction approaches with the newly designed NDWI approach An extensive quantitative evaluation was carried out to test the newly designed NDWI approach for extracting water bodies on Larsemann Hills, eastern Antarctica The results suggest that the modified NDWI approach render intermediate performance with bias error varying from ~1 to ~34 m 2 (least amount of misclassified pixels) We also analyzed the distinctive 8-band capability of WV-2 data coupled with semiautomatic extraction methods to compare their reliability in extracting the water body area The results indicate that the use of the modified NDWI approach on 8-band WV-2 data can significantly improve the semiautomatic extraction of water body features, which can ultimately contribute to an enhanced perceptive of the Antarctic coastal oasis in the context of climate change

© 2015 The Authors Published by Elsevier B.V

Peer-review under responsibility of organizing committee of ICWRCOE 2015

Keywords: Waterbody extraction; WorldView-2; NDWI

* Corresponding author Tel.: +91-832-2525528; fax: +91-832-2520877

E-mail address: shridhar.jawak@ncaor.gov.in

© 2015 Published by Elsevier B.V This is an open access article under the CC BY-NC-ND license

(http://creativecommons.org/licenses/by-nc-nd/4.0/).

Peer-review under responsibility of organizing committee of ICWRCOE 2015

1 Introduction

Water on the Earth’s surface is an essential part of the hydrological cycle Being able to access the spatial distribution and geographical extent information on water bodies in real time has great significance in limnology and for understanding interactions between regional hydrology and climate change (Papa et al., 2008) Satellite remote sensing has advantages because it can track land surface information in real-time macroscopically, multitemporally, multispectrally, dynamically, and repetitively; hence, it is appropriate for surveying and mapping surface water bodies (Chen et al., 2004)

Antarctica’s inclement weather, a few number of fine weather days in summer, and the high cost of ship/ helicopter restricts research trips to Antarctica Therefore, high-resolution satellite remote sensing data and aerial photography are important sources of information for monitoring the short-term and long-term changes that occur at

a specific location in Antarctica over time Although high resolution remote sensing data can never replace aerial photographs, which provide images at a resolution as high as 0.2–0.3 m, the WorldView-2 (WV-2) is found to be suitable for semiautomatic feature extraction in the Antarctic, where frequent aerial photography is difficult because

of the harsh environment and high costs of logistics (Jawak and Luis, 2012; Jawak and Luis, 2011) Hence, development of automated or semiautomated feature extraction methods using high-resolution remote sensing data is much needed to continuously monitor the geographical features in a cryospheric environment

At present, the methods for extracting surface water bodies are prominently based on spectral index or multiband techniques, such as the normalized difference water index (NDWI) (McFeeters, 1996; Lacaux et al., 2007) A novel water extraction index for shoreline delineation by combining the tasseled cap wetness index and the NDWI was proposed by Ouma and Tateishi (2006) Rogers and Kearney (2004) proposed the NDWI for the Moderate Resolution Imaging Spectroradiometer (MODIS) multispectral satellite images (MSI) Lu et al (2011) suggested an integrated water body extraction method with HJ-1A/B satellite imagery by using normalized difference vegetation index (NDVI) and NDWI These modified indices have been frequently used to map surface water bodies using Landsat and MODIS images (Li and Zhou, 2009; Soti et al., 2009) However, because of the complications in cryospheric environments, diverse ground targets may have the similar spectrum characteristics Therefore, only one category of index method can extract water bodies only under certain conditions (Jawak and Luis, 2013a; Jawak and Luis, 2013b) Additionally, these spectral indices were developed for traditional visible near-IR (VNIR) systems Hence, the WV-2 offers a prospect to adapt a modification of NDWI using eight spectral bands (Jawak et al., 2013) Jawak and Mathew (2011) proposed an object oriented method for semiautomatic extraction of roads and water bodies using QuickBird imagery Matched Filtering (MF), Mixture Tuned Matched Filtering (MTMF), Spectral Angle Mapper (SAM), MF/SAM ratio, and Principal Component Analysis (PCA), have been implemented for improved lake feature extraction in cryospheric environment (Jawak and Luis, 2014a; Jawak and Luis, 2014b) Target extraction methods, such as Constrained Energy Minimization (CEM), Adaptive Coherence Estimator (ACE),

Target-Constrained Interference-Minimized Filter (TCIMF), Mixture Tuned TCIMF (MT-TCIMF), and Orthogonal

Subspace Projection (OSP) methods have been used to improve semiautomatic target detection (Jawak and Luis, 2014) In this work, we performed semiautomatic extraction of water bodies in the Larsemann Hills of Antarctica by employing 14 types of pixel-wise methods Our experiment is focused on two objectives: (a) designing a modified NDWI approach to extract water body features, and (b) comparing the performance of supervised feature extraction algorithms with the newly developed modified NDWI approach using visual analysis and statistical accuracy

2 Study area and data

The Larsemann Hills are located on the Ingrid Christensen Coast, Princess Elizabeth Land, eastern Antarctica The Larsemann Hills is roughly situated in the middle of the Amery Ice Shelf and the Vestfold Hills The region hosts a number of water bodies, ranging from small, shallow ponds (<1 m deep) to glacial lakes (~38 m deep), with areas ranging between 100 and 33,000 m2

We employed radiometrically corrected, georeferenced, orthorectified 16-bit standard level 2 (LV2A) WV-2 multisequence data The WV-2 satellite offers images at a spatial resolution of 0.5 m in the panchromatic (PAN) band and 2 m in the multispectral (MS) bands WV-2 MS image consists of four traditional spectral bands, consisting of Band 2, blue (450 to 510 nm); Band 3, green (510 to 580 nm); Band 5, red (630 to 690 nm); and Band

7, near-IR1 (NIR1) (770 to 895 nm), and four new bands: Band 1, coastal (400 to 450 nm); Band 4, yellow (585 to

625 nm); Band 6, red edge (705 to 745 nm); and Band 8, near-IR2 (NIR2) (860 to 1040 nm).The ground truth data sets utilized to support the semiautomatic extraction of water bodies were retrieved from the Australian Antarctic Data Centre (AADC), Indian Scientific Expedition to Antarctica (InSEA), historical Google Earth images, and PAN-sharpened WV-2 images A geodatabase consisting of 110 randomly distributed water bodies in the Larsemann Hills region was generated by manual digitization of PAN-sharpened WV-2 image After a careful visual analysis, only 36 water bodies (out of 110) extracted by manual digitization and confirmed by the field survey data were considered in the present analysis (Figure 1)

Fig 1 A satellite image map showing the spatial distribution of the 36 water bodies under consideration

3 Methods

The data processing protocol for water body feature extraction is shown in Figure 2 The protocol consists of 3 blocks: (a) data calibration, (b) PAN-sharpening and feature extraction, and (c) accuracy assessment Each block of the methodology (Figure 2) is sequentially discussed below

3.1 Data calibration

Calibrating image necessitates radiometric corrections, which require a mathematical function to transfer the digital number (DN) into the at-sensor radiance (Bakker et al., 2009) The calibration procedure was carried out in two steps: (1) converting raw DN values to at-sensor spectral radiance and, (2) converting spectral radiance to top-of-atmosphere reflectance (TOA)

Figure 2 Experimental procedure for semiautomatic extraction of water body information

3.2 PAN-sharpening and feature extraction

First, the 8-band multispectral image was PAN-sharpened from a spatial resolution of 2 m to 0.50 m using 4 most popular and effective algorithms for feature extraction These algorithms are Brovey’s Transform (BT), Hyperspherical Colour Sharpening (HCS), Ehler’s Fusion (EF), and Gram Schmidt (GS) (Jawak and Luis, 2013c; Jawak and Luis, 2013d) After PAN-sharpening, feature extraction procedure was carried out using 14 semiautomatic feature extraction methods categorized as: (a) modified NDWI methods, (b) spectral matching methods, and (c) target extraction methods In this study, modified NDWIs were experimentally estimated by cyclical and robust spectral profile observations in ENVI 5 We recognized the most useful spectral bands for differentiating water bodies through visual inspection The spectral bands were ranked for maximum and minimum response values for water bodies by using the minimum redundancy maximum relevance (mRMR) principle (Peng

et al., 2005) These bands were normalized to generate a modified NDWI for water body extraction (Table 1) To discriminate melt water on the snow surface and water bodies, manual thresholding was carried out Since the water body areas displays NDWI values from 0.60 to 0.92, a scene-dependent threshold was defined and used to discriminate between water and non-water pixels (Jawak and Luis, 2013a; Jawak and Luis, 2013b) After classifying the image by using each modified NDWI, 36 semiautomatically extracted water bodies were vectorized to calculate the area of individual water body The spectral matching methods were implemented using the Spectral Processing Exploitation and Analysis Resource (SPEAR) workflow tools (ENVI 5), which streamline spectral matching methods (MF, MTMF, SAM, MF/SAM ratio, and PCA) for extracting open and obscured water bodies using WV-2 PAN-sharpened data Target detection tools (ENVI 5) were executed for supervised image processing methods into workflows (CEM, ACE, OSP, TCIMF and MT-TCIMF) to extract water bodies

Table 1 Modified NDWIs for effective water body feature extraction ("avg" denotes "average")

3.3 Accuracy assessment:

All the water bodies extracted using the 14-feature extraction methods (4 modified + 10 existing) grouped into 3 classes were vectorized to calculate their areas We compared these calculated areas with that manually digitized reference water body, and we evaluated statistical significance based on the accuracy assessment A geodatabase of water body areas of the 36 extracted water bodies was utilized as reference for assessing the accuracy of the semiautomatic extraction methods Average error or average bias and percent bias between extracted water body area and reference digitized area were calculated For quantifying the uncertainty of our analysis, we calculated the

root mean square error (RMSE) using bias values

4 Results

We evaluated semiautomatic feature extraction methods for delineation of water bodies, which were categorized

as (a) the modified NDWI methods (average RMSE = 234.03 m2), (b) the spectral processing methods (average

RMSE = 243.70 m2), and (c) the target extraction methods (average RMSE = 226.18 m2) Our experiment revealed that existing target extraction approach outperformed the other approaches which employed 4 PAN-sharpening algorithms The modified NDWI approach was intermediate between the other 2 approaches The variation in RMSE between the 3 approaches is summarized in Table 2 The mNDWI3combination yielded superior results, while mNDWI4 produced inferior results, compared to the remaining 2 combinations in a group of 4 practiced NDWIs (Table 2, Average RMSE) On other hand, the differences in average RMSE for mNDWI2 (232.82 m2) and mNDWI1 (236.22 m2), seemed comparable in a given cohort The modified NDWI approach was found to be superior to the spectral matching approach and inferior to the target extraction approach The overall trend of performance of feature extraction methods, based on RMSE, are summarized in Figure 3 and can be ranked as follows: MT-TCIMF > ACE > mNDWI3 > MTMF > TCIMF > mNDWI2 > CEM > mNDWI1 > PCA > mNDWI4

> OSP > MF/SAM > SAM > MF This order suggests that the methods can be ranked as: target extraction approach

> modified NDWI approach > spectral processing approach The average RMSE values for the modified NDWI methods ranged from ~226 to ~240 m2, (bias ~1 to ~34 m2) and for the spectral matching methods and the target extraction methods, RMSE values varied from ~226 to 252 m2 (bias ~7 to ~37 m2) and from ~202 to ~240 m2, (bias

~8 to ~32 m2) respectively (Table 2) The smallest variation in RMSE observed in the modified NDWI approach suggests that it is a more stable and consistent approach than the other two approaches

In this study, we tested 4 PAN-sharpening algorithms (GS, HCS, EF, and BT) for fusing an MS image with PAN

An evaluation of the performance of each PAN-sharpening algorithm for water body extraction was undertaken, since the performance of these algorithms is spectrally and spatially dependent From Table 2, where average RMSE

is depicted, the performance hierarchy of PAN-sharpening algorithms for 14 extraction methods, was HCS > GS >

Modified NDWI Mathematical expression

mNDWI 1

ሼሾሺܤܽ݊݀ͳሻ െ ܽݒ݃ሺܤܽ݊݀͹ ൅ ܤܽ݊݀ͺሻሿሽ ሼሾሺܤܽ݊݀ͳሻ ൅ ܽݒ݃ሺܤܽ݊݀͹ ൅ ܤܽ݊݀ͺሻሿሽ

mNDWI 2

ሼሾሺܤܽ݊݀ʹሻ െ ܽݒ݃ሺܤܽ݊݀͹ ൅ ܤܽ݊݀ͺሻሿሽ ሼሾሺܤܽ݊݀ʹሻ ൅ ܽݒ݃ሺܤܽ݊݀͹ ൅ ܤܽ݊݀ͺሻሿሽ

mNDWI 3

ሼሾܽݒ݃ሺܤܽ݊݀ͳ ൅ ܤܽ݊݀ʹሻ െ ሺܤܽ݊݀͹ሻሿሽ ሼሾܽݒ݃ሺܤܽ݊݀ͳ ൅ ܤܽ݊݀ʹሻ ൅ ሺܤܽ݊݀͹ሻሿሽ

mNDWI 4

ሼሾܽݒ݃ሺܤܽ݊݀ͳ ൅ ܤܽ݊݀ʹሻ െ ሺܤܽ݊݀ͺሻሿሽ ሼሾܽݒ݃ሺܤܽ݊݀ͳ ൅ ܤܽ݊݀ʹሻ ൅ ሺܤܽ݊݀ͺሻሿሽ

EF > BT, which suggests that the HCS algorithm outperformed the others, given its small average RMSE The

HCS-sharpened water body features that were extracted using the 14 different methods are portrayed in Figure 4

Table 2 Statistical evaluation (Bias and RMSE, m2) of 14 method used for water body extraction across 4 PAN-sharpening algorithms The lowest values in each row (italics) and column (bold) are highlighted The row-wise average RMSE, column-wise average RMSE, and local RMSE averages are bolded and underlined

Figure 3 The overall performance trend for all 14 water body extraction methods, in terms of RMSE

Method PAN-sharpening methods RMSE Bias

Spectral Matching

Target Extraction

Total average (RMSE) m2 227.18 245.39 226.13 240.03

Figure 4 A sample of 4 extracted water bodies from a HCS-sharpened image using the 14 extraction methods

The GS-sharpened image performed best for the mNDWI1 method (RMSE = 199.47 m2), while HCS showed the best performance for mNDWI2 (RMSE = 222.65 m2), mNDWI3 (RMSE = 175.24 m2), and mNDWI4 (RMSE = 207.75 m2) In general, HCS and GS are the optimal sharpening algorithms for the modified NDWI methods, while

EF is superior for target extraction methods

5 Discussion

A robust, accurate, and user-friendly method can extremely reduce the laborious manual digitization A semiautomatic water body feature extraction can be applied in an operational environment only if it provides better performance in terms of following quality measures, which are the advantages of our modified NDWI method

x Accuracy: Extraction results should be correct and the geometric errors shall be minimized The result must be better or at least comparable to that from the manual digitizing The RMSE values of modified NDWI shows that our approach could extract water bodies accurately with respect to the manually digitized reference

x Visual comparison: The semiautomatic method should provide the extraction results which can be compared visually against the manual reference data Visual comparison also shows that all the 36 water bodies were detected in the NDWI images, and the boundaries of the extracted water bodies match the actual boundaries of the water bodies in the images or reference digitized data closely

x Water body size: An ideal water body extraction method should extract all sizes of water body features, including small ponds to large glacial lakes NDWI performed better than the other two approaches for extracting even small sized water body features

x Error: The variation in error should be least Our accuracy assessment shows that modified NDWI performed better than the other two methods for extracting water bodies

x Geometric errors: Semiautomatic method should produce the results with actual representation of shapes of water bodies, maintaining the integrity of shape Visual comparison shows that all the three approaches worked better for maintaining the integrity of shape

6 Conclusions

The spectrum characteristics of water bodies from WV-2 images were analyzed by using semiautomatic extraction capability involving a combination of advanced image processing methods The use of popular PAN-sharpening algorithms coupled with a modification of NDWI provided an effective tool to support semiautomatic extraction of Antarctic water body features The use of modified NDWI combinations derived by using the duplet set of Blue and NIR bands offered a precise means for extracting water body areas The magnitude of the spectral and spatial distortions induced by PAN-sharpening, influenced consequent water body extraction processing, and significantly influenced the final accuracy of the analysis Inclusion of the distinctive new spectral WV-2 bands offered a contextual foundation for surface water mapping by using feature extraction methods and scene characterization Different band combinations of NDWI provide a broad vision to resolve minor variations in spectrum properties of various water bodies and consequently the performance of the practiced extraction procedures This would have not been possible if we used other satellite data that contained only the single infrared,

red, and blue bands The different band combinations used in this study also facilitated a deep understanding of the role of specific spectral bands used in varied combinations to produce the best water body extraction from PAN-sharpened images

Acknowledgements

We thank the Australian Antarctic Data Centre for providing us with the supplementary GIS data layers for the study area We acknowledge Mr Parag Khopkar and Mr Tejas Godbole of University of Pune, for their assistance

in initial data processing We acknowledge Dr S Rajan, Director, NCAOR, for encouragement and motivation of this research This is NCAOR contribution No 35/2014

References

Bakker, H.W et al., 2009 Principles of Remote Sensing, The International Institute for GeoInformation Science and Earth Observation (ITC), Enschede

Chen, Q.L., Zhang, Y.Z., Ekroos, A., Hallikainen, M., 2004 The role of remote sensing technology in the EU water framework directive (WFD) Environmental Science and Policy 7, 267–276

Jawak, S.D., Luis, A.J., 2014a A semiautomatic extraction of Antarctic lake features using WorldView-2 imagery Photogrammetric

Engineering & Remote Sensing 80(10), 33-46

Jawak, S.D., Luis, A.J., 2014b Spectral Bands of WorldView-2 Satellite Remote Sensing Data for Semiautomatic Land Cover Extraction in the Antarctic Environment XXXIII SCAR and 6 th Open Science Conference, 25-28 August 2014, Auckland, New Zealand, August 2014

Jawak, S.D., Luis, A.J., Panditrao, S.N., Khopkar, P.S., Jadhav, P S., 2013 Advancement in landcover classification using very high resolution remotely sensed 8-band WorldView-2 satellite data International Journal of Earth Sciences and Engineering 06(02), 1742-1749

Jawak, S.D., Luis, A.J., 2013a Improved land cover mapping using high resolution multiangle 8-band WorldView-2 satellite remote sensing data Journal of Applied Remote Sensing 7(1), 073573

Jawak, S.D., Luis, A.J., 2013b A spectral index ratio-based Antarctic land-cover mapping using hyperspatial 8-band WorldView-2 imagery Polar Science 7(1), 18–38

Jawak, S.D., Luis, A.J., 2013c A comprehensive evaluation of PAN-sharpening algorithms coupled with resampling methods for image synthesis

of very high resolution remotely sensed satellite data Advances in Remote Sensing 2(4), 332-344

Jawak, S.D., Luis, A.J., 2013d Very-high resolution remotely sensed satellite data for improved land cover extraction of Larsemann Hills, east Antarctica Journal of Applied Remote Sensing 0001;7(1), 073460

Jawak, S.D., Luis, A.J., 2012 WorldView-2 satellite remote sensing data for polar geospatial information mining of Larsemann Hills, East Antarctica, Proceedings of 11 th Pacific (Pan) Ocean Remote Sensing Conference (PORSEC), Id: PORSEC2012-24-00006, Kochi, Kerala, India, 05-09 November 2012

Jawak, S.D., Mathew, J., 2011 Semi-automatic extraction of water bodies and roads from high resolution QuickBird satellite data, Geospatial World Forum Hyderabad, India, paper# 263, 247-257

Jawak, S.D., Luis, A.J., 2011 Applications of WorldView-2 satellite data for extraction of polar spatial information and DEM of Larsemann Hills, East Antarctica, International Conference on Fuzzy Systems and Neural Computing (FSNC 2011), Volume- 3, 978-1-4244-9216-9/11/ ©2011 IEEE, pp 148-151, 20-21 February, Hong Kong, China

Lacaux, J.P., Tourre, Y.M., Vignolles, C., Ndione, J.A., Lafaye, M., 2007 Classification of ponds from high-spatial resolution remote sensing: Application to Rift Valley Fever epidemics in Senegal, Remote Sensing of Environment 106(1), 66–74

Lu, S., Wu, B., Yan, N., Wang, H Water body mapping method with HJ-1A/B satellite imagery International Journal of Applied Earth Observation and Geoinformation 13(3), 428-434

Li, B., Zhou, X.Z Research on extraction method of river water-body in mountain area of western Sichuan based on MODIS data Geogr Geo-Inform Sci 25, 59–62

McFeeters, S.K., 1996 The use of the Normalized Difference Water Index (NDWI) in the delineation of open water features International Journal of Remote Sensing 17, 1425–1432

Ouma, Y.O., Tateishi, R., 2006 A water index for rapid mapping of shoreline changes of five East African Rift Valley lakes: an empirical analysis using Landsat TM and ETM+ data Int J Rem Sens 27, 3153–3181

Peng, H., Long, F., Ding, C., 2005 Feature selection based on mutual information: criteria of max-dependency, max-relevance and min-redundancy IEEE Trans Pattern Anal Mach Intel 27(8), 1667–1238

Papa, F., Prigent, C., Rossow, W.B., 2008 Monitoring flood and discharge variations in the large Siberian rivers from a multi-satellite technique Surveys in Geophysics 29, 297–317

Rogers, A.S., Kearney, M.S., 2004 Reducing signature variability in unmixing coastal marsh Thematic Mapper scenes using spectral indices Int

J Rem Sens 20, 2317–2335

Soti, V., Tran, A., Bailly, J., Puech, C , Seen, D.L., Bégué A., 2009 Assessing optical earth observation systems for mapping and monitoring temporary ponds in arid areas Int J Appl Earth Obs 11, 344–351