Dissertation/Thesis Abstract

An Algorithm for Automated Satellite-Based River Ice Identification Using a Local Cloud Mask: Application over the Lower Susquehanna River Using VIIRS and MODIS
by Kraatz, Simon G., Ph.D., The City College of New York, 2017, 147; 10194309
Abstract (Summary)

Spatially detailed characterization of the distribution, amounts and timing of river ice is important for identifying and predicting potential ice hazards. Although information on ice cover over inland water bodies is provided within the VIIRS and MODIS snow products, this information has little practical value for river ice monitoring. First, rivers may not be properly resolved in the respective land/water masks. Second, cloud masks incorporated in the products can be too conservative or contain local errors which may potentially result in reduced effective area and temporal coverage.

This thesis presents an algorithm that identifies river ice as primary product. Notably, it also consists of a two-tiered cloud identification scheme that does not require information from cloud products for accurate cloud characterization. Generally, the developed cloud screening is more liberal than that of other masks, and a larger portion of reasonably cloud-free grid cells can be retrieved at a given time when compared to the MODIS or VIIRS cloud product. This approach avoids cloud masking differences due to snow/cloud confusions and makes it possible to produce ice maps that are highly consistent between the observing platforms. The algorithm is tested with MODIS-Terra (TR), MODIS-Aqua (AQ) and VIIRS (VR) between 2001 and 2016 over part of the Lower Susquehanna River. The algorithm features several unique characteristics and it (1) avoids potential cloud-product related errors by incorporating novel and robust spatiotemporal contrasting and k-means-like data clustering and as result generally samples from high-quality grid cells; (2) uses improved land/water as input, generated from on multi-spectral and multi-temporal data using maximum likelihood classifications to delineate the river channel; (3) produces binary ice maps, one for each confidence level defined by concurrent requirements that must be met by the visible and short wave infrared bands; (4) incorporates viewing geometry information incorporated within the MODIS daily reflectance products for better data clustering, further improving classification accuracy; (5) produces several auxiliary datasets, in particular at-grid ice cover durations, reflectance differences and time-series data that may be used in models.

The algorithm first estimates cloudiness via ”test 1” (T1) within the region, and if it fails no data is obtained for that day. If T1 conditions are met, river grid cells are classified as water, cloud and ice cover. Owing to this approach the algorithm observes 23% (MODIS-Terra), 45% (MODIS-Aqua) and 10% (VIIRS) more of the river on days T1 is passed. Although the algorithm frequently rejects data outright, an overall improvement in the amount of available data for MODIS-Aqua (39%) is realized in comparison to the incorporated cloud mask, which substantially reduces its effective revisit time. However, when applied to MODIS-Terra and VIIRS the algorithm has fewer overall data by 6% and 26%, respectively. The effective revisit times in days are 4.0 & 4.2 (MODIS-Terra), 5.2 & 3.7 (MODIS-Aqua) and 2.9 & 3.6 (VIIRS), for the cloud mask and algorithm, respectively. Daily effective revisit times may be further improved by compositing ice maps between the platforms, in particular between the AM and PM observations. While the sample involving VIIRS is limited to 2015 and 2016, using multiple platforms (i.e. TR vs TR/AQ, AQ vs AQ/VR and TR vs TR/VR) significantly increases the number of unique days the algorithm may observe by 37%, 23% and 42%, respectively.

Observations are highly consistent between the independently observing platforms. Comparisons of the normalized river ice cover fraction (RIF) and river ice amount (RIA) between platforms show correlations of 95% or more and mean absolute differences (MAD) near 5%, with the best agreement between MODIS-Aqua and VIIRS. Ice cover outputs were evaluated against the discharge data quality flag (DQF) at the USGS gauge, taken to suggest ice cover. While this may lead to occasional errors, river observations by traffic cameras for 2016 support that the DQF are quite accurate. Visual comparisons with Landsat 8 and the CRIOS river ice product also show good correspondence. RIF time series for the algorithm and its equivalent for CRIOS are nearly identical. The POD for ice and PC range from 87–91% and 91–94%. Despite the liberal cloud screening and inclusion of lower likelihood ice layers, errors are few. False detections range from 4–7% while non-detections range between 0–2%. The higher false detection rate is due to the chosen approach being more liberal with cloud screening, resulting in occasional misclassification of cloud as ice. Produced river ice maps are also consistent and in good agreement with traffic camera imagery.

Reflectance difference maps (ΔR) show some promise in helping distinguish between solid ice cover and mobile ice, and also generally indicate regions of ice accumulations. (Abstract shortened by ProQuest.)

Indexing (document details)
Advisor: Khanbilvardi, Reza
Commitee: Devineni, Naresh, Krakauer, Nir, Restrepo, Pedro, Romanov, Peter
School: The City College of New York
Department: Civil Engineering
School Location: United States -- New York
Source: DAI-B 79/08(E), Dissertation Abstracts International
Subjects: Hydrologic sciences, Civil engineering, Remote sensing
Keywords: Algorithm, Ice jam, MODIS, Susquehanna, VIIRS
Publication Number: 10194309
ISBN: 9780355768428
Copyright © 2019 ProQuest LLC. All rights reserved. Terms and Conditions Privacy Policy Cookie Policy