Improved machine-learning-based open-water–sea-ice–cloud discrimination over wintertime Antarctic sea ice using MODIS thermal-infrared imagery

Abstract. The frequent presence of cloud cover in polar regions limits the use of the Moderate Resolution Imaging Spectroradiometer (MODIS) and similar instruments for the investigation and monitoring of sea-ice polynyas compared to passive-microwave-based sensors. The very low thermal contrast between present clouds and the sea-ice surface in combination with the lack of available visible and near-infrared channels during polar nighttime results in deficiencies in the MODIS cloud mask and dependent MODIS data products. This leads to frequent misclassifications of (i) present clouds as sea ice or open water (false negative) and (ii) open-water and/or thin-ice areas as clouds (false positive), which results in an underestimation of actual polynya area and subsequently derived information. Here, we present a novel machine-learning-based approach using a deep neural network that is able to reliably discriminate between clouds, sea-ice, and open-water and/or thin-ice areas in a given swath solely from thermal-infrared MODIS channels and derived additional information. Compared to the reference MODIS sea-ice product for the year 2017, our data result in an overall increase of 20 % in annual swath-based coverage for the Brunt Ice Shelf polynya, attributed to an improved cloud-cover discrimination and the reduction of false-positive classifications. At the same time, the mean annual polynya area decreases by 44 % through the reduction of false-negative classifications of warm clouds as thin ice. Additionally, higher spatial coverage results in an overall better subdaily representation of thin-ice conditions that cannot be reconstructed with current state-of-the-art cloud-cover compensation methods.

Download Full-text

Novel machine-learning based cloud mask and its application for Antarctic polynya monitoring using MODIS thermal-infrared imagery

10.5194/egusphere-egu21-9603 ◽

2021 ◽

Author(s):

Stephan Paul ◽

Marcus Huntemann

Keyword(s):

Machine Learning ◽

Sea Ice ◽

False Positive ◽

Cloud Cover ◽

Near Infrared ◽

False Negative ◽

Open Water ◽

Thermal Infrared ◽

Spatial Coverage ◽

Cloud Mask

The frequent presence of cloud cover in polar regions limits the use of the Moderate-Resolution Imageing Spectroradiometer (MODIS) and similar instruments for the investigation and monitoring of sea-ice polynyas compared to passive-microwave-based sensors. The very low thermal contrast between present clouds and the sea-ice surface in combination with the lack of available visible and near-infrared channels during polar nighttime results in deficiencies in the MODIS cloud mask and dependent MODIS data products. This leads to frequent misclassifications of i) present clouds as sea ice/open water (false-negative) and ii) open-water/thin-ice areas as clouds (false-positive), which results in an underestimation of actual polynya area and subsequent derived information. Here, we present a novel machine-learning based approach using a deep neural network that is able to reliably discriminate between clouds, sea-ice, and open-water/thin-ice areas in a given swath solely from thermal-infrared MODIS channels and derived additional information. Compared to the reference MODIS sea-ice product for the year 2017, our data results in an overall increase of 20% in annual swath-based coverage for the Brunt Ice Shelf polynya, attributed to an improved cloud-cover discrimination and the reduction of false-positive classifications. At the same time, the mean annual polynya area decreases by 44% through the reduction of false-negative classifications of warm clouds as thin ice. Additionally, higher spatial coverage results in an overall better sub-daily representation of thin-ice conditions that cannot be reconstructed with current state-of-the-art cloud-cover compensation methods.

Download Full-text

Improved machine-learning based open-water/sea-ice/cloud discrimination over wintertime Antarctic sea ice using MODIS thermal-infrared imagery

10.5194/tc-2020-159 ◽

2020 ◽

Cited By ~ 1

Author(s):

Stephan Paul ◽

Marcus Huntemann

Keyword(s):

Machine Learning ◽

Sea Ice ◽

Cloud Cover ◽

Near Infrared ◽

Open Water ◽

Thermal Infrared ◽

Polar Regions ◽

Antarctic Sea Ice ◽

Spatial Coverage ◽

Thermal Infrared Imagery

Abstract. The frequent presence of cloud cover in polar regions limits the use of the Moderate-Resolution Imageing Spectroradiometer (MODIS) and similar instruments for the investigation and monitoring of sea-ice polynyas compared to passive-microwave-based sensors. The very low thermal contrast between present clouds and the sea-ice surface in combination with the lack of available visible and near-infrared channels during polar nighttime results in deficiencies in the MODIS cloud mask and dependent MODIS data products. This leads to frequent misclassifications of i) present clouds as sea ice and ii) open-water/thin-ice areas as clouds, which results in an underestimation of polynya area and subsequently derived information. Here, we present a novel machine-learning based approach using a deep neural network that is able to reliably discriminate between clouds, sea-ice, and open-water/thin-ice areas in a given swath solely from thermal-infrared MODIS channels and additionally derived information. Compared to the reference MODIS sea-ice product, our data results in an overall increase of 31 % in annual swath-based coverage, attributed to an improved cloud-cover discrimination. Overall, higher spatial coverage results in a better sub-daily representation of thin-ice conditions that cannot be reconstructed with current state-of-the-art cloud-cover compensation methods.

Download Full-text

Combined SMAP–SMOS thin sea ice thickness retrieval

The Cryosphere ◽

10.5194/tc-13-675-2019 ◽

2019 ◽

Vol 13 (2) ◽

pp. 675-691 ◽

Cited By ~ 11

Author(s):

Cătălin Paţilea ◽

Georg Heygster ◽

Marcus Huntemann ◽

Gunnar Spreen

Keyword(s):

Soil Moisture ◽

Sea Ice ◽

Brightness Temperature ◽

Incidence Angle ◽

Data Availability ◽

Ice Thickness ◽

Polar Regions ◽

Sea Ice Thickness ◽

Brightness Temperatures ◽

Spatial Coverage

Abstract. The spaceborne passive microwave sensors Soil Moisture Ocean Salinity (SMOS) and Soil Moisture Active Passive (SMAP) provide brightness temperature data in the L band (1.4 GHz). At this low frequency the atmosphere is close to transparent and in polar regions the thickness of thin sea ice can be derived. SMOS measurements cover a large incidence angle range, whereas SMAP observes at a fixed 40∘ incidence angle. By using brightness temperatures at a fixed incidence angle obtained directly (SMAP), or through interpolation (SMOS), thin sea ice thickness retrieval is more consistent as the incidence angle effects do not have to be taken into account. Here we transfer a retrieval algorithm for the thickness of thin sea ice (up to 50 cm) from SMOS data at 40 to 50∘ incidence angle to the fixed incidence angle of SMAP. The SMOS brightness temperatures (TBs) at a given incidence angle are estimated using empirical fit functions. SMAP TBs are calibrated to SMOS to provide a merged SMOS–SMAP sea ice thickness product. The new merged SMOS–SMAP thin ice thickness product was improved upon in several ways compared to previous thin ice thickness retrievals. (i) The combined product provides a better temporal and spatial coverage of the polar regions due to the usage of two sensors. (ii) The radio frequency interference (RFI) filtering method was improved, which results in higher data availability over both ocean and sea ice areas. (iii) For the intercalibration between SMOS and SMAP brightness temperatures the root mean square difference (RMSD) was reduced by 30 % relative to a prior attempt. (iv) The algorithm presented here allows also for separate retrieval from any of the two sensors, which makes the ice thickness dataset more resistant against failure of one of the sensors. A new way to estimate the uncertainty of ice thickness retrieval was implemented, which is based on the brightness temperature sensitivities.

Download Full-text

The 2018 North Greenland polynya observed by a newly introduced merged optical and passive microwave sea-ice concentration dataset

The Cryosphere ◽

10.5194/tc-13-2051-2019 ◽

2019 ◽

Vol 13 (7) ◽

pp. 2051-2073 ◽

Cited By ~ 6

Author(s):

Valentin Ludwig ◽

Gunnar Spreen ◽

Christian Haas ◽

Larysa Istomina ◽

Frank Kauker ◽

...

Keyword(s):

Sea Ice ◽

Growth Yield ◽

Thermal Infrared ◽

Passive Microwave ◽

Pressure System ◽

Sea Ice Concentration ◽

Sea Ice Thickness ◽

Ice Growth ◽

Spatial Coverage ◽

Ice Concentration

Abstract. Observations of sea-ice concentration are available from satellites year-round and almost weather-independently using passive microwave radiometers at resolutions down to 5 km. Thermal infrared radiometers provide data with a resolution of 1 km but only under cloud-free conditions. We use the best of the two satellite measurements and merge thermal infrared and passive microwave sea-ice concentrations. This yields a merged sea-ice concentration product combining the gap-free spatial coverage of the passive microwave sea-ice concentration and the 1 km resolution of the thermal infrared sea-ice concentration. The benefit of the merged product is demonstrated by observations of a polynya which opened north of Greenland in February 2018. We find that the merged sea-ice concentration product resolves leads at sea-ice concentrations between 60 % and 90 %. They are not resolved by the coarser passive microwave sea-ice concentration product. The benefit of the merged product is most pronounced during the formation of the polynya. Next, the environmental conditions during the polynya event are analysed. The polynya was caused by unusual southerly winds during which the sea ice drifted northward instead of southward as usual. The daily displacement was 50 % stronger than normal. The polynya was associated with a warm-air intrusion caused by a high-pressure system over the Eurasian Arctic. Surface air temperatures were slightly below 0 ∘C and thus more than 20 ∘C higher than normal. Two estimates of thermodynamic sea-ice growth yield sea-ice thicknesses of 60 and 65 cm at the end of March in the area opened by the polynya. This differed from airborne sea-ice thickness measurements, indicating that sea-ice growth processes in the polynya are complicated by rafting and ridging. A sea-ice volume of 33 km3 was produced thermodynamically.

Download Full-text

Limitations imposed by cloud cover on multitemporal visible band satellite data sets from polar regions

Annals of Glaciology ◽

10.1017/s0260305500012696 ◽

1993 ◽

Vol 17 ◽

pp. 113-120 ◽

Cited By ~ 7

Author(s):

G.J. Marshall ◽

W.G. Rees ◽

J.A. Dowdeswell

Keyword(s):

Cloud Cover ◽

Near Infrared ◽

Meteorological Data ◽

Regional Climate ◽

Data Sets ◽

Polar Regions ◽

Ice Caps ◽

Low Cloud ◽

Multitemporal Data ◽

Snow And Ice

The study of surface features and processes on glaciers, ice caps and ice sheets often requires multitemporal data in order to take account of the dynamic behaviour of snow and ice. Satellite remote sensing provides an important means of acquiring such data sets from polar regions, but the effectiveness of visible (VIS) and near-infrared (NIR) wavebands is severely limited by the frequent presence of cloud cover. As well as preventing direct imaging of the surface, cloud cover can be difficult to discriminate from snow and ice because of its similarly high reflectivity. This paper describes a detailed quantitative study of the limitations imposed by cloud cover. The number of potentially useful images which could have been acquired from a study area in Svalbard for the April-September period for 1980–89 is determined. Meteorological data from two stations analyzed over this study period show that fewer than 10% of the available satellite overpasses occurred during cloud-free periods. A cloud cover level of three octas or less, used to define high quality imagery, occurred 29% and 32% of the time at the two stations. Markedly seasonal variations are found which can be explained in terms of the regional climate such that during the months August and September the probability of obtaining a cloud-free Landsat image every year is effectively zero. Analysis of several years of Landsat Quick Look images for other regions of Svalbard confirms these findings. However, a similar series for the Scoresby Sund region of East Greenland shows a much higher percentage of low cloud scenes and no marked seasonality.

Download Full-text

Surface Albedo of the Antarctic Sea Ice Zone

Journal of Climate ◽

10.1175/jcli3489.1 ◽

2005 ◽

Vol 18 (17) ◽

pp. 3606-3622 ◽

Cited By ~ 124

Author(s):

Richard E. Brandt ◽

Stephen G. Warren ◽

Anthony P. Worby ◽

Thomas C. Grenfell

Keyword(s):

Sea Ice ◽

Surface Albedo ◽

Near Infrared ◽

Field Experiments ◽

Climate Models ◽

Solar Spectrum ◽

Open Water ◽

Visual Observation ◽

Antarctic Sea Ice ◽

The Antarctic

Abstract In three ship-based field experiments, spectral albedos were measured at ultraviolet, visible, and near-infrared wavelengths for open water, grease ice, nilas, young “grey” ice, young grey-white ice, and first-year ice, both with and without snow cover. From the spectral measurements, broadband albedos are computed for clear and cloudy sky, for the total solar spectrum as well as for visible and near-infrared bands used in climate models, and for Advanced Very High Resolution Radiometer (AVHRR) solar channels. The all-wave albedos vary from 0.07 for open water to 0.87 for thick snow-covered ice under cloud. The frequency distribution of ice types and snow coverage in all seasons is available from the project on Antarctic Sea Ice Processes and Climate (ASPeCt). The ASPeCt dataset contains routine hourly visual observations of sea ice from research and supply ships of several nations using a standard protocol. Ten thousand of these observations, separated by a minimum of 6 nautical miles along voyage tracks, are used together with the measured albedos for each ice type to assign an albedo to each visual observation, resulting in “ice-only” albedos as a function of latitude for each of five longitudinal sectors around Antarctica, for each of the four seasons. These ice albedos are combined with 13 yr of ice concentration estimates from satellite passive microwave measurements to obtain the geographical and seasonal variation of average surface albedo. Most of the Antarctic sea ice is snow covered, even in summer, so the main determinant of area-averaged albedo is the fraction of open water within the pack.

Download Full-text

Extraction of Sea Ice Cover by Sentinel-1 SAR Based on SVM with Unsupervised Generation of Training Data

10.20944/preprints202005.0336.v1 ◽

2020 ◽

Author(s):

Xiaoming Li ◽

Yan Sun ◽

Qiang Zhang

Keyword(s):

Machine Learning ◽

Sea Ice ◽

Learning Algorithm ◽

Texture Features ◽

Open Water ◽

Ice Cover ◽

Training Data ◽

Support Vector ◽

Training Samples

In this paper, we focus on developing a novel method to extract sea ice cover (i.e., discrimination/classification of sea ice and open water) using Sentinel-1 (S1) cross-polarization (vertical-horizontal, VH or horizontal-vertical, HV) data in extra wide (EW) swath mode based on the machine learning algorithm support vector machine (SVM). The classification basis includes the S1 radar backscatter coefficients and texture features that are calculated from S1 data using the gray level co-occurrence matrix (GLCM). Different from previous methods where appropriate samples are manually selected to train the SVM to classify sea ice and open water, we proposed a method of unsupervised generation of the training samples based on two GLCM texture features, i.e. entropy and homogeneity, that have contrasting characteristics on sea ice and open water. We eliminate the most uncertainty of selecting training samples in machine learning and achieve automatic classification of sea ice and open water by using S1 EW data. The comparison shows good agreement between the SAR-derived sea ice cover using the proposed method and a visual inspection, of which the accuracy reaches approximately 90% - 95% based on a few cases. Besides this, compared with the analyzed sea ice cover data Ice Mapping System (IMS) based on 728 S1 EW images, the accuracy of extracted sea ice cover by using S1 data is more than 80%.

Download Full-text

Limitations imposed by cloud cover on multitemporal visible band satellite data sets from polar regions

Annals of Glaciology ◽

10.3189/s0260305500012696 ◽

1993 ◽

Vol 17 ◽

pp. 113-120 ◽

Cited By ~ 1

Author(s):

G.J. Marshall ◽

W.G. Rees ◽

J.A. Dowdeswell

Keyword(s):

Cloud Cover ◽

Near Infrared ◽

Meteorological Data ◽

Regional Climate ◽

Data Sets ◽

Polar Regions ◽

Ice Caps ◽

Low Cloud ◽

Multitemporal Data ◽

Snow And Ice

Download Full-text

Radiometrie measurements of sea-ice surface temperature in East Antarctica

Annals of Glaciology ◽

10.3189/1998aog27-1-466-470 ◽

1998 ◽

Vol 27 ◽

pp. 466-470

Author(s):

Kelvin J. Michael ◽

Clemente S. Hungria ◽

R. A. Massom

Keyword(s):

Sea Ice ◽

Surface Temperature ◽

Snow Cover ◽

Open Water ◽

Thermal Infrared ◽

Near Surface ◽

Video Footage ◽

Antarctic Sea Ice ◽

Ice Surface ◽

Ice Conditions

This paper presents surface temperature data collected over East Antarctic sea ice by two thermal infrared radiometers mounted on the RSV Aurora Australis in March-May 1993. Operating at wavelengths equivalent to those utilised by channels 4 and 5 of AVHRR and similar channels of ATSR, the radiometers provided high-reso-lution data on surface (skin) temperature along the ship track. Additional information on the sea-ice conditions was obtained from hourly observations made from The ship's bridge, video footage and direct measurements made at ice stations. Following calibration, time series of temperatures from each of the radiometers were compared wi th ice-surface and near-surface air temperatures. Observed changes in the surface temperature are related to different snow and ice conditions. For a given air temperature, the surface temperature depends upon the thickness of ice and its snow cover. While open water areas (leads) have temperatures near -2.0°C, thick ice is characterised by surface temperatures which approximate those of the air. Taken as a whole, the along-track profile of surface temperature provides a proxy estimate of The proportion of open water and thin ice with in the pack. The presence of a snow cover has a significant effect on the surface temperature. It is anticipated that the results will be of use in the validation of sea-ice models and satellite thermal infrared data.

Download Full-text

MACHINE LEARNING FOR SEA ICE MONITORING FROM SATELLITES

ISPRS - International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences ◽

10.5194/isprs-archives-xlii-2-w16-83-2019 ◽

2019 ◽

Vol XLII-2/W16 ◽

pp. 83-89

Author(s):

C. O. Dumitru ◽

V. Andrei ◽

G. Schwarz ◽

M. Datcu

Keyword(s):

Machine Learning ◽

Sea Ice ◽

Weather Conditions ◽

Representation Learning ◽

Image Understanding ◽

Machine Learning Techniques ◽

Polar Regions ◽

Learning Approaches ◽

Automated Processing ◽

Modern Machine

Abstract. Today, radar imaging from space allows continuous and wide-area sea ice monitoring under nearly all weather conditions. To this end, we applied modern machine learning techniques to produce ice-describing semantic maps of the polar regions of the Earth. Time series of these maps can then be exploited for local and regional change maps of selected areas. What we expect, however, are fully-automated unsupervised routine classifications of sea ice regions that are needed for the rapid and reliable monitoring of shipping routes, drifting and disintegrating icebergs, snowfall and melting on ice, and other dynamic climate change indicators. Therefore, we designed and implemented an automated processing chain that analyses and interprets the specific ice-related content of high-resolution synthetic aperture radar (SAR) images. We trained this system with selected images covering various use cases allowing us to interpret these images with modern machine learning approaches. In the following, we describe a system comprising representation learning, variational inference, and auto-encoders. Test runs have already demonstrated its usefulness and stability that can pave the way towards future artificial intelligence systems extending, for instance, the current capabilities of traditional image analysis by including content-related image understanding.

Download Full-text