Machine learning approaches to retrieve pan-Arctic melt ponds from visible satellite imagery

Melt ponds on sea ice play an important role in the seasonal evolution of the summer ice cover. In this study we present two machine learning algorithms, one (multi-layer neural network) for the retrieval of melt pond binary classification and another (multinomial logistic regression) for melt pond...

Full description

Bibliographic Details
Main Authors: Lee, S, Stroeve, J, Tsamados, M, Khan, AL
Format: Article in Journal/Newspaper
Language:English
Published: 2020
Subjects:
Online Access:https://discovery.ucl.ac.uk/id/eprint/10103879/1/RSE_lee_for_MT.pdf
https://discovery.ucl.ac.uk/id/eprint/10103879/
Description
Summary:Melt ponds on sea ice play an important role in the seasonal evolution of the summer ice cover. In this study we present two machine learning algorithms, one (multi-layer neural network) for the retrieval of melt pond binary classification and another (multinomial logistic regression) for melt pond fraction using moderate resolution visible satellite imagery from the Moderate Resolution Imaging Spectroradiometer (MODIS). To minimize the impact of the anisotropic reflectance characteristics of sea ice and melt ponds, normalized MODIS band reflectance differences from top-of-the-atmosphere (TOA) measured reflectances were used. The training samples for the machine learning were based on MODIS reflectances extracted for sea ice, melt ponds and open water classifications based on high resolution (~2 m) WorldView (WV) data. The accuracy assessment for melt pond binary classification and fraction is further evaluated against WV imagery, showing mean overall accuracy (85.5%), average mean difference (0.09), and mean RMSE (0.18). In addition to cross-validation with WV, retrieved melt pond data are validated against melt pond fractions from satellite and ship-based observations, showing average mean differences (MD), root-mean-square-error (RMSE), and correlation coefficients (R) of 0.05, 0.12, and 0.41, respectively. We further investigate a case study of the spectral characteristics of melt ponds and ice during refreezing, and demonstrate an approach to mask out refrozen pixels by using yearly maps of melt onset and freeze-up data together with ice surface temperatures (IST). Finally, an example of monthly mean pan-Arctic melt pond binary classification and fraction are shown for July 2001, 2004, 2007, 2010, 2013, 2016, and 2019. Bulk processing of the entire 20 years of MODIS data will provide the science community with a much needed pan-Arctic melt pond data set.