Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

An Autoencoder Architecture for L-band Passive Microwave Retrieval of Landscape Freeze-Thaw Cycle

Divya Kumawat, Ardeshir Ebtehaj, Xiaolan Xu, Andreas Colliander, Vipin Kumar

TL;DR

This work addresses the challenge of mapping landscape freeze-thaw cycles over the Northern Hemisphere using L-band SMAP brightness temperatures. It introduces the FTC-Encoder, a supervised convolutional autoencoder trained with a contrastive loss to produce a probabilistic FT-state indicator p(F) by learning from peak winter (frozen) and peak summer (thawed) segments labeled via ERA5 data. The approach yields improved FT-state retrieval accuracy across Alaska’s diverse land-cover types and sub-grid lake ice scenarios, reducing thaw false detections relative to the SMAP NPR-based product. The method is robust to radiometric complexity from snow, vegetation, and lakes, and is validated against in-situ ISMN observations, highlighting potential for operational and large-scale applications with future multi-frequency extensions and global-scale testing.

Abstract

Estimating the landscape and soil freeze-thaw (FT) dynamics in the Northern Hemisphere is crucial for understanding permafrost response to global warming and changes in regional and global carbon budgets. A new framework is presented for surface FT-cycle retrievals using L-band microwave radiometry based on a deep convolutional autoencoder neural network. This framework defines the landscape FT-cycle retrieval as a time series anomaly detection problem considering the frozen states as normal and thawed states as anomalies. The autoencoder retrieves the FT-cycle probabilistically through supervised reconstruction of the brightness temperature (TB) time series using a contrastive loss function that minimizes (maximizes) the reconstruction error for the peak winter (summer). Using the data provided by the Soil Moisture Active Passive (SMAP) satellite, it is demonstrated that the framework learns to isolate the landscape FT states over different land surface types with varying complexities related to the radiometric characteristics of snow cover, lake-ice phenology, and vegetation canopy. The consistency of the retrievals is evaluated over Alaska, against in situ ground-based observations, showing reduced uncertainties compared to the traditional methods that use thresholding of the normalized polarization ratio.

An Autoencoder Architecture for L-band Passive Microwave Retrieval of Landscape Freeze-Thaw Cycle

TL;DR

This work addresses the challenge of mapping landscape freeze-thaw cycles over the Northern Hemisphere using L-band SMAP brightness temperatures. It introduces the FTC-Encoder, a supervised convolutional autoencoder trained with a contrastive loss to produce a probabilistic FT-state indicator p(F) by learning from peak winter (frozen) and peak summer (thawed) segments labeled via ERA5 data. The approach yields improved FT-state retrieval accuracy across Alaska’s diverse land-cover types and sub-grid lake ice scenarios, reducing thaw false detections relative to the SMAP NPR-based product. The method is robust to radiometric complexity from snow, vegetation, and lakes, and is validated against in-situ ISMN observations, highlighting potential for operational and large-scale applications with future multi-frequency extensions and global-scale testing.

Abstract

Estimating the landscape and soil freeze-thaw (FT) dynamics in the Northern Hemisphere is crucial for understanding permafrost response to global warming and changes in regional and global carbon budgets. A new framework is presented for surface FT-cycle retrievals using L-band microwave radiometry based on a deep convolutional autoencoder neural network. This framework defines the landscape FT-cycle retrieval as a time series anomaly detection problem considering the frozen states as normal and thawed states as anomalies. The autoencoder retrieves the FT-cycle probabilistically through supervised reconstruction of the brightness temperature (TB) time series using a contrastive loss function that minimizes (maximizes) the reconstruction error for the peak winter (summer). Using the data provided by the Soil Moisture Active Passive (SMAP) satellite, it is demonstrated that the framework learns to isolate the landscape FT states over different land surface types with varying complexities related to the radiometric characteristics of snow cover, lake-ice phenology, and vegetation canopy. The consistency of the retrievals is evaluated over Alaska, against in situ ground-based observations, showing reduced uncertainties compared to the traditional methods that use thresholding of the normalized polarization ratio.
Paper Structure (15 sections, 3 equations, 9 figures, 1 table)

This paper contains 15 sections, 3 equations, 9 figures, 1 table.

Table of Contents

  1. Introduction
  2. Data and Study Area
  3. SMAP data
  4. ERA5 land data sets
  5. In situ data
  6. Methodology
  7. Autoencoder Models and Loss Function
  8. Model Implementation and Training
  9. Results and Discussion
  10. Performance Statistics
  11. Time Series Analysis
  12. Spatiotemporal Assessments
  13. Time series analyses
  14. Spatial analyses
  15. Conclusion

Figures (9)

  • Figure 1: (a) The study area overlaid with IGBP land-cover types, water-fraction classes, the spatial distribution of 34 available ISMN stations; and (b) the probability density function (pdf) of SMAP TBs in 2018 at vertical polarization for all pixels with varying water fractions. The land-cover map represents open shrublands (OS), woody savannas (WS), savannas (S), grasslands (G), Permanent Ice/snow (SI), and Barren (B) for areas with water fraction less than 0.05.
  • Figure 2: A schematic of the FTC-Encoder architecture: A supervised convolutional autoencoder framework that can learn from the peak winter and summer TB time series with variable lengths to probabilistically determine the FT state.
  • Figure 3: Freeze and thaw retrieval accuracy for the ISMN network sites (Fig. \ref{['fig:01']}) grouped by land-cover classes and sub-pixel water fractions. The boxes represent the median and interquartile range while the whiskers extend the 5th and 95th percentiles.
  • Figure 4: Time series of (a,b) SMAP TBs at vertical (V-pol) and horizontal (H-pol) polarization channels with the FTC-Encoder frozen probability $p(F)$, (c,d) in-situ air and soil temperatures from ISMN ground-based stations as well as ERA5 snow depth, and (e,f) the NPR ratio together with SPL3FTP_E FT states over two SMAP pixels centered at (65.06°N, 146.71°W) and (63.92°N, 160.72°W) with sub-grid water fraction 0 (left column) and 15% (right column).
  • Figure 5: The Sentinel-2 false composite and true color imageries over the SMAP footprint centered at 60.3398°N and 162.02°W in the calendar year 2019 (first and second rows, respectively), (a) the SMAP TBs at horizontal (H) and vertical (V) polarization with $p(F)$, (b) snow depth, lake ice depth, ground and air temperature measurements obtained from the ERA5 reanalysis dataset and (c) the NPR ratio with the SPL3FTP_E freeze and thaw labels. The pink areas in panel (b) mark the ERA5 ground temperatures greater than 0° C.
  • ...and 4 more figures