A Machine Learning Algorithm Improves Surface Freeze-Thaw Classification

• Main Author: Bunt, Fredrick
• Format: Thesis
• Language: unknown
• Published: University of Montana 2022
• Subjects: Machine Learning; Freeze-Thaw Classification; U-Net; FT-ESDR; Brightness Temperature; Data Science; permafrost
• Online Access: https://scholarworks.umt.edu/etd/11944; https://scholarworks.umt.edu/context/etd/article/13061/viewcontent/thesis_fred_bunt_w_cover.pdf

The frozen or thawed state of the land surface is an important factor affecting a wide range of natural processes such as surface water movement, the carbon cycle, and ecosystem development. It is also important for human endeavors such as permafrost engineering and agricultural planning. This makes having an accurate record important. The Freeze-Thaw (FT) Earth System Data Record (FT-ESDR) is a global, daily product that strives to be a reliable record of the FT ground state. In its current form, the FT-ESDR uses annual regression analysis of reanalysis surface air temperatures (SAT) and brightness temperatures (Tb) at each grid cell to produce a FT record. This has great accuracy (>85%) at middle latitudes and during the summer and winter seasons. Unfortunately, the FT-ESDR has degraded accuracy (<75%) in much of the polar regions as well as during the transitional seasons. The product is derived from the vertically polarized 37 GHz band of global Tb satellite retrievals. We present a new method for generating FT records over the Northern Hemisphere that uses all polarizations for the 19, 22, and 37 GHz Tb bands and a global elevation map. This method uses a fully convolutional neural network model for its classification. The neural network is trained using the Tb bands, elevation map, reanalysis SAT, and global automated weather station data from the 10 year period 1998–2007. The classifications are validated against the World Meteorological Organization's global automated weather station network's SAT record over the combined 20 year period of 1988–1997 and 2009–2020. Our new Northern Hemisphere product shows significantly improved classification accuracy (as much as 6.1% points) over the FT-ESDR record in both higher latitudes and the transitional seasons. The model that this method produces is much faster at generating prediction records for new data. It also has the advantage of producing probability maps along with the classification predictions.