Estimating Arctic sea ice surface roughness by using back propagation neural network

Abstract:

Sea ice surface roughness is an important diagnostic proxy that can be used to characterize sea ice type, age, thickness, and surface albedo. In this work we demonstrate an innovative method for mapping sea ice surface roughness from NASA’s Multi-angle Imaging SpectroRadiometer (MISR) at 275-m spatial resolution over the Arctic. We use a back propagation neural network (BPNN) to estimate sea ice surface roughness for each pixel of a MISR image. To calibrate the model we use lidar-derived roughness measurements from the Airborne Topographic Mapper (ATM). This machine-learning approach builds a direct relationship between multi-angular reflectance data from MISR and roughness data from ATM to estimate sea ice surface roughness. We take advantage of the red spectral band from all nine of MISR cameras and develop two separate models for summer and winter time, respectively to compare the performance of the models in capturing the sea ice surface properties in different seasons. The purpose of this paper is to verify the feasibility of the proposed method and to compare the final results of this method with previous sea ice surface roughness determinations.