Fusion of ALOS-2/PALSAR-2 and Sentinel-1 Polarimetric Imagery with Explainable Neural Network-based Confidence Level Estimation for Glacier Surface Classification

Glacier monitoring and classification play a vital role in understanding the impact of climate change. Remote sensing imagery, particularly synthetic aperture radar (SAR) data, offers valuable information for glacier surface classification due to its all-weather and day-and-night imaging capability. This paper proposes an approach that combines dual-pol ALOS-2/PALSAR-2 and dual-pol Sentinel-1 polarimetric SAR imagery using the inverse-mapping explainable neural network for glacier surface classification. The findings indicate improved training and validation accuracies for the combination of the two SAR datasets as compared to their individual training and validation accuracies. This paper also presents a method for determining the confidence levels of the neural network in its own results. The generated confidence maps have the potential to provide valuable insights into the decision-making process of the neural network and can assist in the selection of suitable training inputs for improving the performance of the network.

I. Introduction

Glaciers play an important role in the Earth’s climate system, and their monitoring and analysis are critical to understanding the impacts of climate change [1], [2]. Over the years, the field of glacier monitoring has witnessed an increased use of SAR technology [3]. SAR sensors operate irrespective of sunlight and weather conditions, enabling continuous monitoring. This all-weather, all-day capability is particularly advantageous in high mountain areas, where persistent and dense cloud cover presents challenges to traditional optical remote sensing techniques [4]. In addition, microwave signals from SAR can penetrate the surface of glaciers and provide valuable information about their internal structure and dynamics [5].