Estimation of area and volume change in the glaciers of the Columbia Icefield, Canada using machine learning algorithms and Landsat images

Abstract

Machine learning image classification algorithms offer a potential for effective and efficient classification of remotely sensed images covering glaciated areas. The Columbia Icefield in Canada is one such place where glaciers are retreating and losing mass over the years. The Columbia Icefield plays an important role in the region's water budget. In this study, the accuracy of three machine learning algorithms, namely, SVM, RF and MLC, were assessed for the classification of snow/ice area on 2020 Landsat 8 OLI image. All three algorithms classified the image with over 99 percent accuracies, but the SVM classifier showed a higher accuracy in debris covered areas on glaciers. Further, we used SVM algorithm to classify Landsat 5 TM - Sept 10, 1985, Landsat 5 TM - Sept 27, 1991, Landsat 8 OLI - Aug 22, 2013, and Landsat 8 OLI - Sept 10, 2020 images in the Columbia Icefield. Among nine glaciers, Saskatchewan (- 4.57 km2), Dome (−2.03 km2), Columbia (−2.06 km2), Stutfield (−2.17 km2), G242655E52112N (−1.39 km2), Athabasca (−1.39 km2), Castleguard (−1.3 km2), and G242614E52109N (−0.54 km2) measured less ice and snow-covered areas between 1985 and 2020. For these nine glaciers, there was a total decrease of 2.01 ± 0.24 km3 volume between 1985 and 2020, which is about 1.81 ± 0.22 km3 water equivalent or 0.12 ± 0.015 km3 water equivalent per year. On average, Saskatchewan (−0.699 km3) and Columbia (−0.307 km3) Glaciers lost the highest volume of snow and ice between 1985 and 2020. This study also concluded that all SVM, RF and MLC produce highly accurate satellite image classification in the glaciated areas.