Evaluation of machine learning algorithms to classify and map landforms in Antarctica

• Published in: Earth Surface Processes and Landforms
• Main Authors: Siqueira, Rafael G., Veloso, Gustavo V., Fernandes‐Filho, Elpídio I., Francelino, Márcio R., Schaefer, Carlos Ernesto G. R., Corrêa, Guilherme R.
• Other Authors: Coordenação de Aperfeiçoamento de Pessoal de Nível Superior
• Format: Article in Journal/Newspaper
• Language: English
• Published: Wiley 2021
• Subjects: Antarc*; Antarctic; Antarctic Peninsula; Antarctica; Vega Island
• Online Access: http://dx.doi.org/10.1002/esp.5253; https://onlinelibrary.wiley.com/doi/pdf/10.1002/esp.5253; https://onlinelibrary.wiley.com/doi/full-xml/10.1002/esp.5253

Abstract The detailed geomorphology of ice‐free landscapes of Antarctica is key to understanding how their highly fragile environments respond to climate change, at different temporal and spatial scales. Despite the recent advances in geomorphological studies of ice‐free areas, machine learning applications to produce landform maps are still scarce on the Antarctic continent. In this study, we evaluated the predictive performance of different supervised machine learning algorithms to produce digital geomorphological maps in Vega Island—Antarctic Peninsula region. We tested six different models: average artificial neural networks, C5.0 decision tree, random forest, support vector machine, supervised self‐organizing map and weighted k ‐nearest neighbours. We used an initial set of 54 geomorphometric and spectral predictors, from which redundant variables with Pearson correlation coefficient >|0.95| were removed, and only the most important predictors for each model were selected using recursive feature elimination. For training, we ran each model 100 times and predictions were assessed by the kappa and global accuracy values. The best predictors were the Red Edge 6 and SWIR 11 bands, roughness concentration index, elevation and drainage density. The decision trees C5.0 and random forest had the best performance, with average validation kappa of 0.85 ± 0.03 and 0.84 ± 0.03, respectively, evidencing excellent prediction. Despite the similar performance, random forest showed greater uncertainty degree and accuracy when classifying complex landforms, attesting to its great robustness. From sensitivity and specificity values, we observed that the glaciers and talus showed higher accuracy, whereas cryoplanated platforms and scree slopes had the worst classification. The presented methodology optimized the classification by selecting the most important predictors, assessing accuracy and evaluating uncertainty. The results indicated that machine learning methods have great potential to produce geomorphological mappings ...