Enhancing Snow Detection through Deep Learning: Evaluating CNN Performance Against Machine Learning and Unsupervised Classification Methods

Snow cover is a crucial natural water resource that provides water storage. Monitoring snow cover is essential for sustainable water resource management. Accurate and efficient snow cover mapping is vital for water resource planning, flood control, and climate impact assessments. Remote sensing provides a cost-effective approach for monitoring snow cover dynamics, but classification accuracy remains a challenge. This study introduces a deep learning-based 1D convolutional neural network (1D CNN) approach for improved snow cover classification in mountainous regions utilizing Landsat 5 and 8 satellite images. The effectiveness of 1D CNN was compared with unsupervised classification (UC) methods (Normalized Difference Snow Index (NDSI), band ratio-based Red/Short wave infrared (R/SWIR), and Near infrared/Short wave infrared (NIR/SWIR)) and supervised machine learning classification methods (support vector machines (SVM) and classification and regression trees (CART)) in this research. The study was performed on Erciyes Mountain, Turkey, across four different years. Classifier performances were assessed using various quantitative (precision, recall, F-score, overall accuracy, kappa) and qualitative (mean squared error, Pearson correlation coefficient, structural similarity index measure, peak signal-to-noise ratio) evaluation metrics. Pairwise statistical analysis using McNemar’s test revealed significant differences between 1D CNN and the other classifiers. The findings demonstrated that the 1D CNN approach outperformed all other methods in achieving the highest accuracy in snow cover classification. These results indicate the potential of deep learning for enhancing snow cover assessment and supporting more effective water resource management strategies and policy-making.