SRCNN-Based High-Resolution Downscaling Forecasting of Arctic Sea Ice Concentration

Fine-scale downscaling forecasts of Arctic sea ice concentration are of strategic importance for ensuring navigational safety and supporting resource development. However, traditional approaches struggle to effectively characterize the complex mapping between low- and high-resolution images in regions lacking high-resolution remote sensing data, and they remain insufficient in handling extreme climate conditions. To address these limitations, this study proposes an innovative solution based on a Super-Resolution Convolutional Neural Network (SRCNN), leveraging 0.05 \(^\circ \) resolution OSTIA sea ice data (2015–2024, Chukchi Sea). The model employs a three-layer core architecture—patch extraction for spatial feature learning, nonlinear mapping for capturing complex relationships, and reconstruction for generating high-resolution outputs—thereby achieving an end-to-end mapping. Validation results demonstrate that, in the 2023–2024 test environment, the SRCNN consistently maintained mean squared error (MSE) below 0.001, peak signal-to-noise ratio (PSNR) up to 50 dB (average 35 dB), and structural similarity index (SSIM) above 0.98. Notably, during the rare ice-remnant event in the western Chukchi Sea in the summer of 2024, the model improved the accuracy of ice fracture detection and reduced land–sea boundary artifacts. This research provides a high-precision, low-cost, and operationally feasible forecasting tool for Arctic development, with future potential to integrate ConvLSTM models to enhance spatiotemporal prediction capabilities.