Deep Learning-Based Meteorological Data Downscaling: A Comparative Study with Physics Informed CNN and a Component-Level Ablation Analysis
Meteorological statistical downscaling is a critical technique for deriving high-resolution climate information from coarse global reanalysis products. This experiment investigates the application of five representative deep learning architectures for downscaling ERA5 reanalysis 2m temperature fields from 1° to 0.25° spatial resolution (4x upscaling factor) over the China region, spanning the period 2010 to 2020. The five baseline models evaluated are Convolutional Neural Network (CNN), Generative Adversarial Network (GAN), Long Short-Term Memory network (LSTM), Vision Transformer (ViT-style), and Denoising Diffusion Probabilistic Model (DDPM). Building upon the analysis of baseline strengths and weaknesses, a sixth model is proposed, the Physics-Informed CNN (PICNN), which integrates multi-scale feature extraction, spatial attention mechanisms, and a composite physics-informed loss function incorporating mean squared error, Laplacian spatial smoothness regularization, and spatial energy conservation constraints. All models were trained on an NVIDIA GeForce RTX 5050 Laptop GPU with 8GB VRAM under computational constraints, with training epochs reduced from default values to accommodate hardware limitations. Experimental results on the held-out 2020 test set demonstrate that the proposed PICNN achieves the best performance among all models, with RMSE of 0.665 K, MAE of 0.435 K, PSNR of 43.43 dB, and SSIM of 0.976, representing improvements of 7.7% in RMSE and 8.7% in MAE over the strongest CNN baseline. To determine whether this improvement reflects the architectural design or simply the 84% increase in parameter count relative to the CNN baseline, a controlled ablation study was conducted comprising a capacity-matched plain CNN and three component-ablated PICNN variants, evaluated across three random seeds for the ablated variants. The results show that approximately 65.5% of PICNN's improvement over the CNN baseline is explained by parameter count alone. Among PICNN's three architectural components, spatial attention accounts for the majority of the remaining gain, the physics-informed loss contributes a small but consistent improvement, and the multi-scale feature extraction head shows no measurable benefit, with its ablated variant statistically indistinguishable from the full model. These findings refine the paper's central claim from a general endorsement of physics-informed design toward a specific, evidence-backed identification of which architectural choices are responsible for the observed gains. The Transformer model, despite having 139 million parameters, performed significantly worse than the CNN, highlighting the data inefficiency of attention-based architectures on moderate-sized meteorological datasets. The LSTM performed worst overall due to its inherent inability to preserve spatial structure.