Precipitation downscaling for climate characterization with scarce data using transfer learning

High-resolution precipitation data is essential for applications such as agriculture, infrastructure planning, and climate risk assessment, particularly in regions with complex terrain. Conventional climate datasets, including satellite and reanalysis products, are often too coarse to capture local variability. Downscaling techniques refine these data, but traditional dynamical methods are computationally intensive, and statistical methods require long-term observations that are often unavailable. We present a deep learning-based downscaling framework using transfer learning. A base model is pre-trained on large datasets and then fine-tuned to regions with limited local data, reducing the need for extensive weather station networks. Applied to Ecuador, the method produces accurate long-term precipitation estimates with only a few years of observations. This approach is flexible, extendable to other climate variables, and suitable for downscaling future climate projections, demonstrating the potential of transfer learning to improve the accessibility and quality of high-resolution climate information in data-scarce, topographically complex regions.