Incorporación de información temporal de precipitación en un método de downscaling basado en datos

Abstract

This study evaluates the impact of incorporating temporal information into deep learning models for the downscaling of daily precipitation in regions with complex topography, using satellite data as the primary information source. The objective is to compare the performance of purely spatial U-Net architectures with spatiotemporal architectures in terms of statistical agreement and their ability to represent hydrometeorological variability. To this end, three architectures were implemented: a two-dimensional convolutional U-Net (Conv2D) without temporal information, and two U-Net models that incorporate temporal information, one based on three-dimensional convolution (Conv3D) and the other combining two-dimensional convolution with Long Short-Term Memory (LSTM) operators along the temporal dimension. All models were trained under homogeneous configurations, and their performance was evaluated through spatial analyses of the mean, standard deviation, and error metrics. The results indicate that the inclusion of temporal information improves overall performance by reducing errors and increasing correlation with the reference data. The ConvLSTM2D architecture exhibited the best overall performance, particularly in the representation of extreme values. Nevertheless, the analysis suggests that performance gains do not depend solely on the inclusion of the temporal dimension, but also on how temporal information is integrated into the model architecture. In summary, spatiotemporal deep learning models show strong potential for satellite precipitation downscaling in mountainous regions and highlight the need to further explore strategies that optimize the use of temporal information.