Improving perceptual quality in spatiotemporal timeseries forecasting

Representing spatiotemporal time series data as image sequences and learning data-driven models that capture the dynamics of such data offers an efficient method for spatiotemporal forecasting. While pixel-wise metrics such as Mean Squared Error (MSE) are commonly used to train and evaluate these forecasting models, this study reveals that such distortion metrics struggle to capture localized variations, particularly when forecasting large spatial domains with high-frequency, irregular components leading to blurry predictions that average out temporal variations. This phenomenon is attributed to the uncertainty in the data and its relation to widely used loss functions, thus highlighting a tradeoff between distortion and perceptual quality metrics in spatiotemporal forecasting. We propose custom loss functions to control this tradeoff and evaluate it with perceptual metrics on irregular spatiotemporal timeseries forecasting tasks. We also show that this tradeoff can be implicitly affected through model design decisions. Based on forecasting applications in the transportation and weather domain, our experiments demonstrate that these approaches enhance the model's ability to capture the spatiotemporal dynamics represented in image timeseries datasets, in terms of the proposed perceptual metrics while maintaining competitive performance on traditional distortion measures such as MSE.