Resumen de Deep convolutional neural networks for statistical downscaling of climate change projections
- español
Las proyecciones climáticas a escala local y/o regional son muy demandadas por diversos sectores socioeconómicos para elaborar sus planes de adaptación y mitigación al cambio climático. Sin embargo, los modelos climáticos globales actuales presentan una resolución espacial muy baja, lo que dificulta enormemente la elaboración de este tipo de estudios. Una manera de aumentar esta resolución es establecer relaciones estadísticas entre la variable local de interés (por ejemplo la temperatura y/o precipitación en una localidad dada) y un conjunto de variables de larga escala (por ejemplo, geopotencial y/o vientos en distintos niveles verticales) dadas por los modelos climáticos. En particular, en esta Tesis se explora la idoneidad de las redes neuronales de convolución (CNN) como método de downscaling estadístico para generar proyecciones de cambio climático a alta resolución sobre Europa. Para ello se evalúa primero la capacidad de estos modelos para reproducir la variabilidad local de precipitación y de temperatura en un período histórico reciente, comparándolas contra otros métodos estadísticos de referencia. A continuación, se analiza la idoneidad de estos modelos para regionalizar las proyecciones climáticas en el futuro (hasta el año 2100). Además, se desarrollan diversos estudios de interpretabilidad sobre redes neuronales para ganar confianza y conocimiento sobre el uso de este tipo de técnicas para aplicaciones climáticas, puesto que a menudo son rechazadas por ser consideradas “cajas negras”.
- English
Regional climate projections are very demanded by different socioeconomics sectors to elaborate their adaptation and mitigation plans to climate change. Nevertheless, the state-of-the-art Global Glimate Models (GCMs) present very coarse spatial resolutions what limits their use in most of practical applications and impact studies. One way to increase this limited spatial resolution is to establish empirical/statistical functions which link the local variable of interest (e.g. temperature and/or precipitation at a given site) with a set of large-scale atmospheric variables (e.g. geopotential and/or winds at different vertical levels), which are typically well-reproduced by GCMs. This is called statistical downscaling [1]. To date, a variety of techniques which include (generalized) linear models [2], analogs [3], support vector machines [4], random forests [5] and shallow neural networks [6] have been employed to establish the link between the large and the local-scale. Despite their several merits, none of these methods has the capability to automatically handle high-dimensional input spaces without leading to overfitting, reason why the predictor space needs to undergo tedious and “human-guided" feature selection and/or reduction procedures before entering the statistical model for training/calibration/fit [2,7,8]. This typically implies a loss of information which can be relevant to explain the local variability of the target variable.
In this context, this Thesis explores the suitability of deep learning [9], and in particular modern Convolutional Neural Networks (CNNs, [10]), as statistical downscaling techniques to produce regional climate change projections over Europe. To achieve this ambitious goal, the capacity of CNNs to reproduce the local variability of precipitation and temperature fields in present climate conditions is first assessed by comparing their performance with that from a set of traditional statistical methods [2]. The results indicate the ability of CNNs to deal with high-dimensional input spaces inherent in any downscaling application, and reproduce better the local variability than the generalized linear models used as benchmark. Subsequently, the suitability of CNNs to produce plausible future (up to 2100) high-resolution scenarios is put to the test by comparing their projected signals of change with those given by a set of state-of-the-art GCMs from CMIP5 [11] and Regional Climate Models (RCMs) from the flagship EURO-CORDEX initiative [12], which are used as “pseudo-reality” since there are no observations into the future [13]. As compared to classical statistical methods, the CNNs produce more plausible climate signals, however, some differences are observed between the CNN and GCM/RCM projections. Further research is needed to properly establish conclusions regarding the suitability of statistical methods to downscale global climate simulations. To this aim, a merit of this Thesis is the publication of a dataset, named DeepESD, which consists of an ensemble of regional precipitation and temperature projections (up to 2100) over Europe, and is expected to fasten the analysis of SD-based climate projections with views to a possible integration of these products in climate impact studies.
Also, a variety of interpretability techniques [14] are also carried out to gain confidence and knowledge on the use of CNNs for climate applications, which have typically discarded until now for being considered as "black-boxes".
[1] Douglas Maraun and Martin Widmann. Statistical Downscaling and Bias Correction for Climate Research. Cambridge University Press, January 2018 [2] J. M. Gutierrez, D. Maraun, M. Widmann, R. Huth, E. Hertig, R. Benestad, O. Roessler, J. Wibig, R. Wilcke, S. Kotlarski, D. San Martin, S. Herrera, J. Bedia, A. Casanueva, R. Manzanas, M. Iturbide, M. Vrac, M. Dubrovsky, J. Ribalaygua, J. Portoles, O. Raty,J. Raisanen, B. Hingray, D. Raynaud, M. J. Casado, P. Ramos, T. Zerenner, M. Turco, T. Bosshard, P. Stepanek, J. Bartholy, R. Pongracz, D. E. Keller, A. M. Fischer, R. M.Cardoso, P. M. M. Soares, B. Czernecki, and C. Page. An intercomparison of a large ensemble of statistical downscaling methods over Europe: Results from the VALUE perfect predictor cross-validation experiment. International Journal of Climatology, 39(9):3750–3785, 2019.
[3] Bruce C Hewitson and Robert George Crane. Climate downscaling: techniques and application. Climate Research, 7(2):85–95, 1996 [4] Shivam Tripathi, Srinivas Venkata, and Ravi Nanjundiah. Downscaling of Precipitation for Climate Change Scenarios: A Support Vector Machine Approach. Journal of Hydrology,330:621–640, November 2006 [5] Christopher Hutengs and Michael Vohland. Downscaling land surface temperatures at regional scales with random forest regression. Remote Sensing of Environment, 178:127–141,2016 [6] Dannell Quesada-Chacon, Klemens Barfus, and Christian Bernhofer. Climate change projections and extremes for Costa Rica using tailored predictors from CORDEX model output through statistical downscaling with artificial neural networks. International Journal of Climatology, 41(1):211–232, 2021 [7] D San-Martin, R Manzanas, Swen Brands, S Herrera, and Jose M Gutierrez. Reassessing model uncertainty for regional projections of precipitation with an ensemble of statistical downscaling methods. Journal of Climate, 30(1):203–223, 2017 [8] Jose M Gutierrez, D San-Martin, Swen Brands, R Manzanas, and S Herrera. Reassessing statistical downscaling techniques for their robust application under climate change conditions. Journal of Climate, 26(1):171–188, 2013 [9] Ian Goodfellow, Yoshua Bengio, and Aaron Courville. Deep Learning. MIT Press, 2016 [10] Yann LeCun, Yoshua Bengio, et al. Convolutional networks for images, speech, and timeseries. The handbook of brain theory and neural networks, 3361(10):1995, 1995 [11] Taylor, K. E., R. J. Stouffer, and G. A. Meehl, 2012: An overview of cmip5 and the experiment design. Bulletin of the American meteorological Society,93 (4), 485–498.
[12] Jacob, D., et al., 2020: Regional climate downscaling over Europe: perspectives from the EURO-CORDEX community. Regional Environmental Change,20 (2), 51 [13] Mathieu Vrac, ML Stein, K Hayhoe, and X-Z Liang. A general method for validating statistical downscaling methods under future climate change. Geophysical Research Letters,34(18), 2007.
[14] Zintgraf, L. M., T. S. Cohen, T. Adel, and M. Welling, 2017: Visualizing deep neural network decisions: Prediction difference analysis. arXiv preprint arXiv:1702.04595.
