Generative deep learning models are increasingly adopted for climate downscaling due to their ability to efficiently produce high-resolution outputs from coarse-resolution boundary conditions. However, unlike dynamical downscaling models, which explicitly represent physical mechanisms governing regional climate behavior, generative models remain largely opaque in terms of whether and how they reflect known or plausible regional physical processes. This lack of domain knowledge connection and understanding limits scientific trust and hinders the use of generative downscaling in risk-sensitive and decision-making situations. In this work, we present a visual analytics approach that enables domain experts to explore regional physical processes reflected in trained generative climate downscaling models by examining how spatially localized, multivariable input patterns relate to model outputs across cohorts of similar predictions.
WilbyRLWigleyTMLConwayD, et al. Statistical downscaling of general circulation model output: a comparison of methods. Water Resour Res1998; 34(11): 2995–3008.
4.
XuZHanYYangZ.Dynamical downscaling of regional climate: a review of methods and limitations. Sci China Earth Sci2019; 62(2): 365–375.
ChiangCHHuangZHLiuL, et al. Climate downscaling: a deep-learning based super-resolution model of precipitation data with attention block and skip connections. arXiv preprint, arXiv:240317847, 2024.
9.
WangFTianDLoweL, et al. Deep learning for daily precipitation and temperature downscaling. Water Resour Res2021; 57(4): e2020WR029308.
10.
KumarBAteyKSinghBB, et al. On the modern deep learning approaches for precipitation downscaling. Earth Sci Inform2023; 16(2): 1459–1472.
11.
WangFTianDCarrollM.Customized deep learning for precipitation bias correction and downscaling. Geosci Model Dev2023; 16: 535–556.
12.
HarrisLMcRaeATTChantryM, et al. A generative deep learning approach to stochastic downscaling of precipitation forecasts. J Adv Model Earth Syst2022; 14(10): e2022MS003120.
13.
Mosqueira-ReyEHernández-PereiraEAlonso-RíosD, et al. Human-in-the-loop machine learning: a state of the art. Artif Intell Rev2023; 56(4): 3005–3054.
14.
CastelvecchiD.Can we open the black box of AI?Nature News2016; 538(7623): 20–23.
15.
KimBWattenbergMGilmerJ, et al. Interpretability beyond feature attribution: quantitative testing with concept activation vectors (TCAV). In: International conference on machine learning, 2018, pp. 2668–2677. PMLR.
16.
WangYZhangTGuoX, et al. Gradient based feature attribution in explainable ai: a technical review. arXiv preprint arXiv:240310415, 2024.
BaldassarreFAzizpourH.Explainability techniques for graph convolutional networks. arXiv preprint, arXiv:190513686, 2019.
19.
CortezPEmbrechtsMJ.Using sensitivity analysis and visualization techniques to open black box data mining models. Inf Sci2013; 225: 1–17.
20.
SmilkovDThoratNKimB, et al. SmoothGrad: removing noise by adding noise. arXiv preprint, arXiv:170603825, 2017.
21.
BaehrensDSchroeterTHarmelingS, et al. How to explain individual classification decisions. J Mach Learn Res2010; 11: 1803–1831.
22.
FongRCVedaldiA.Interpretable explanations of black boxes by meaningful perturbation. In: Proceedings of the IEEE international conference on computer vision, 2017, pp. 3429–3437. IEEE.
23.
SelvarajuRRCogswellMDasA, et al. Grad-CAM: visual explanations from deep networks via gradient-based localization. In: Proceedings of the IEEE international conference on computer vision, 2019, pp. 618–626. IEEE.
24.
BinderAMontavonGLapuschkinS, et al. Layer-wise relevance propagation for neural networks with local renormalization layers. In: International conference on artificial neural networks, 2016, pp. 63–71. Springer.
25.
ShrikumarAGreensidePKundajeA.Learning important features through propagating activation differences. In: International conference on machine learning, 2017, pp. 3145–3153. PMlR.
26.
AntamisTDrosouAVafeiadisT, et al. Interpretability of deep neural networks: a review of methods, classification and hardware. Neurocomputing2024; 601: 128204.
27.
CrabbéJSchaarMVD. Concept activation regions: a generalized framework for concept-based explanations. Adv Neural Inf Process Syst2022; 35: 2590–2607.
28.
BaiAYehCKRavikumarP, et al. Concept gradient: concept-based interpretation without linear assumption. arXiv preprint, arXiv:220814966, 2022.
29.
GhorbaniAWexlerJZouJY, et al. Towards automatic concept-based explanations. In: Advances in neural information processing systems. Curran Associates, Inc., 2019, p. 32.
30.
YehCChenYWuA, et al. AttentionViz: a global view of transformer attention. IEEE Trans Vis Comput Graph2024; 30(1): 262–272.
31.
DeRoseJFWangJBergerM.Attention flows: analyzing and comparing attention mechanisms in language models. IEEE Trans Vis Comput Graph2021; 27(2): 1160–1170.
32.
JiXTuYHeW, et al. USEVis: visual analytics of attention-based neural embedding in information retrieval. Vis Inform2021; 5(2): 1–12.
33.
CosciaAEndertA.Knowledgevis: interpreting language models by comparing fill-in-the-blank prompts. IEEE Trans Vis Comput Graph2024; 30(9): 6520–6532.
34.
HeWWangJGuoH, et al. CECAV-DNN: collective ensemble comparison and visualization using deep neural networks. Vis Inform2020; 4(2): 109–121.
35.
LiYWangJAboagyeP, et al. Visual analytics for efficient image exploration and user-guided image captioning. IEEE Trans Vis Comput Graph2024; 30(6): 2875–2887.
36.
WangJZhangWYangH.Scanviz: interpreting the symbol-concept association captured by deep neural networks through visual analytics. In: 2020 IEEE Pacific visualization symposium (PacificVis), 2020, pp. 51–60. IEEE.
37.
ZhaoZXuPScheideggerC, et al. Human-in-the-loop extraction of interpretable concepts in deep learning models. IEEE Trans Vis Comput Graph2022; 28(1): 780–790.
38.
LiGWangJShenHW, et al. CNNPruner: pruning convolutional neural networks with visual analytics. IEEE Trans Vis Comput Graph2021; 27(2): 1364–1373.
39.
WangJGouLYangH, et al. GANViz: a visual analytics approach to understand the adversarial game. IEEE Trans Vis Comput Graph2018; 24(6). 1905–1917. https://doi.org/10.1109/TVCG.2018.2816223
40.
LiuMShiJLiZ, et al. Towards better analysis of deep convolutional neural networks. IEEE Trans Vis Comput Graph2017; 23(1). 91–100. https://doi.org/10.1109/TVCG.2016.2598831
41.
La RosaBBlasilliGBourquiR, et al. State of the art of visual analytics for explainable deep learning. Comput Graph Forum2023; 42: 319–355.
42.
AchantaRShajiASmithK, et al. SLIC superpixels compared to state-of-the-art superpixel methods. IEEE Trans Pattern Anal Mach Intell2012; 34(11): 2274–2282.
43.
WardJH.Hierarchical grouping to optimize an objective function. J Am Stat Assoc1963; 58(301): 236–244.
44.
PearsonK.LIII. On lines and planes of closest fit to systems of points in space. Lond Edinb Dubl Philos Mag J Sci1901; 2(11): 559–572.
45.
HersbachHBellBBerrisfordP, et al. The ERA5 global reanalysis. Q J R Meteorol Soc2020; 146(730): 1999–2049.
46.
BrookeJ, et al. SUS-a quick and dirty usability scale. Usability Eval Ind1996; 189(194): 4–7.
47.
BangorAKortumPMillerJ.Determining what individual sus scores mean: adding an adjective rating scale. J Usability Stud2009; 4(3): 114–123.
Supplementary Material
Please find the following supplemental material available below.
For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.
For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.
