NERSCPowering Scientific Discovery Since 1974

ClimateNet

Bringing the power of Deep Learning to the climate community via open datasets and architectures

(Click here for The ClimateContours labeling tool)

The Mission

The ClimateNet Project seeks to address an major open challenge in bringing the power of Deep Learning to the climate community, viz. that of creating community-sourced open-access expert-labeled datasets and architectures for improved accuracy and performance on a range of supervised learning problems, where plentiful reliable labelled training data is a requirement. 

The Motivation

Pattern recognition tasks such as classification, localization, object detection and segmentation have remained challenging problems in the weather and climate sciences (see figure below for analogues between classic computer vision tasks and corresponding climate science tasks).

ComputerVisionClimateScienceAnalogue

 

While there exist many heuristics and algorithms for detecting weather patterns or extreme events in a dataset, the disparities between the output of these different methods even for a single class of event are huge and often impossible to reconcile. See, for example, the ARTMIP project that attempts to compare, contrast and reconcile the dozen or so atmospheric river tracking algorithms. 

Meanwhile, Machine Learning and Deep Learning based approaches for pattern recognition and pattern discovery tasks are showing great promise across the sciences. However, for supervised ML and DL methods, availability of plentiful reliable training data is key. Given the scarcity of reliable labelled training data for supervised Deep Learning based approaches and the pressing need to address this problem, we propose creating a community-sourced expert-labelled database.

The Vision and Proposed Approach

The figure below captures our overall vision for a unified Deep Learning workflow, as relevant to climate science. The workflow can be split into two pieces: the Training phase and the Inference phase. The goal of the Training phase is to produce a single, unified Deep Network that is trained by examples from `hand'-labeled examples by human experts. In the inference phase, the unified Deep Network is applied to archives of multi-resolution, multi-modal datasets from climate model output or reanalyses or observational datasets. 

ClimateNetWorkflow

 

A fundamental challenge in training Deep Networks is the availability of reliable suitably labeled training data. We propose creating a `ClimateNet' dataset, with an accompanying schema that can capture information pertaining to pattern labels, bounding boxes and segmentation masks. Domain experts (atmospheric and climate scientists) contribute `hand'-labeled information to the ClimateNet dataset -- the proposal was first presented by Prabhat, Karthik Kashinath et al. at AGU 2017 and a project update was presented at AGU 2018.

We have developed a web interface called ClimateContours based on the "Label-Me" tool developed at the MIT computer vision lab to crowdsource the `hand'-labeling task. Note that the overall spirit of the Deep Learning methodology is to avoid the prescription of heuristics for defining weather patterns; it has been established by the computer vision community that Deep Learning is effective at learning relevant features for pattern classification tasks, without requiring application-specific tuning or feature engineering. The figure below shows a snapshot of the labelling tool on the web interface mentioned earlier.  

Once the ClimateNet dataset has over a thousand labelled images (multi-variate snapshots from high-res 25-km CAM5 data), we will extend state-of-the-art segmentation CNN architectures to learn unified representations for various weather patterns, including atmospheric rivers, tropical cyclones, extra-tropical cyclones, and weather fronts. This will build upon our previous work, notably -- (1) Liu et al. (2016), Application of deep convolutional neural networks for detecting extreme weather in climate datasets, arXiv preprint arXiv:1605.01156, (2) Mudigonda et al. (2017), Segmenting and tracking extreme climate events using neural networks, Deep Learning for Physical Sciences, NIPS workshop, (3) Racah et al. (2017), ExtremeWeather: A large-scale climate dataset for semi-supervised detection, localization, and understanding of extreme weather events, 31st conference on Neural Information Processing Systems, and (4)  Kurth, T., Treichler, S., Romero, J., Mudigonda, M., Luehr, N., Phillips, E., Mahesh, A., Matheson, M., Deslippe, J., Fatica, M., Prabhat, and Houston, M. Exascale deep learning for climate analytics. In Proceedings of the International Conference for High Performance Computing, Networking, Storage, and Analysis, pp. 51. IEEE Press, 2018. (Also winner of the Gordon Bell Prize in 2018)

We will then apply the unified network to archives of multi-resolution, multi-modal datasets from climate model output, reanalyses and observations to seamlessly extract class labels, bounding boxes and segmentation masks.

The ClimateContours labeling tool

 

ClimateContours ClimateNet