Description
Date depot: 16 juin 2023
Titre: Modeling the carbon impacts of land-use changes via machine learning
Directrice de thèse:
Claire MONTELEONI (Inria-Paris (ED-130))
Domaine scientifique: Sciences et technologies de l'information et de la communication
Thématique CNRS : Intelligence artificielle
Resumé: Land-use change has the potential to reduce CO2 emissions, which is key to mitigating the most severe risks of of climate change. However the possible impacts of land-use changes on CO2 emissions and their relevant uncertainties are not well understood. Machine learning can help address this problem in a variety of ways. This doctoral project will focus on improving the modeling of land-use change impacts on CO2 and quantifying the relevant uncertainties, towards accelerating the reduction of CO2 emissions over land. The primary ways machine learning will be applied to this problem are:
a. Developing data-driven emulators of physical processes relevant to modeling land-cover change
b. Downscaling climate variables currently modeled at course scale in global climate models to the fine-scales needed to quantify local impacts.
a. Running a (physics-driven) Earth System Model (ESM) is quite computationally expensive. The goal of developing machine learning emulators for ESMs is in part to reduce the compute time needed, because after training the machine learning model, it is much quicker to run than existing physics-based models. Machine learning is also useful for modeling complex systems where the physics may be unknown and/or involve spatiotemporal scales that are difficult to model. We will explore the state-of-the-art in machine-learning-based emulators of ESMs, in order to develop emulators for the specific processes relevant for the task of modeling the effects of land-use change on CO2 emissions. Initially this can be framed as a supervised learning task. In the second part of the thesis we will look at pipelines involving downscaled data, which we will also do via machine learning, both supervised and unsupervised.
b. We can first train machine learning models to downscale the course resolution ESM outputs, using a supervised learning framework. In particular, the data from the physics-driven ESMs, which is at course resolution in both time and space, will be used as the training data, which will be “labeled” by fine resolution (in space and time) ground-truth data types. Since the data are in the form of spatial fields, we will compare the performance of convolutional neural networks (CNNs) for the task. Recent work has also shown the benefit of Conv-LSTM (which uses convolutions in a form of recurrent neural network) for downscaling one data field using other data fields for training (Harilal et al., 2021). Meanwhile, since there are non-linear relationships in these fields over both space time, as well as across different fields, we will also use self-supervised learning methods involving masking parts of the data, and learning attention-based models (e.g. Transformer; Vaswani et al., 2017) to estimate it from the remaining data. Our team has demonstrated the promise of unsupervised deep learning methods for domain alignment for the general task of downscaling geospatial data (Groenke et al., 2020). We are currently developing an ML-based method for temporal downscaling. The proposed doctoral thesis project will apply and extend these methods to improve our modeling of land-use change impacts.
Résumé dans une autre langue: Anglais
Doctorant.e: Clyne Graham