Multitask Convolutional Neural Network Emulators for Global Crop Models - Supervised Deep Learning in Large Hypercubes of Non-IID Data
Abstract: The aim of this thesis is to establish whether a neural network (NN) can be used for emulation of simulated global crop production - retrieved from the computationally demanding dynamic global vegetation model (DGVM) Lund-Potsdam-Jena General Ecosystem Simulator (LPJ-GUESS). It has been devoted to elaboration with various types of neural network architectures: Branched NNs capable of processing inputs of mixed data types; Convolutional Neural Network (CNN) architectures able to perform automated temporal feature extraction of the given weather time series; simpler fully connected (FC) structures as well as Multitask NNs. The NN crop emulation approach was tested for approximation of global annual spring wheat responses to changes in carbon dioxide, temperature, precipitation, and nitrogen fertilizer levels. The model domain is a multidimensional hyperspace of non-IID samples that can be blocked into climate-, temporal- and global position factor levels. NNs tend to focus on the normally behaving samples and take less notice of rare behaviors and relations. In this vast data set, the varying characteristics and relations can thus be hard to detect - even for a large and complex neural network. Due to these complexities in the LPJ-GUESS sample distributions and because neural network learning is heavily reliant on the data, a fair share of the thesis has been dedicated to network training and sample selection - with the purpose of improving learning without causing the network to overfit. Further, the Köppen climate classification system, based on historical climate and vegetation, was used for aggregation of the emulator domain - in order to form smaller homogeneous groups. It is easier to dissect the data when these disjoint groups are analysed in isolation, which in turn can facilitate input variable selection. Moreover, by aggregating the model domain and allowing for separate deep learning of each domain-fraction, several sub-models can be constructed and trained for a specific Köppen climate region. These can then be combined into an integrated composite emulator. In contrast to an emulator trained to model crop production for the whole domain, a model composition emulator does not have to account for the differences between the sub-domains and hence only has to focus on learning the within-group relations and patterns in the disjoint climate classes.
AT THIS PAGE YOU CAN DOWNLOAD THE WHOLE ESSAY. (follow the link to the next page)