Please use this identifier to cite or link to this item: http://hdl.handle.net/2307/40931
Title: Neural-Network modelling for meteorological and climatological applications
Authors: Amendola, Stefano
Advisor: Pettinelli, Elena
Keywords: CLIMATE
ATTRIBUTION
Issue Date: 16-Mar-2020
Publisher: Università degli studi Roma Tre
Abstract: This thesis, entitled Neural network modelling for meteorological and climatological applications, aims to give new insight in open problems in the atmospheric sciences with a statistical method of investigation, based on neural network analysis. As we will see, the use of a data-driven method can be considered both as an alternative and a complementary tool of analysis respect to the usual approach based on dynamical modelling. During the last decades the field of the physics of the atmosphere obtained a rising interest by research centres worldwide and by the general public. Meteorology and climate are of primary importance for our lives, just think about the influence on sectors as civil protection, agriculture, tourism, transport or industry. Furthermore, climate change is a matter of deep discussion for both scientists and policymakers. The study of climate requires long data series, so this science rapidly grows hand by hand with the availability of meteorological data. Furthermore, thanks to the improvement of computer capabilities, in the second part of the last century operational meteorological-climatic models were born. They can be considered as a virtual laboratory in which the (complex) climate system is reproduced – with all its subcomponents and related interactions – by the equations of fluid dynamics and thermodynamics. In the application of these models we must keep in mind the limitations typical of numerical simulation of chaotic systems. In general, topics as weather, monthly and seasonal forecasts, climate projections and related impacts are approached by dynamical modelling. In some respects, for example for climate change attribution – i.e. the effort to scientifically establish causes for the recent warming of the Earth – this approach is a matter of discussion. In fact, the problem of attribution is mainly addressed by the so-called Global Climate Models (GCMs) – a type of meteo-climatic model – and, although in last decades great improvements are achieved, these models still have some limitations. In particular, the use of GCMs for attribution only partially allows to apply a robustness scheme. Robustness of the results relies on a condition of independence among the different models employed, i.e. reliability improves if the same results are obtained using different models, if independence between them holds. As in many cases GCMs have a common ancestor, the condition of independence is questioned. Furthermore, the unavoidable abstraction and idealization included in the physics of dynamical models does not make them able to completely represent the climate system. Consequently, in order to obtain stronger results, data-driven models have been used in atmospheric sciences. Among others, we underline the application of neural network (NN) analysis and Granger causality, which have been shown to be particularly useful in many applications. A multi-approach strategy permits to satisfy the just described condition of robustness: results obtained by different ways may complement each other and data-driven models can also improve the performance of dynamical models. Based on these arguments, data-driven approaches have found large application for the study of different topics related to climate and meteorology. Here, our interest is principally focused on neural network analysis, an artificial intelligence method. NNs can find nonlinear relationships between a set of predictors and a fixed target. The use of different sets of predictors permits to investigate the causes that may have generated the behaviour of the target. In the light of what has been said, the objective of this PhD thesis is to develop a neural network tool able to investigate some open topics about atmospheric sciences. In the context of the rapid changes to which we are witnessing, physical information of great usefulness can be achieved. We will see that this tool can be applied to a wide range of topics, covering a wide spatial scale from global to local. Furthermore, the consideration of multi-linear regression analysis parallelly to the neural network tool will be useful to both underline the goodness of the choice of a non-linear method and to deeply analyse the importance of linear or non-linear mechanisms. So, in Chapter 2 we furnish some generalities about climate, with considerations on its energy budget and the subcomponents in which it can be divided – obviously from a theoretical point of view. Then we introduce also the hot topic of climate change, related to the greenhouse effect. We conclude the first Chapter by introducing also the item of weather and climate modelling. Then, in Chapter 3 we will explain the Neural Network tool. Due to the central importance of neural analysis for this thesis, we start with a historical explanation of the development of this kind of method, from the origins up to the typology related to this research. It is a technical Chapter, useful to fix the idea about neural network analysis. It is closed by a section that explains in detail the tool developed during this three-year activity, with its peculiarities and novel characteristics. Each application has its proper NN model, all the differences respect to the explanation of Chapter 3 will be highlighted. With Chapter 4 the second part of the thesis starts, in particular this Chapter shows the application of the NN model to the debated theme of attribution of the mean global temperature behaviour on the last 150 years. We find a lot of interesting results with a completely independent mean of investigation respect to the dynamical models. Our tool will also be able to investigate about the causes that have determined the behaviour of global temperature in sub-intervals into the considered period. More insight about the role of sun is also given. Then, Chapter 5 is dedicated to the analysis of the causes behind the behaviour of the Atlantic Multidecadal Oscillation (AMO). AMO is a mode of variability of the sea surface temperature anomalies over the North Atlantic Ocean. For many years it has appeared as a natural component of the climate system, but recently several works linked it to anthropogenic activity appears. We insert in this open debate with our NN tool, i.e. with a completely different approach. In Chapter 6 we will show the first application of our neural network tool to study impacts related to weather/climate conditions. Here we focus on the analysis of the causes that could influence forced migrations from the Sahelian countries to Italy. It is a hot topic in the political agenda of many European countries. We will see impressive results, that relate climate variables and harvest yields to the migration rate to Italy. Finally, in Chapter 7 we conclude this work with an application of our tool to study the impact of meteorological variables on the dynamics of the observed quantity of a kind of sandfly responsible for the leishmanias spread. In fact, in recent years we observed an unusual expansion of such sandflies over areas previously considered immune. The results show that we are able to explain a wide majority of the variance in the data of population density. This is obtained through the application of our NN model driven in input by averaged mean temperature, relative humidity and temperature at 10 cm below ground during oviposition, larval and adult stages. Interesting results are achieved, showing the power of our model.
URI: http://hdl.handle.net/2307/40931
Access Rights: info:eu-repo/semantics/openAccess
Appears in Collections:Dipartimento di Matematica e Fisica
T - Tesi di dottorato

Files in This Item:
File Description SizeFormat
Neura_Network_Modelling_for_meteorological_and_climatological_applications.pdf4.7 MBAdobe PDFView/Open
Show full item record Recommend this item

Page view(s)

48
checked on Nov 29, 2023

Download(s)

13
checked on Nov 29, 2023

Google ScholarTM

Check


Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.