The use of multivariate techniques and neural network predictions to the greenhouse gases database, climate parameters and biomass burning at Amazon

• Grant number: 19/21789-8
• Support type: Research Grants - Visiting Researcher Grant - Brazil
• Duration: February 01, 2020 - January 31, 2021
• Field of knowledge: Physical Sciences and Mathematics - Geosciences
• Principal researcher: Luciana Vanni Gatti
• Grantee: Luciana Vanni Gatti
• Visiting researcher: Sergio Machado Correa
• Visiting researcher institution: Universidade do Estado do Rio de Janeiro (UERJ). Faculdade de Tecnologia (FAT), Brazil
• Home Institution: Instituto Nacional de Pesquisas Espaciais (INPE). Ministério da Ciência, Tecnologia, Inovações e Comunicações (Brasil). São José dos Campos , SP, Brazil
• Associated research grant: 16/02018-2 - Interannual variation of Amazon Basin greenhouse gas balances and their controls in a warming and increasingly variable climate – Carbam: the Amazon carbon balance long-term study, AP.PFPMCG.TEM

Abstract

The project focus is the statistical treatment of the data obtained since 2010, of the greenhouse gas balance interannual variation in the Amazon Basin. In the first stage, the information already collected will be integrated into a single big database, with columns referring to variables such as CO2, CH4, N2O and SF6 concentration, other gases, meteorological parameters and data regarding the sample collection, with the database rows refer to observations for each time interval. In the second stage, the data will be validated, using descriptive statistics, to evaluate the minimum, maximum, average, median values, among other properties, with a view to deciding whether or not to remove anomalous data, according to criteria to be defined, either by events occurred during sampling or by operating limits of analytical equipment. Then the data will still be treated descriptively, but in such a way that the entire data can be evaluated visually, for example in the form of boxplots. In the third stage multivariate statistics will be used. The first evaluation tool is the correlation matrix of all data using Pearson's method or similar, to see the correlation between the entire data and hierarchical groupings. Then the data will be treated to create the principal component analysis (PCA), which uses grouping criteria to identify stronger relationships between groups of variables, either directly or indirectly. Further the hierarchical groupings will be classified by different methods, such as Euclidean distances, for the assembly of dendrograms in three dimensions. Finally the data will be treated by classifying algorithms, such as Boruta, in order to evaluate the degree of importance in the change of key properties (eg temperature, precipitation, among others) against all variables in the database. In the fourth stage of the work, the database will be used to feed a neural network algorithm to predict future scenarios against events that may occur in the Amazon Basin, such as increased burns, drought periods, changes in land use, weather events, among others. All project steps will be conducted using the open source R Language. (AU)