G5 Doctoral dissertation (article)
Machine Learning Approaches for Natural Resource Data

List of Authors: Pohjankukka Jonne
Publisher: Turku Centre for Computer Science
Place: Turku
Publication year: 2018
eISBN: 978-952-12-3710-2


Real life applications involving efficient management of natural resources are dependent on accurate geographical information. This information is usually obtained by manual on-site data collection, via automatic remote sensing methods, or by the mixture of the two. Natural resource management, besides accurate data collection, also requires detailed analysis of this data, which in the era of data flood can be a cumbersome process. With the rising trend in both computational power and storage capacity, together with lowering hardware prices, data-driven decision analysis has an ever greater role. 

In this thesis, we examine the predictability of terrain trafficability conditions and forest attributes by using a machine learning approach with geographic information system data. Quantitative measures on the prediction performance of terrain conditions using natural resource data sets are given through five distinct research areas located around Finland. Furthermore, the estimation capability of key forest attributes is inspected with a multitude of modeling and feature selection techniques. The research results provide empirical evidence on whether the used natural resource data is sufficiently accurate enough for practical applications, or if further refinement on the data is needed. The results are important especially to forest industry since even slight improvements to the natural resource data sets utilized in practice can result in high saves in terms of operation time and costs. 

Model evaluation is also addressed in this thesis by proposing a novel method for estimating the prediction performance of spatial models. Classical model goodness of fit measures usually rely on the assumption of independently and identically distributed data samples, a characteristic which normally is not true in the case of spatial data sets. Spatio-temporal data sets contain an intrinsic property called spatial autocorrelation, which is partly responsible for breaking these assumptions. The proposed cross validation based evaluation method provides model performance estimation where optimistic bias due to spatial autocorrelation is decreased by partitioning the data sets in a suitable way. 

Last updated on 2019-20-07 at 15:17