aalto1 untyped-item.component.html
Prediction of forest variables and its reliability with a hyperspectral deep learning model
Loading...
URL
Journal Title
Journal ISSN
Volume Title
Perustieteiden korkeakoulu |
Master's thesis
Unless otherwise stated, all rights belong to the author. You may download, display and print this publication for Your own personal use. Commercial use is prohibited.
Authors
Date
Department
Mcode
SCI3044
Language
en
Pages
63
Series
Abstract
With three-quarters of the land surface area is covered by forests, Finland is the most heavily-forested country in Europe. Forests have been the major source of both social well-being and economic of Finland. Therefore, it is important to understand the forest development, vegetation features and quality to help preserve the biological diversity of forests while securing different economic uses of forest in a good manner. In order to understand the forest development, the hyperspectral remote sensing has been used with the artificial intelligence methods to predict and analyze the forest variables. A novel convolutional neural network has been developed in the AIROBEST project to predict various the forest variables simultaneously. Despite of decent results for both categorical and continuous variables, there are a few problems with the predictions obtained from the model.
This thesis tackles the existing challenges in the prediction of forest variables from hyperspectral images by suggesting a new procedure for data processing and data split to utilize the large amount of hyperspectral data. Based on the new data split, many experiments were carried out to create a new baseline model. To review and assess the predictions of the baseline model, new evaluation metrics are proposed for stand-level analysis of the predictions. The predictions are also studied and quantified to understand their variation and reliability.
Overall, the new baseline model achieves a good overall accuracy 83.48\% and mean class accuracy 75.13\% for the categorical variables. For the continuous tasks, the model reaches 17.24\% relative root mean square error for four main continuous variables with a high coefficient of determination 0.75. At the end, the shortcomings of this thesis work are discussed and suggestions are given to improve the results as well as the reference data.