Pro gradu, diplomityö, ylempi amk-opinnäytetyö (G2)

Environmental damage assessment based on satellite imagery using machine
learning





Julkaisun tekijät: Zelioli Luca

Kustantaja: Åbo Akademi

Julkaisuvuosi: 2020

Sivujen määrä: 95

Verkko-osoite: https://urn.fi/URN:NBN:fi-fe202003259258


Tiivistelmä

The aim of this thesis is to provide a source of information about damage assessment in forestry using deep learning. A large source of environmental information is provided by satellites imagery. Orbital devices are equipped with sensors that read the frequency variations
in the terrestrial electromagnetic field. The information obtained by these devices is composed by collections of dots. Machine learning methodologies, however, have the ability to transform raw data into human-understandable output.

Cloud and blur represent artefacts that need to be tackled to obtain high-quality imagery. For instance, a deep learning neural network, a Generative Adversarial Neural Network, can extrapolate the cloud compound from the image. Moreover, resampling techniques are used to improve their resolution. In this way, it is possible to correct the overall quality of satellite data.

The Finnish Kvarken Region, situated in the province of Vaasa, comprises a delicate forestry zone. Climate changes and the rise of temperature are influencing the forest quality negatively. Moreover, the public company in charge of the operational management needs new tools in order to enhance the environment condition.

Plenty of satellite data analysis frameworks are available for the consumer. In particular, SNAP, QGIS and ArcGIS offer capabilities to analyze environmental damage. Moreover, Google Earth Engine uses powerful programming languages such as Python to elaborate information from the Kvarken Region. It is also possible to study the historic forestry change from the past years until today. Unsupervised and supervised machine learning models are used to underline the difference between techniques. Deforested areas in the
Kvarken Region are mapped using state-of-the-art deep learning architectures for image segmentation. The implementation is done using Python programming language and open-source libraries such as TensorFlow and Keras.


Last updated on 2022-30-01 at 11:50