Forest Health Monitoring by Detection of Biotic Agents in Satellite Images

Abstract

Since 1999, the Pine Wood Nematode (PWN) has spread, infected and damaged growing areas of pine forest in Portugal. The pest is subject to strict quarantine measures, that require forest owners to register, fell and dispose of infected trees. As remote sensing from satellites provides repetitive and consistent data sets with high temporal resolution and large spatial coverage, a surveillance based on satellite images would be advantageous. Moreover, since multispectral images from the European Space Agency's (ESA's) Sentinel-2 mission are freely available, a monitoring based on these data is preferrable. The most commonly used tool for monitoring vegetation by remote sensing is the Normalized Difference Vegetation Index (NDVI). However, the infection by PWN may appear more prominent in other combinations of channels, and by introducing machine learning techniques, allowing for the exploration of changes in all spectral bands, one may end up with a better discrimination. This thesis is intended as a feasibility study that aims to explore the viability of an operational PWN detection system based on Sentinel-2 data. To this purpose, traditional feature extraction algorithms, including Fisher's linear discriminant analysis (FLDA), and sparse linear discriminant analysis (SLDA) are examined, as well as some spectral unmixing algorithms, including iterated constrained energy minimization (ICEM), mixture-tuned matched filtering (MTMF) and the 2D-Corr-NLS algorithm. The methods are examined with respect to the impact of spatial and spectral resolution, and the performances are measured and compared using a constant miss rate (CMR) detector. Multisource data with different spatial and spectral resolution are used in experiments to investigate how these resolutions constrain successful PWN detection. The results show that the spatial resolution of the Sentinel-2 data is too low for the selected methods to be useful for PWN detection, whilst it remains an open question whether the spectral resolution of Sentinel-2 is a limitation.