| Summary: | The goal of this thesis is to build a near-real time defoliation detector that could be used, early in the spring, to find out which areas of a birch forest are being affected by an insect outbreak. The importance of a reliable detection at the beginning of the spring lies on the possibility for an early intervention. The data on which the study is based is the Normalized Difference Vegetation Index (NDVI), which is calculated from the 8-days interval measurements, obtained from the MODerate resolution Imaging Spectroradiometer (MODIS). Further available information that is included in the analysis are the forest fraction and the altitude of each pixel. The first step of the analysis is fitting a function to the measurements of each pixel whose parameters could capture the important aspects of the changes in NDVI values. Among them, the final NDVI value that is reached during the summer is of particular interest since an abnormally low value could be the indication of an insect infestation. Different assumptions are made on the error distribution. The first ones, more simplystic, do not manage to counteract the noise. A more complex error structure is thus taken into account, leading to a better estimate which is then used to build the estimator. The idea behind the detector is to identify those pixels for which the NDVI does not reach its high late spring/summer values fast enough, with respect to other pixels and previous years. It is known that the two years, among those available in the dataset, that have suffered a moth outbreak are 2004 and 2013. Hence, the estimation of the fitting function is run for 2000-2003, the detector is tried on 2004 and finally tested on 2013, since for this year a few locations of the outbreak were known. The discrepancy between field data and the results generated by the detector suggests further adjustments that would improve the capacity to detect moth infestation. Early and reliable detection of pest infestation in forests is crucial to protect the health of trees. Insect outbreaks are an important cause of defoliation: they delay the blossoming of leaves and thereby affect the growth of trees. Such delays lower the economic potential of forests as well as their capacity to absorb atmospheric CO 2 . The aim of this thesis is to use mathematical models to build a near real-time detector that can identify abnormal behavior in forest growth, which can be the sign of pest infestation. In order for the detector to provide effective early warnings, the analysis uses high time-resolution satellite images. In particular, the detector is based on data collected at the Abisko National Park in the north of Sweden. The insect that is known to threaten this birch forest is a moth called Epirrita Autumnata, whose larvae feed on the bursting leaves early in the spring. The data used to study the behaviour of green mass is the Normalized Difference Vegetation Index (NDVI). A dataset of 14 years of measurements was available over an area of 350 km 2 . The index is calculated based on images collected by the MODerate resolution Imaging Spectroradiometer (MODIS), which is placed on two satellites that orbit Earth and collects images of the surface daily. The analysis also includes information on forest fraction and altitude for each pixel of the area. The first step of the analysis is fitting a function to the NDVI measurements of each pixel. The chosen function captures all important aspects of the change in NDVI during the spring. Different methods are used to fit the function to the NDVI. The first, simplistic models fail to fit the fuction because of the strong noise that affects the measurements. Therefore more robust and complex estimators are tried out. The final, best performing technique is used to construct the detector. The idea behind the detector is to identify those pixels for which NDVI grows slower than expected, based on the values of other pixels and previous years, during the spring. In particular, it is known that, within the available dataset, the two years that have suffered a moth outbreak are 2004 and 2013. Hence, the chosen function is fitted to the data from 2000-2003, in order to get a sense of the behaviour of NDVI during ”healthy” years. The detection of abnormal behaviour is done for 2004 and it is then tested on 2013. For this last year, a few locations of the outbreak were known, so that the results generated by the detector could be verified. The discrepancy between field data and the results generated by the detector suggests further adjustments that would improve the capacity to detect moth infestation.
|
|---|