This dissertation presents methods for automatic analysis of end faces from Scots pine (Pinus sylvestris L.) and Norway spruce (Picea abies (L.) Karst). Two features are extracted: the pith (centre core) position and the number of annual rings. The pith detection uses local orientations to estimate the centre of the annual rings in a manner that is robust to disturbances as knots and cracks as well as partial coverage of dirt or snow. The number of annual rings is counted using the polar distance transform, a tool developed here. This transform combines the image intensity and the circular shape of the rings so that the annual ring pattern can be outlined in rough and noisy images. First the marks on end faces from uneven sawing are removed using an automatic method developed in this work.
The data are images of untreated end faces mostly acquired at sawmills. A large amount of the data was imaged using a camera mounted above a conveyor belt at a sawmill, collecting images every month during one year. In total, the data consists of over 4000 images of pine and spruce. In this dissertation an algorithm for generating synthetic log end face images is also presented. The synthetic data can be used as a tool for developing image analysis methods.
The annual ring measurements were thoroughly evaluated on pine end face images acquired using the mounted end face camera. This evaluation shows that the method performs equally well as an experienced manual grader for grading the logs into quality classes. The method can thus be used as a component of an automatic grading system, overseen by a manual grader.
The hyperspectral reflectance measurements in the open field took the form of instantaneous spectra recording using an existing method called feature vector based analysis (FVBA), which was applied on disease severity. A new method called iterative normalisation based analysis (INBA) was developed and evaluated on disease severity and plant biomass. The methods revealed two different spectral signatures in both disease severity and plant density data. By concentrating the analysis on a 12% random subset of the hyperspectral field data, the unknown part of the data could be estimated with 94-97% coefficient of determination.
The empirical model for site-specific weed control combined a model for weed competition and a dose response model. Comparisons of site-specific and conventional uniform spraying using model simulations showed that site-specific spraying with the uniform recommended dose resulted in 64% herbicide saving. Comparison with a uniform dose with equal weed control effect resulted in 36% herbicide saving.
The methods developed in this thesis can be used to improve systems for site-specific plant protection in precision agriculture and to evaluate site-specific plant protection systems in relation to uniform spraying. Overall, this could be beneficial both for farm finances and for the environment.