The impact of the spectral dimension of hyperspectral datasets on plant disease detection

Abstract

Precision Agriculture as an information based approach requires explicit spatial information about the within field heterogeneities for site-specific applications. Thus, the usage of cost-intensive agrochemicals and the impact on the environment can be significantly reduced. Spectroscopic approaches are thereby a promising tool for providing fast and precise information on a local to regional level. In this thesis, the impact of hyperspectral near-range and remote sensing data for crop stress detection will be observed since spectroscopic approaches are of great interest for Precision Agriculture. Two greenhouse experiments and three field experiments were conducted with spectroscopic measurements to examine possibilities and limitations of hyperspectral data. The data were acquired using a near-range non-imaging spectrometer (ASD Fieldspec 3) and a near-range imaging spectrometer (ImSpec V10E) in the greenhouse, or were acquired by the airborne sensor systems HyMap^TM, ROSIS or AISA for the field experiments. The methodical foci thereby are the improvement of binary detection approaches, discriminating 'vital' and 'infected' wheat stands or parts of wheat stands, and quantification approaches to estimate disease severities at canopy level.<br /> This thesis examines the spectral dimension of hyperspectral data for crop stress detection by assessing data redundancy and defining spectral necessities. Different feature selection methods were tested for their suitability in reducing the high amount of spectral data without losing significant information. Conventional classification approaches and recent developments, such as support vector machines for classification (SVM), were thereby tested based on the identified spectral subsets to assess the status of different wheat stands. By focusing on phenomenon-specific spectral bands, stressed wheat stands could successfully be identified with high accuracies. Using optimal band combinations could even increase classification accuracies. The results showed that not the entire spectrum of hyperspectral data is necessary for the detection of fungal infections in wheat. These findings are particularly interesting for future spectral sensor design and remote sensing missions that are aiming at the provision of spatial information for agricultural practice.<br /> The ability of hyperspectral data in quantifying the severity of fungal diseases was observed. Site-specific fungicide treatments based on application maps are technically possible and doses can be adjusted if the maps provide information about the health status of the crops. Crop growth anomalies caused by fungal infections were observed, which differed significantly within one field. The derivation of disease severities based on hyperspectral near-range and remote sensing data were examined using classification approaches and support vector machines for regression (SVR). Fungal infections of wheat stands in the greenhouse and wheat stands in the field could be quantified with a high level of certainty. The results are very promising and the findings may be of great interest for agricultural questionnaires and automatic phenotyping approaches, since the presented approaches are fast and non-destructive. Spatial maps with continual disease severity data could be derived, which can be used to generate application maps for agricultural practice.<br /> Since the study shows that a reduction of hyperspectral data to a few but specifically selected spectral bands can improve the classification accuracies or regression analyses, a preliminary feature selection should be considered when working with hyperspectral remote sensing data. Agricultural and geographical approaches that are based on spatial-spectral information may thus profit from a faster and more reliable extraction of information. The study shows great advantages of the usage of hyperspectral imaging data but also the necessity of advanced and innovative analyzing methods.