Detection of fungal infection in wheat with high-resolution multispectral data

Abstract

The exact knowledge of the spatiotemporal dynamics of crop diseases for an implementation of a site-specific fungicide application is fundamental. Remote sensing is an appropriate tool to monitor the heterogeneity of fungal diseases within agricultural sites. However, the identification of an infection at an early growth stage is essential. This study assesses the potential of multispectral remote sensing for multitemporal analyses of crop diseases. Within an experimental test site near Bonn (Germany) a 6-ha sized plot with winter wheat was created, containing crops with each possible infection stage of three different pathogens. Two multispectral QuickBird images (04/22/2005 and 06/20/2005) and a spectrally resampled HyMap image (05/28/2005) were used to analyse the spatiotemporal dynamic of infection. The data preprocessing comprised a radiometric and a precise geometric correction by using DGPS-measurements that is an important requirement for Precision Agriculture applications. Ground truth data, in particular infection severity, growth stage/height, and spectroradiometer measurements were collected. A decision tree, using mixture tuned matched filtering results and a vegetation index was applied to classify the data (infected and non-infected areas). Classification results were compared to ground truth data. The classification accuracy of the first scene was only 56.8% whereas the scene of 28 May (65.9%) and the scene of 20 June (88.6%) achieved considerably higher accuracies. The results showed that high-resolution multispectral data are generally suitable to detect in-field heterogeneities of vegetation vitality though they are only moderately suitable for early detection of stress factors.