Automated detection of Pueraria montana (kudzu) through Haar analysis of hyperspectral reflectance data

Abstract

The automated detection of noxious weeds using remote sensing techniques would be of great benefit for their monitoring and control. In this article, the Haar discrete wavelet transform (DWT) method is investigated for extracting pertinent features from hyperspectral signatures. Based on the Haar DWT features, a fully automated detection system is designed and evaluated to determine its performance for the practical use of kudzu detection. For performance evaluation, the authors use a leave-one-out test of a nearest mean classifier to compute classification accuracies and the corresponding 95% confidence intervals. When the system was tested to determine its ability to classify each of five classes of weeds, including kudzu and four similar broadleaf weeds, the classification accuracy was 90.2%/spl plusmn/4.4%. When the system was tested to determine its ability to detect kudzu among a mixture of the four weed types, the classification accuracy was 100%.