Enhanced duckweed detection using bootstrapped SVM classification on medium resolution RGB MODIS imagery

Abstract

From early 2004, Lake Maracaibo (northwest Venezuela) experienced an unprecedented invasion of duckweed Lemna obscura. Recurrent blooms of the plant in the past 2 years illustrate the need for an automatic monitoring method to follow the plant cover with time and to plan contingency measures. We present an approach that allows the cover of the duckweed to be quantified through the classification of MODIS 250 m RGB composite images available from the internet. The method improves the accuracy of the results of the Support Vector Machine (SVM) algorithm for classification by including a bootstrap step during the training phase. Using only 200 pixels for training (<0.05% of the total), the bootstrapped SVM method allows a better identification of the duckweed class, reducing the number of false negatives by half and improving the KHAT statistic by almost 40% in comparison to the standard SVM method. This method has proved to be a reliable solution in cases where rapid responses are needed and only medium‐resolution, free satellite imagery is available.

Acknowledgements

This work was partially supported by the Decanato de Investigación y Desarrollo of the Simón Bolívar University, grant DID‐GID‐03/2004. We thank Fermín Avila and Federico Troncone from the Instituto para la Conservación del Lago de Maracaibo (ICLAM) for their helpful insights on the evolution of the duckweed invasion. We also thank Judd Taylor at IMaRS for maintaining the MODIS products for the Maracaibo Lake, Chuanmin Hu and Brock Murch for their help with the description of the MODIS product and Frank Müller‐Karger for his valuable comments on the Lemna invasion. The system described is available online, and details can be provided on request.