An Automatic Method for Screening Water Hyacinth from Optical Satellite Imagery

Abstract

The availability of reliable and real time remote sensing data becomes permissible in addressing the spread of water hyacinth aquatic over freshwater bodies and in a wide geographic distribution area. However, some of the methods used in to discriminate water hyacinth were limited in yielding greater accuracy, some of the procedures produce errors that stem from the observer misidentification, poor estimation, and location accessibility bias. In addition, the existence of noninvasive plant species mixed with water hyacinth make the estimation precision more complicated. As a result, several studies implemented several techniques but yielded quite a different result of water hyacinth coverage over Lake Tana even with in the same season and years. Therefore, the aim of this study was to develop an automatic method for screening water hyacinth from water surfaces by discriminating invasive species from noninvasive co-existing species. To design an automatic screening method, 8 Landsat 8 imagers were used. In connection, a total of 450 GCP points was used for threshold selection, optimization and accuracy assessment purposes. ArcMap 10.5 was used to compute cell statistic and to extract cell values to the GCP points; while ERDAS imagine 2015 was used to conduct image preprocessing, signature evaluation and accuracy assessment. In addition, STATA 14 was also employed to calculate independent sample t-test, bivariate probit, binary logistic regression, RMSE and R2 values. Independent sample t-test was used to examine whether there is a difference in spectral reflectance values in between water hyacinth and other noninvasive vegetation with GCP. Bivariate probit and binary logistic regression tests was used to identify the best indices and optimized threshold value. Moreover, PyCharm 3 was used to write the algorithm and the prototype. In this case GDAL/OGR, Tkinter, Scipy, Matplotlib, and Numpy libraries were adopted to build the algorithm and the prototype. The result of this study confirmed that though NDVI was the best indices based on 1% amplitude of global thresholding values than other indices (NDWI and SAVI), the estimation of water hyacinth coverage based on 1% amplitude of threshold needs optimization process. Accordingly, binary, logistic regression model result revealed that NDVI based estimation above 0.5 pixel values yielded odds ratio of 6.88 indicating that probability of successful prediction of water hyacinth increases 6.88 times or by 588% when water hyacinth is predicted from NDVI value above 0.5. Again, this threshold has an overall accuracy of 98 % with a kappa coefficient of 0.94 implying that this threshold had a higher accuracy compared to other proposed optimization thresholds. Thus the result also indicated water hyacinth coverage varies from season to season and this variation ranges from 26,332,336(summer) – 9,142,164 (winter) m2 of land over lake Tana and currently, the areas which have been invaded by the water hyacinth are located on the northeastern shore of the lake. Based on the finding the study concluded that the NDVI thresholding approach above 0.5 pixel value allowed Landsat 8 image based detection by circumventing the labelling costs associated for regular water hyacinth by discriminating other aquatic vegetation over Lake Surface. Moreover, NDVI threshold-based detection technique enables an automatic screening of water hyacinth from a raw satellite images starting from radiometric correction to visualizing hyacinth coverage by simply providing a raw cloud free satellite images. Based on the finding the study recommends for environmental protection bureau and environmental expertise to use the proposed algorithm for regular monitoring purposes of water hyacinth.