COS 66-10
Using computer vision techniques to test the accuracy and precision of plant dieback estimates
Plant dieback is measured by visual inspection and estimation of the percentage of dead material. Although this method of estimation is widespread, two issues are unclear: a) the level of accuracy in any estimate of a particular plant’s dieback, and b) the amount of variation in estimation of dieback between different observers. In seeking answers to these questions, we developed an application in Python utilizing functions from the OpenCV library, which we used to estimate the percentage of dieback in high-resolution JPEG top view images of 10 Azorella compacta specimens. By sweeping through an image with a customized patch-based back-projection in HSV color space using the Earth Mover’s Distance, we computed an estimate of the percentage of dieback area to total area. These estimates were then compared to 19 human estimates of the same percentages, and we measured the relative error for each human estimate of dieback. Subjects visually estimated the proportion of dieback to growth in 10 images. The first 5 were simple images composed of green circles and ellipses with brown regions formed by circles, ellipses, and irregular patches representing “dieback”, and the last 5 were images of actual A. compacta specimens with various distributions of dieback.
Results/Conclusions
Regarding the level of accuracy in any estimate of a particular plant’s dieback, the relative error in a particular estimate ranged from 21% for the simplest abstract image to 330% for the more complex specimen images. For the second issue of variation in estimates of dieback between different observers, the mean relative estimation error for n=190 estimations was 113.5% with a standard error of 16%. Of these estimates, 164 were over-estimates and 26 were under-estimates. Over the 10 estimations, the most accurate estimator had a mean relative error of 31.3% with a standard deviation of 105.2% and the least accurate had an error of 379.2% with a standard deviation of 688.3%. There was a significant (p < 0.001) correlation between the mean relative error and standard deviation in a particular individual’s estimates with a correlation coefficient of r = 0.9349. This indicates that there is a consistent, natural tendency to significantly over-estimate the percentage of dieback. Thus, ideally ecologists could be trained in the field with known dieback images and calibrate their estimates accordingly, which would reduce error among observers. This would improve both accuracy and precision of dieback estimates.