A comparison of manual interpretation and object-oriented analysis of digital photographs to assess relative cover of phenologically distinct plant groups in small-scale field plots

Background/Question/Methods

            Remote sensing and image analysis have proven to be a valuable means of analyzing vegetation dynamics over large areas.  Tools used to assess vegetation patterns at the landscape scale can also be applied to analysis of smaller-area images, such as digital photographs of meter-scale field plots, and offer the potential for increased speed and consistency in analysis.   One promising new image-processing approach is object-oriented analysis in which computer software, such as Definiens, identifies and classifies polygons (objects) using a process tree that considers spectral properties, texture, shape, size, and other characteristics, in a manner analogous to the way humans identify objects. In this work, we compared the relative effectiveness of manual interpretation and object-oriented analysis of digital photographs taken of 140 meter-scale experimental plots in annual grasslands.  The plots were installed in natural grassland patches to study interactions between resident forage grasses and invasive grasses with a distinct, late-season phenology.  Our aim was to evaluate the patch structure of each plot, specifically the relative distribution and fractional cover of each of these two plant types. 

Results/Conclusions

            We found that manual interpretation and object-oriented analysis of images exhibited different strengths and weaknesses for our application.  One strength of manual interpretation is that it requires less training for operators and is easier to do.   Manual interpretation also results in less misclassification of large polygons.  However, manual interpretation is tedious, results in greater misclassification of small polygons, and exhibits greater variability among time and among operators.  In contrast, the object-oriented approach allowed development of standardized process trees that with minor adjustments could be applied consistently across images.  This approach also more consistently classified small polygons correctly.  Misclassification of large polygons was more frequent with object-oriented analysis, but could be reduced with operator review.   The greatest weakness of the object-oriented approach is that the software is very expensive and requires at least several weeks to a month for a novice to master.  Overall, when resources are available and there is a large volume of images to evaluate, we recommend the object-oriented approach with operator review, due to the consistency and repeatability of results obtained.