I recently stumbled upon an interesting comparative analysis of several software packages that offer Object-Based Image Analysis (OBIA) analysis approaches. No surprise, Trimble eCognition was one of them.
Lourenço et al. 2020 examined the potential of eCognition Developer, Orfeo Toolbox/Monteverdi and ESRI ArcGIS within the context of identifying invasive species in their paper “Assessing the performance of different OBIA software approaches for mapping invasive alien plants along roads with remote sensing data”.
For the purpose of the study, the authors chose a 112 km length of national roads in Portugal that are “characterized by a diverse mosaic of land-cover types” including several Invasive Alien Plant (IAP) species such as “Silver wattle (Acacia dealbata), Australian blackwood (Acacia melanoxylon), Black locust (Robinia pseudoacacia), Tree of heaven (Ailanthus altissima) and Giant reed (Arundo donax)”.
The study aimed at mapping these species based on VHR aerial images acquired in 2016 with a spatial resolution of 10 cm and spectral resolution that included Red, Green Blue and Near-Infrared bands.
A truly 1-to-1 comparison of the software packages is difficult due to the differences in the tools each includes. For example, the first step the authors discuss is the segmentation – the creation of image objects. A Mean Shift Segmentation (MSS) was available and used in the Orfeo Toolbox/Monteverdi and ESRI ArcGIS applications, whereas the Multiresolution Segmentation (MRS) algorithm was used in eCognition.
In terms of the classification approach, three supervised classification models were used, all trained based on field data and expert knowledge and executed on 15 different images. A Support Vector Machine (SVM) was used in Orfeo Toolbox/Monteverdi (also available in eCognition, but not used by the authors); a Nearest Neighbor classifier (KNN) was used in eCognition; and a Maximum Likelihood classifier (MLC) was used in the ArcGIS analysis.
According to the article, the authors did not perform any additional image object refinement or OBIA steps post-classification.
The classification results were then run through an accuracy assessment that considered overall accuracy (OA) as well as the Kappa coefficient. The authors determined that “eCognition obtained the higher accuracy when classifying land-cover classes (OA = 95.7%; Kappa = 0.95; PA = 93.0%; UA = 96.9%)” followed by Orfeo Toolbox/Monteverdi and ArcGIS.
Furthermore, the “best classification map of IAP species, i.e. the result of the second segmentation and classification map, was obtained with the eCognition” with an OA of 92.8%. Again, eCognition outperformed both the other software packages, but with an even greater gap than was observed in the general land cover classification. Orfeo Toolbox/Monteverdi was a distant second with an OA of 63.3% and ArcGIS was the lanterne rouge with an OA of 45.7% for the target IAP classes.
Our results corroborate the effectiveness of the OBIA methods of eCognition in producing high-quality classification maps
Lourenço et al. 2020
It is alway great to see our Trimble products perform so well, which I hope is a reflection of our dedication to quality and accuracy for geo-spatial professionals. I only wish the authors had utilized the most current version of all software packages they examined. Imagine what would be possible if today’s version of eCognition was used with access to the most cutting edge tools! Nevertheless, an interesting read and well documented study on an ever growing topic of interest.