I am excited to announce that Chiara Zarro, a recent graduate of Unisannio (Università degli Studi del Sannio), in Benevento, Italy, has been awarded the IEEE GRS-Geoscience and Remote Sensing award for her thesis work entitled “Extraction of features on urban area from WorldView-2 images and from LiDAR data through object-oriented classification techniques. Case study: Lioni (AV)”!
Zarro conducted the work during her internship at Mapsat in Benevento through the coordination of Professor Silvia Ullo and has also been granted funding for her doctorate which provides for a period abroad at the DLR (German Space Agency) in Germany.
As you can imagine, Zarro leveraged the Trimble eCognition Developer software for her research. She states in her thesis abstract, that “the objective pursued in this thesis work has been to develop methodologies for extracting features in a semi-automatic way from satellite images to provide the information needed by policy-makers to help them develop correct and sustainable environmental action plans. The idea is to transform the proposed techniques in an industrial product (automatic and unsupervised) to be offered for analysis and decision to interested final users.”
Her work uses the unique data fusion capabilities of eCognition to combine raster data with LiDAR derivatives. She used the LiDAR data to generate DSM and DTM layers and ultimately a nDSM. In addition, a slop map was calculated from the nDSM.
These elevation products were then used as input in the Contrast Split segmentation algorithm to identify “potential buildings”. Two iterations of the Contrast Split segmentation are performed: the first considers the nDSM and creates objects for “ground” and “no ground”; the second round uses the slope layer as input and generates the classes “steep areas” and “no steep areas” – all objects with a slope <15 degrees are assigned to “no steep areas” and all objects with a slope >30 degrees are consider “probable buildings”.
In the next stage of the analysis a Multiresolution segmentation is carried out on the “probable buildings” class to generate detailed objects for the subsequent analysis steps.
Zarro’s method demonstrates a nice example of utilizing multiple segmentations within the rule set to efficiently generate good image objects. These objects then flow into the classification and the authors hierarchical classification approach allows for the use of transferable class descriptions. Zarro uses a combination of NDVI and elevation values to further discriminate between objects and generate an accurate building class.
The transferability of the method was tested by performing the analysis in two different study areas of Lioni and “Users Accuracy values calculated by the error matrix have been 81% and 97% respectively”.
Zarro’s PhD work explores the “reliability of the process of classification on asbestos roofs, where the results on some cases of interest will be compared with some surveys carried out by drones in the same areas”. She is also planning to evaluate combining OBIA classification techniques with deep learning and other knowledge-based methods. I look forward to seeing her results as the strengths of fusing OBIA and deep learning within eCognition become more widespread.