1313 Rate this article:

Student Spotlight Fall 2021

Andrew Fore

ENVI Deep Learning Helps Map The World

Bujar Fetai, the newest L3Harris Geospatial Student Spotlight, is a PhD student in the Department of Geodesy, and is on the faculty of Civil and Geodetic Engineering, University of Ljubljana, Slovenia. Fetai is passionate about researching land administration systems and his research focus is on improving cadastral maps. As part of his PhD, he is working on automating the process of delineating visible land boundaries from UAV imagery. “ENVI® is an important tool for my research. I use ENVI to process and analyse all remote sensing data, including UAV imagery.” 

Fetai’s research on visible land boundary detection with ENVI Deep Learning (featured below) was recently published in Remote Sensing for Land Administration 2.0. (Fetai, B.; Račič, M.; Lisec, A. “Deep Learning for Detection of Visible Land Boundaries from UAV Imagery” https://doi.org/10.3390/rs13112077). “In terms of deep learning, ENVI plays a crucial role and can be considered as a rigid mapping approach that does not require programming. The latter is very important as not all land administrators or managers have programming skills and ENVI Deep Learning is a great alternative."

Fetai holds a Master of Science in Geoinformation Science and Earth Observation from the Faculty of Geo-Information Science and Earth Observation at the University of Twente in the Netherlands. His research interests include land tenure, cadastre, geographic information systems and UAV and satellite image analysis. His PhD research is under the supervision of Dr. Anka Lisec, and collaborates with Dr. Krištof Oštir and Dr. Mojca Kosmatin Fras in the field of image processing, remote sensing and photogrammetry.

An example of visible a cadastral boundary.

The State of the World’s Cadastral System

More than 70% of the world's land rights are unregistered and not part of a cadastral system. This makes establishing land tenure difficult, and in many cases impossible. Mapping the boundaries of land rights, creating a complete cadastral system, and being able to keep it up to date, is a top priority in land administration. Multiple challenges exist to achieving this priority, however, low cost and fast cadastral survey and mapping techniques can overcome many of these.

Indirect mapping techniques are based on delineation of visible cadastral boundaries from high-resolution remote sensing images. The application of image-based cadastral mapping is based on the recognition that many cadastral boundaries coincide with natural or man-made boundaries, such as hedges, land cover boundaries, building walls, roads, etc., and can be easily identified from remote sensing or UAV imagery.

Apart from high-visibility UAV imagery, many previous case studies have reported manual delineation and a limited number of studies have investigated the automatic approach to extract visible cadastral boundaries. Mainly, customized image segmentation and edge detection algorithms have been used to automate cadastral mapping.

ENVI Deep Learning Makes Quick Work of Cadastral Mapping

The application of deep learning for visible cadastral boundary extraction is becoming increasingly important, especially for UAV-based cadastral mapping. As part of his PhD thesis, Fetai used ENVI Deep Learning to automate the extraction of visible cadastral boundaries from UAV imagery. For training and validation of the model, three different UAV image datasets from different locations with scenes from rural areas were used. The training of the model was patch-based.

The results for the test UAV images are shown in the figure below.

Left: Testing UAV imagery; Middle: Ground truth visible boundaries; Right: Extracted visible boundaries with ENVI Deep Learning - ENVINet5

The large amount of training data and data preparation required for deep learning can prove to be a major hurdle. However, Fetai applied a data augmentation technique which was possible using ENVI Deep Learning (version 1.1.2). The raster labels were created in ENVI by uploading reference cadastral boundaries as Regions of Interest (ROIs) – a quick approach compared to manually creating raster labels.

Workflow for the detection of visible land boundaries based on the ENVINet5 model – ENVI Deep Learning


Testing the Model In Slovenia

The model was tested in a case study in Odranci, Slovenia. The results showed that ENVI Deep Learning has the potential to extract visible cadastral boundaries, especially in rural areas, resulting the an 84% accuracy rate. The predicted boundary maps were georeferenced and no additional post-processing step was required.

(a) UAV imagery for Odranci – Slovenia, divided into areas for training and testing. (b) UAV imagery for Ponova vas — Slovenia, used for training. (c) UAV imagery for Tetovo — North Macedonia, used for training. (a, b, c) 0.25m ground sample distance.


What Findings Mean for Cadastral System Going Forward

“These findings indicate that ENVI Deep Learning can be effectively used as a rigid mapping approach,” said Fetai. “Because ENVI Deep Learning does not require any programming, this approach can become an important tool for land administrators who are usually not skilled in programming.”

Boundary maps predicted with ENVI Deep Learning could be used in developing regions with low cadastral coverage to speed up cadastral mapping. This will be particularly useful in rural areas where the visibility of cadastral boundaries on images is higher, compared to dense urban areas. In more developed regions where a complete cadastre already exists, ENVI Deep Learning could be used to automate the revision of existing cadastral maps to automatically define areas where cadastral maps need to be updated.

However, it should be noted that the automation of cadastral boundaries is not at an end. This is due to the complexity of cadastral boundaries which can have very simple geometry, but can be difficult to interpret. The predicted visible boundaries should be further validated by landowners and other beneficiaries before they are considered as final cadastral boundaries. However, the boundaries extracted using ENVI Deep Learning are accelerating the pace towards automation of cadastral mapping, and the extracted boundaries can be considered as preliminary cadastral boundaries.

ENVI Deep Learning is commercial off-the-shelf technology that is specifically designed to work with remotely sensed imagery to solve geospatial problems. For more information click here or email GeospatialInfo@L3Harris.com.


You could be the next L3Harris Geospatial Student Spotlight!





Students, Researchers and Instructors


Using ENVI and IDL for research and to prepare students for successful careers.


ENVI Software


Process and analyze all types of imagery resources, and data



Extract meaningful visulizations from complex numerical data.