New routes to affordable land administration

According to estimates, 70% of the world’s population cannot access a legal land administration system. This often makes it hard to determine which land belongs to whom. It also causes problems in vital areas such as productive farming, urban planning, and sustainable development. That’s why the UN incorporated the formal registration of property rights, ownership, and value into the sustainable development goals (SDGs) of 2030. 

Fit-for-purpose land administration

Most rich countries have a highly accurate land administration system, created using ground-based survey methods. For most mid- and low-income countries, however, this approach is too labour-intensive and expensive. To remedy this, the concept of fit-for-purpose land administration (FFPLA) has been introduced and tested. FFPLA focuses on identifying visible cadastral boundaries and delineating them with remotely sensed data. Deep learning and remote sensing imagery can play an important part here. 

Faster and cheaper

FFPLA is based on the notion that land administration in mid- and low-income countries should be adapted to the needs of the inhabitants instead of focusing on accuracy and top-end technical applications. Cadastral boundaries are mapped by defining general boundaries from satellite, aerial, and unmanned aerial vehicles (UAVs) imagery. These sources make it easy to delineate visible cadastral boundaries that coincide with physical objects such as buildings, fences, hedges, rivers, and roads. The main advantage of this approach is that it's faster and cheaper than using ground-based methods. 

Automating things

The principles of FFPLA have resulted in a growing demand for automated extraction of visible cadastral boundaries. According to estimates, 30-60% of an initial land administration project is spent on cadastral mapping. Therefore, expenses and time could be saved by automating a part of the cadastral mapping process. However, manual and traditional methods tend to take lots of time and effort in terms of setting the required parameters. 

Deep learning

A way to overcome labour-intensive parameter setting is using deep learning. Deep learning doesn't rely on predefined segmentation parameters but learns the required information from the input data fed into it. In recent years, deep learning methods have proven to be very successful at transforming input data into actionable output data. Compared with traditional methods, the most significant advantage of deep learning models is that output is generated in a supervised way and not handcrafted by a human operator.  

Different architectures

In the paper, researchers briefly describe several deep learning architectures commonly used in this regard: convolutional neural networks CNN), fully convolutional networks (FCN), and vision transformers (ViT). Several studies would lead to suggest that the latter often outperform the convolution-based models.

Conclusion

The researchers note that, up until now, few studies have really investigated deep learning combined with remote sensing imagery. What studies there are seem to show that this approach could significantly improve the automated extraction of visible cadastral boundaries. The authors’ final remarks and recommendations include the following: 

• Using vector format instead of raster output saves (postprocessing) time.
• Specific benchmark datasets for cadastral boundary delineation (non-existing so far) would help the community to train and test state-of-the-art models. 
• Studies have been based on (too) small datasets, which raises the question of whether trained algorithms would perform in other geographic locations as well. 
• Studies indicate that automated extraction methods will likely be more successful if a large portion of the cadastral boundaries is visible. Although there appears to be a substantial overlap between cadastral reference data and topographic objects, no large-scale research has been executed on the physical properties of cadastral boundaries. 

This is a teaser of a journal article entitled “Cadastral boundary delineation using deep learning and remote sensing imagery: state of the art and future developments”