Copyright © Panorama Group 1991 - 2022
"Panorama Vision" automatic recognition and vectorization system
The "Panorama Vision" system is designed for automatic recognition and vectorization of data from satellite images, aerial photographs, and UAV data. The system is based on artificial intelligence (AI) using trained neural networks. "Panorama Vision" can recognize contours of field plots, hydrographic objects (rivers, ponds, lakes, and others), buildings, and structures. The system can be accessed through a client module in the professional GIS "Panorama". The client application creates data processing requests to the server part of the system, hosted on KB "Panorama" technical facilities. The recognition process uses either user-provided images or data from open sources. The spatial resolution of the data must be at least 10 meters/pixel. Image processing results in a vector map with contours of recognized objects, such as agricultural field boundaries.
The "Panorama Vision" system performs semantic segmentation of objects in images using deep neural networks based on attention mechanisms. The interpretation models were trained using satellite images and aerial photographs from various regions of the Russian Federation, along with vector data created by professional cartographers. The total volume of training data exceeds 500 GB. The recognition accuracy of all models on test images is approximately 90%.
For working with artificial neural networks, the "Panorama Vision" complex uses:
To accelerate computations, GPU resources with NVIDIA CUDA technology support are utilized.
To process data, the user needs to perform several actions:
As a result of data processing, the complex will generate a layer with a vector map. Data is transmitted to the client via WFS protocol. Users can copy or edit the resulting map, as well as convert it to other formats.
The processing stages consist of two main parts:
Client-side:
Server-side:
Data recognition examples
For instance, recognizing building contours from satellite images of Moscow with 0.6 meters/pixel resolution within an area of 25 square kilometers took 2 minutes and 20 seconds. The processing resulted in the creation of 2,895 objects
When recognizing hydrographic features, such object types as lakes, ponds, rivers, and streams are identified. For training the neural network on hydrographic objects, images from forest, mountainous terrain, urban development, steppe, and agricultural territories of the Russian Federation were used. To generate hydrographic contours covering a total area of 100 km2 and raster data of 80 MB, it took 1 minute, resulting in the creation of 416 unique objects.
Source image
Layer with raster field masks
Layer with final vector map
For creating field contours, the agro10t.rscz, classifier is used, objects are created with code 79000002 - agricultural land boundaries. For creating hydrographic object contours, the map10000.rscz, classifier is used, objects are created with code 31410000 - rivers, streams. For example, it took 2 minutes to form field contours with a total area of 452 km2 and raster data of 112 MB. As a result of processing, 758 unique fields were created.
p.u. - processing units
m/p - meters per pixel
Maximum data volume available in the tariff (km2) = Available volume (p.u.) / Cost per 1km2 (p.u.)
Image Resolution |
|||||||||
Model Type | 10 m/p | 5 m/p | 2 m/p | 1 m/p | 0.5 m/p | 0.25 m/p | 0.1 m/p | 0.05 m/p | 0.01 m/p |
Cost of field contour recognition per 1 km2 in p.u. | 1 | 1 | 1 | 1 | 1 | ||||
Cost of hydrographic objects recognition per 1 km2 in p.u. | 1 | 1 | 1 | 1 | 1 | ||||
Cost of buildings/structures recognition per 1 km2 in p.u. | 5 | 5 | 5 | 5 | 20 | 20 | 20 |
High quality (90% Accuracy) | |
Medium quality (75% Accuracy) | |
Below average (50% Accuracy) | |
Not supported |
*When calculating, the area is always rounded to whole numbers, for example 0.12 km2 will be rounded to 1 km2, 1.37 km2 will be rounded to 2 km2.