Deep Neural Network Regression for Normalized Digital Surface Model Generation With Sentinel-2 Imagery

[BillFox]

Deep Neural Network Regression for Normalized Digital Surface Model Generation With Sentinel-2 Image... 

Reference:

Print ISSN: 1939-1404

Electronic ISSN: 2151-1535

[ShareUser]

This is a link to an IEEE article that describes approaches using deep learning models for the extraction of high-resolution (<=10m) normalized digital surface models (nDSMs) from low-resolution Sentinel-2 (SAR) data.  Today, high spatial resolution nDSMs cannot be queried globally on an open-source basis.   The approach starts with a neural network architecture based on an enhanced U-net approach. 

U-net is available in ArcGIS, and can be modified using the ArcGIS.learn module.  See How U-net works? | ArcGIS API for Python | Esri Developer.

[ShareUser]

This is a link to an IEEE article that describes approaches using deep learning models for the extraction of high-resolution (<=10m) normalized digital surface models (nDSMs) from low-resolution Sentinel-2 (SAR) data.  Today, high spatial resolution nDSMs cannot be queried globally on an open-source basis.   The approach starts with a neural network architecture based on an enhanced U-net approach. 

U-net is available in ArcGIS, and can be modified using the ArcGIS.learn module.  See How U-net works? | ArcGIS API for Python | Esri Developer.

[ShareUser]

Thanks for sharing this! The use of deep neural network regression, especially the enhanced U-net architecture, for generating high-resolution nDSMs from Sentinel-2 imagery is quite impressive—especially considering the challenge of working with lower-res SAR data.

For those exploring practical implementations, the integration with ArcGIS.learn really opens up doors for custom training and deployment. Definitely worth checking out for remote sensing professionals.

At Community Improvement Group, we're actively exploring AI-driven geospatial solutions, and advancements like this align well with our work in digital terrain modeling and remote sensing analysis.

[ShareUser]

This research presents a highly innovative approach to generating normalized Digital Surface Models (nDSMs) by applying deep neural network regression to Sentinel-2 satellite imagery. The methodology is significant because it offers a cost-effective and scalable alternative to traditional methods like LiDAR, making frequent 3D monitoring of urban and natural landscapes more accessible. By leveraging free, widely available satellite data, this technique can greatly enhance applications in urban planning, forestry, and environmental change detection.