In this project Machine Learning was used for surface reconstruc-
tion to a super-resolution Digital Elevation Model. The resolution was improved
to 10 meters from 30 meters from the fusion of both LIDAR photon data and the
existing DEM
In this project, we wish to improve the clarity of available DEMs, but how do we do this?
To get an intuitive sense, we consider the two pictures given above that depict the same street. However, while the picture on the left is discontinuous and seems to be made of big cubes(pixels), the image on the right is much more smooth with all the details/features clearly visible.
This difference in clarity is due to the massive difference in pixel density. That leads to the fact that the size of each pixel is much smaller for the right picture.
Therefore, the size of the pixels (which we call resolution) is the deciding factor in the clarity of an image.
In this project, we try to refine the existing DEM models using the additional input from LiDAR photon elevations to create a super-resolution DEM.
DEMs have a huge number of applications like-
Engineering and infrastructure design.
Military applications like line-of-sight analysis.
Modeling water flow for hydrology. High clarity DEMs can be used to estimate flood hazards accurately.
Therefore, its relevance cannot be underestimated. However, poor clarity of current globally available DEMs does not allow us to apply them in the above fields.
This is where our project, which aims to improve the resolution of existing DEMs from 30m resolution to a 10m DEM, comes into the picture.
The SRTM is an international research effort that obtained DEMs on a near-global scale. SRTM consisted of a specially modified radar system that flew onboard the space shuttle Endeavour during an eleven-day mission back in February 2000. SRTM gives us near-global access to DEMs of 30metre resolution. Because the SRTM data is outdated(close to 20 years old) and has a poor resolution(30 meters), we need to correct and improve the DEMs we get from here.
The Ice, Cloud and land Elevation Satellite-2(ICESat-2) data-
Processing Photon data
The Photon data came in the form of multiple CSV files, with every point having a confidence code. The CSV files were converted using Pandas in Python and points with only a confidence code of 4 were selected. Once this was done, all the csv files were combined to make one single csv file.
Processing DEM data
Processing of the data was done using GDAL. GDAL helped us convert the data into a form of an array. Later we used GDAL to get the latitude and the longitude of that particular point. We got all the nearest points of a particular data point. We took the mean of all the lidar points lying in the same box
In this project, a method of DEM super-resolution based on a regression model is used. Our approach can effectively extract the feature mapping relationship between a low- and high-resolution DEM (input features) to output features, i.e., the LiDAR data using regression.
The results show that regression has a more significant impact on DEM super-resolution. Compared with the other similar regression based models, Linear Regression significantly improves the DEM’s reconstructed details and recovered textures.
The model is ready to be used for creation of a super resolution and accurate DEM. Our model should make accurate height predictions for all the input data.
These predictions would definitely be more accurate than the Digital Elevation Model data that we have, and should provide height values, as close to the ground-truth as possible.
Contact errijuldahiya(at)gmail.com to get more information on the project