A Workflow for applying a Convolutional Neural Network to Geospatial data. Input data is a multi layer geotiff, output data is a geotiff. This demo is semantic segmentation, however regression is also possible.
If using this code for research, please cite
Google Colab lets you run interactive python scripts (Jupyter Notebooks). This demo has been set up to run by simply executing all cells one after another. To run each cell press Shift-Enter
If you do not want to train the network skip the "Training" section. If you wan to train the network from scratch, skip the cell that loads the pre-trained model.
Lidar data is converted into multiple raster layers. We used LasTools. Specifically LasCanopy. The Lidar point cloud (pcl) was flattened to move all ground points to 0m. The resulting flattened pcl is split into slices at differing heights above ground. rasters where generated using relative point densisty for ground layer ( # of points in slice/ # of total points) and normalized point density(# of points in slice/ # of points in and below slice) with a grid of 4m. for the canopy and upper vegetation structure, percentile heights where used with a 1m grid.
The resulting rasters should be merged into a single Geotiff (future updates may allow for independent files)
Training / Test data is created by drawing polygons using GIS software. Each polygon is labeled with a class (integer) two attributes "Training_Class" and "Test_Class" are used to split polygons between training and test.
It is recommended to mount and store datasets in a google drive folder, this will allow automatic saving of snapshots and output data to a persistent storage location.
