Deep Learning Testing
STEC project
Lorena Abad
March 20, 2021
Deep Learning tests I
With the gully features from 1988 I ran some tests to train a deep learning model based on the RGB imagery The model is a Mask RCNN (details here), which trains a model on imagery based on masked polygons around ground-truth features to detect objects with similar characteristics.
- Export samples
Some samples of the resulting chips:
- Train model
- Detect objects
Considerations at this point:
Results have not been successful so far with the trained models. Further points to test:
- Use DEM and derivatives as input data
- Check how to use polygons as masks to generate labeled chips
- Test different parameters on the literature to adjust the models
Deep Learning tests II
I managed to make the workflow work for the DEM data and derivatives. So far for one derivative at a time, eventually it would be useful to use a combination of distinct derivatives that allow the differentiation of the gully features on LiDAR derived data.
The training is now performed with a combination of all gully features for 1939, 1957, 1960, 1970, 1988, 1997 from Marden et al. 2012, 2014. The preparation of the samples is documented here
I tested the approach with the raw DEM values and with the Terrain Ruggedness Index (TRI) derivative
Results with DEM values
This are examples of the created chips used for training:
The chip size was increased to 512x512 and the reference features were combined to include all the active gullies detected from 1939 to 1997.
The model characteristics show more insightful results compared to the RGB model:
However, applying the model on a subset of the study area was unsuccessful. Running the model for the whole area is still needed, but it is already noticeable that DEM values vary greatly among gully features, and a more standard measure is needed. This is why the TRI is used next.
Results with TRI values
This are examples of the created chips used for training:
