Improvements - Githubissues

ainazHjm / LandslidePrediction

Classification task for predicting landslides based on GIS maps using locally aligned convolutional neural networks. This work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 3.0 License.

https://creativecommons.org/licenses/by-nc-sa/3.0/

37 stars 11 forks source link

Improvements #15

Open ainazHjm opened 5 years ago

ainazHjm commented 5 years ago

[x] train the standardized features on the new data with no-data points for Veneto without oversampling
- [ ] train the standardized features on the new data with no-data points for Veneto with oversampling
[ ] train the rotation dataset for the rotation model without oversampling
[ ] train the rotation dataset for the rotation model with oversampling
[ ] train using multiple regions

ainazHjm commented 5 years ago

I trained the FCNwPool model on the new dataset for Veneto (with no-data points) without oversampling but the results are not good. It doesn't seem like a coherent susceptibility map and I don't know the reason why. I'm starting to think that my model might not be a good model to use for this task. I'm going to try this model pixel wise as well. to see how the results change. Results on validation set: FCNwPoolMultiRes_test_noOversample Results on train set: FCNwPoolMultiRes_train_noOversample

ainazHjm commented 5 years ago

I found a major bug in my data: my slope feature (which is between 0 and 360) has large real values that I thought I was normalizing but I wasn't! So I guess one reason that the model is doing worse than anticipated is this. Rest of my data is categorical that I'm using them as one hot. I processed the dataset again and normalized the slope features. I'm going to retrain the previous model (FCNwPool) on both large images (200x200) and pixel-wise and see how the results change. One good thing about the pixel-wise training is that I can actually oversample the landslide points.