This is a repository demonstrating how to detect clouds in Landsat 8 images via semantic segmentation and a UNET CNN. We use the SPARCS Dataset to train and validate the model.
Ideally this would be run in a docker container, but I do not have Windows Professional on my personal computer, and so I cannot activate Hyper-V. As for using docker through WSL, CUDA is embarassingly slow via WSL, and so is not appropriate for GPU-based neural net computations. So we use conda in Windows.
You can obtain the project environment via:
conda env create -f requirements.yml
conda activate cloud-detection
Hardware Environment
- CUDA 11.6
- CUDNN 8.3
- RTX 2070 Super, 8 GB VRAM
Use the demo notebook detect_clouds_tensorflow.ipynb. You must have unzipped the SPARCS data into the ./data folder
- Need k-fold validation (would use k=5 and ensure each image ends up in a single fold test set)
- Would employ hyperparameter optimization
- Different achitectures could be examined (for example, Mask R-CNN)
- Ensemble could be created using multiple networks
- Confirm that CNN performs better as compared to pixel-based classifiers
- Could try different loss functions like stochatic gradient descent
- Some labels look quite poor, figure out when these are the culprit
- Identify under which conditions the model performs most poorly
- Test the elimination of classes that are not clouds or cloud shadows
- Data augmentation (more examples)!
- Learning rate decay tuning
- Fix the missing edge of all scenes
- Make implementation more OOP
- Make tests