Awesome

MSc Data Science Dissertation/Project

An Attention-Based U-Net for Detecting Deforestation Within Satellite Sensor Imagery. https://www.sciencedirect.com/science/article/pii/S0303243422000113

Datasets

Amazon 1 (Regular 3-dim Dataset) -- https://zenodo.org/record/3233081

Amazon 2 (Larger 4-band Amazon and Atlantic Datasets) -- https://zenodo.org/record/4498086#.YMh3GfKSmCU

Files

• dataset -- Folder of original dataset from Regular Dataset.
• figures -- Figures for report (amazon-atlantic-forest-mapjpg.jpg from https://pubmed.ncbi.nlm.nih.gov/20433744/).
  • shapefiles -- Shapefiles for map. Amazon Shapefile from: (http://worldmap.harvard.edu/data/geonode:amapoly_ivb), rest from: (http://terrabrasilis.dpi.inpe.br/en/download-2/).
• models -- Folder of each of the three types of Attention U-Net model; load into Keras using 'load_model([modelfilename])'.
• metrics -- Folder of metrics (accuracy, precision, recall, F1-score) for each result.
• Experimentation.ipynb -- Jupyter notebook of data processing, augmentation, model training and testing.
• Figures.ipynb -- Jupyter notebook of figures found in figures.
• predictor.py -- Takes any input RGB or 4-band image and outputs Attention U-Net-predicted deforestation mask to file.
• preprocess-4band-amazon-data.py -- Python script to preprocess GeoTIFFs from 4-band Amazon Dataset and export as numpy pickles.
• preprocess-4band-atlantic-forest-data.py -- Python script to preprocess GeoTIFFs from 4-band Atlantic Forest Dataset and export as numpy pickles.
• preprocess-rgb-data.py -- Python script to preprocess data in RGB Dataset and export as numpy pickles.
• requirements.txt -- Required Python libraries.

How to use

Obtaining Attention U-Net Deforestation Masks

• Run pip -r requirements.txt to install libraries.
• Download 'unet-attention-3d.hdf5', 'unet-attention-4d.hdf5' and 'unet-attention-4d-atlantic.hdf5' models, and place in same directory as script.
• Run 'python predictor.py [MODEL IDENTIFIER] [INPUT IMAGE PATH]' or 'python3 predictor.py [MODEL IDENTIFIER] [INPUT IMAGE PATH]'.
  • Model identifier for RGB is 1, 4-band Amazon-trained is 2, 4-band Atlantic Forest-trained is 3.
  • e.g. Get mask prediction of image named 'test.tif' from 4-band Amazon model: 'python predictor.py 2 test.tif'.

Obtaining Pre-Processed Data

• Run pip -r requirements.txt to install libraries.
• Run 'preprocess-4band-amazon-data.py' to pre-process 4-band Amazon data.
• Run 'preprocess-4band-atlantic-forest-data.py' to pre-process 4-band Atlantic Forest data.
• Run 'preprocess-rgb-data.py' to pre-process RGB Amazon data.