This projet entails the application of deep learning algorithms i.e. CNN to detect deforestation using satellite imagery.
I analysed Sentinel-2 images (cloud free) of three different regions in the State of Mato Grosso in Brazil and derived a total of 5,122 image chips and masks of size 256 * 256. See a sample of training images and masks in the sample data folder.
Image chips were flipped left and right before fed into the network for training. See preprocess.py
Unet model was implemented by using MobileNetV2 as the feature extraction with pre-trained weights.
More on Unet model U-Net: Convolutional Networks for Biomedical Image Segmentation
The model was trained first using the pre-trained weights of the base model for 15 epochs with early stopping. The best result was achieved by re-training the whole model for 20 epochs and attained a binary accuracy of 0.9780.
- tensorflow
- keras
- numpy
- matplotlib
- pandas
- cv2
