multiclass-multiband-classification of satilite data

This repository shows you how to handle multiband satellite data and train convolution neural network to predict 10 different types of surface.

PREPARE THE ENVIRONMENT

• install python 3.6.7 or higher is required
• install requirements.txt
• add absolute path to src root to .env (.env.example for example)

DATA

The dataset is based on Sentinel-2 satellite images covering 13 spectral bands and consisting out of 10 classes with in total 10,000 labeled and geo-referenced images. link to original paper

easy download from google drive

PREPARE THE DATA

• download the data to src root
• run the data_prep.ipynb in src root
  • train dataset shape is (7500, 12)
  • validate dataset shape is (1500, 12)
  • test dataset shape is (1000, 12)
• to visualize the data with RGB bands run the data_vis.ipynb

• to visualize all bands run the "notebooks/generator test.ipynb"

TRAIN THE MODEL

There are few experiments:

• costume resnet for multispectral data - "exp_resnet" (test metrics precision is 0.81, recall is 0.8, f1 is 0.8)
• simple convolution nn - "exp_main" with ReduceLROnPlateau learning rate annealing. (test metrics precision is 0.91, recall is 0.91, f1 is 0.91)
• simple convolution nn - "exp_cos_lr_ann" with CosineAnnealingLR learning rate annealing.
  • Both works pretty well, but with CosineAnnealingLR training is longer.
• unet + fully connected layers - "exp_unet_fcl.2". (test metrics precision is 0.92, recall is 0.92, f1 is 0.92)

to run the training:

• run the train.py. Note that augmentation does not work with Sentinel-2 multispectral data.
• run tensorboard:
  • enter src folder
  • run command "tensorboard --logdir logs/scalar"

TEST THE MODEL

to run the testing:

• run "notebooks/model test.ipynb"
• to run resnet based model, first download weights from gdrive
• to run unet based model, first download weights from gdrive

Metrics on test dataset: precision is 0.92, recall is 0.92, f1 is 0.92

Confusion matrix on test dataset:

Easy to notice model struggling with vegetation recognition and industrial surface.