DESTRUCTION FROM SKY: WEAKLY SUPERVISED APPROACH FOR DESTRUCTION DETECTION IN SATELLITE IMAGERY

WSAN Network Flow diagram.

Image showing detection of destruction by WSAN model.

Dataset Link: https://drive.google.com/file/d/180aF7cR4mICNPN5PnR_jLDw1a_vF1V4L/view?usp=sharing

Requirenments:

Keras 2.1.6
Tensorflow-gpu 1.12.0
numpy 1.14.1

Dataset Link: https://drive.google.com/drive/folders/1PhWY3pUtrERSjK88jSEvIW5sFOHqNN7K?usp=sharing
Project Link: http://im.itu.edu.pk/destruction-detection/
Paper Link: http://im.itu.edu.pk/wp-content/uploads/2020/02/id_compressed.pdf

1. Training

Before Training either download Features from this link https://www.dropbox.com/sh/y55nyifuimkzp37/AADjZCESS8W2VKkn8n-dZLWia?dl=0
or run this command:

!python feature_extractor.py

To train model just enter this below command:

!python train.py --trainfeatures_filename trainfeatures.pickle --epochs 500

Recommended iterations = 500

2. Retraining

To retrain our network put this command:

!python retraining.py --model_name Model1_AttentionNetwork_500.h5 --trainfeatures_filename trainfeatures.pickle --epochs 500

It will retrain the first model using Hard negative mining approach to 500 epochs

3. Testing

Before testing the model, there is need to generate segmentation masks.To generate mask enter these two commands one by one:

!python segmentation.py --model_name Model1_AttentionNetwork_500.h5 --test_path Data/test --apply_CRF no

!python segmentation.py --model_name Model2_retrain_AttentionNetwork_500.h5 --test_path Data/test --apply_CRF yes

Now to see Testing results. Enter this command:

!python testing.py --model1_name Model1_AttentionNetwork_500.h5 --model_retrain_name Model2_retrain_AttentionNetwork_500.h5 --features_filename testfeatures.pickle