Automated disaster detection of satellite images
This is my master project applying deep learning model to the satellite images.
The aim of this project is to detect the affected area by natural disasters by using the way of semantic segmentation and change detection method. The model of semantic segmentation in this project is the U-Net which has been widely used semantic segmentation network. The proposed approach achieved an overall accuracy 80% and an F1 score of 55%.
Table of content
- Dataset
- Preprocessing
- Proposed Method
- Build U-Net Model
- Change Detection Method
- Result
- Conclusion
1. Dataset
Training and Validation Data - ISPRS Potsdom dataset(Dataset consists of 36 images and each image size is 6000 × 6000)
Disaster Image Data - DigitalGlobe's Open data platform (The image is taken in Indonesia and the disaster types are Earthquake and Tsunami)
2. Preprocessing
Scaling the training images into the sizes of 6000 × 6000, 2560 × 2560, 1792 × 1792, 1024 × 1024
After the scale task, splitting the scaled images into the small patches size of 256 × 256
3. Proposed Method
The flow of proposed method is as follows:
After the preprocessing task was carried out, the segmentation model was trained using the preprocessed images. Then, the trained model was applied to the pre and post disaster images in order to classify all pixels into 6 classes(Impervious road, Buildings, Low vegetation, Trees, cars, Clutter). Finally, the segmented output images from the trained model were assesed with the change detection method to create the final output.
The detail of way to train the model is as follows:
In this project, four types of methods were used for training models, then comparing each model performance of disaster ditection task and choosing the best model.
4. Build U-Net Model
Model architecture used in this project is as follows:
5. Change Detection Method
After the predicted segmentation masks of pre and post disaster images were created by the trained semantic segmentation model, the impacted regions are extracted to compare these two images by using the pixel-based change detection method with a sliding window. The areas that satisfy below equation can be considered as the damaged area by disasters.
In this project, the threshold and window size are set as 0.7 and 64×64 pixels respectively for the models to avoid recognizing the few changes that occurred due to the time series as the impacted area. The window is slid over the segmentation masks both of pre and post disaster images, and if the area within the window satisfies the above formula, it is labelled as the damaged zone by disasters.
6. Result
The results which trained models applied to pre and post disaster images are shown as follows:
The final results which was applied by change detection method are shown as follows:
The quantitative result is shown as follows:
6. Conclusion
In this project, the approach was presented to identify the damaged area caused by large scale disasters, such as an earthquake and tsunami. To detect such impacted zones from the aerial imagery, the semantic segmentation is performed to all pixels in the pre and post disaster images and the changes occurred before and after the disasters are monitored by comparing those two mapped images. For this approach, the U-Net based model, which is the most major semantic segmentation network, was applied to classify every pixel into either the impervious surfaces, buildings, low vegetation, trees, cars and clutter classes from pre and post disaster images. Furthermore, when training the U-Net based model, the resolution between training and disaster images differed, hence, the models were trained at different scales of training image resolution. After that process, the pixel-based change detection method using a sliding window was implemented to the two predicted mask images in order to extract the impacted zones as a final result.
The results indicated that most of the proposed methods performed the semantic segmentation task well for all classes, besides the clutter class, to the test data of the ISPRS Potsdam dataset. The F1 score for each class were also relatively high. On the other hand, although the method 2 accomplished the highest F1 score of 55% in the damaged class among all methods, it would be considered that this performance could not be enough to apply this model to the real-time disaster assessment task. However, this study demonstrated that the proposed method to combine the different scale images with the training images would be effective for the different resolution unseen images in the semantic segmentation task.