Urban areas are rapidly expanding in developing countries. One of goals of the United Nations Human Settlement Programme (UN-Habitat) is to understand and guide urban development for some developing regions.
Currently, the approaches that UN-Habitat is using cost plenty of workforce, material, and time. Therefore, UN-Habitat is interested in exploring new approaches on how to drive down costs and time, which would not only allow for faster responses but expanding their analysis. Since UN-Habit is already using satellite imagery for urban mapping, our research question is formulated as: Can we develop an automated system that provides valuable information about urban development for the UN-Habitat from satellite image data (e.g. building detection)?
After examining the satellite imagery provided by UN-Habitat and those available publicly (crowd AI and Inria Areial datasets), we define the main task as a building segmentation task. In this research, we study deep learning techniques for building segmentation on satellite image data.
Duo to the fact that the number of images and the quality available for the region of interest (Middle East) for UN-Habitat are insufficient to solely rely on for training. Therefore, we use some public datasets (crowd AI and Inria Areial datasets) for training and evaluation, whose regions and construction practice are different.
Starting with testing several classic segmentation algorithms (FCN8S, SegNet, Deep\_Lab and U-Net), from the experiment results, we find that the performance can still be improved. Then, we propose two novel data reweighing methods, named border weight and inter-building distance weight, to improve the detection performance. By increasing the weights of the pixels outside but close to the border of the buildings, the model is encouraged to learn those information and thus performs better. Inspired by the idea of reweighing the non-building pixels, we investigate whether modifying building pixels can achieve further improvement. We propose a new label representation -- multi-level boundary label that does help to improve the segmentation results. Based on the distance to the building boundary, we can divide building pixels into multiple classes, as their pixel values can be affected by some factors such as trees and shadows. From the experiment result, we can see that the performance is improved since the model captures more information about the buildings. Next, we propose a new neural network architecture that utilizes the two pixel weights, and the multi-level boundary label explained above. Our proposed model achieves state-of-the-art building segmentation performance compared with several classic segmentation methods.
For example, the proposed model's mean intersection of union on the test dataset is 3\% higher than that of FCN8S.
Our model also uses fewer number of parameters (~16 million in total) because we only use the first 13 layers of the VGG16 as the encoder and we do not use any convolutional layers in the decoder part.
The results using the publicly available datasets show that with enough good quality input the building segmentation is possible, hence should be possible in other regions as well.
To see the performance of our proposed model on the UN-Habitat dataset, we train our model with public datasets (crowd AI and Inria Areial datasets) and then use transfer learning to fit the UN-Habitat dataset.
The building detection performance is reduced still good results are obtained. For achieving comparable performance in the region of interest for UN-Habitat more labelled data is needed. Based on the results using the publicly available datasets, we are confident that a comparable performance is attainable.
Regarding the research question, our answer is definitely yes. We not only show that it is possible to obtain information about urban development from satellite image but also propose a new model with great performance in our work.