Find Airports

Semantic segmentation of airports from satelite images using deeplabV3+ and resNet50

Check it out

Introduction

Accurate geodata is required when developing tomorrows telecom network, where an increased demand are seen in several fields such as video, 5G and IoT solutions. By using geodata, it is possible to plan and design telecom networks with optimal performance. Telecom networks are often overloaded in crowded areas, and by localizing social structures that contains a lot of people, it is possible to distribute resources more efficiently.
Together with Vricon we have developed a product utilizing deep learning convolutional neural networks to efficiently detect and mark airports in satelite images. This webpage describes the task tackeled, how the challange was accomplished and what results were obtained.
Welcome to the future of computer vision.

Concept

Our product constitutes a way to efficiently train deep neural networks to perform semantic segmentation on satelite images. The product is modular, and can easily be adapted to detect more types of objects and landmarks.
Provided is also pre-trained networks, based on the award-winning ResNet50 and DeepLabV3+ designs, which are out-of-the-box capeable of detecting airports from satellite images. Classified Atlanta

Above is shown an image describing the concept of semantic segmentations. Illustrated is a satellite image, which has been fed to a deep convoluational neural network, and classified to several diferent types of objects and landmarks, reining from rivers and images to roads and bushes.

DeepLabV3+

DeepLabV3+ is a deep convolutional neural network designed to perform Semantic Segmentation. It makes use of Spatial Pyramid Pooling to preserve semantic information. We have implemented DeepLabV3+ in our system, designing it to perform Semantic Segmentation on satellite images, with focus on detecting and classifying airports. A resulting image when using the system on a satellite image is visible below. Classified Atlanta

As ilustrated in the above image, our implementation of DeepLabV3+ can successfully mark an airport on the given map. It does so with an F1-score of 98.3 % and 75.3 %, for non-airport and airport pixels respectively, resulting in an average F1-score of 86.8 %.

ResNet50+

ResNet50 is deep neural network designed to perform image classification. It utilizes Residual Skips to achieve more efficient training and higher accuracy. We have modified the original design to perform semantic segmenation instead of image classification, while still making use of the benificial residual Skips. Below is shown a resulting image from applying our implementation to a satellite image. Classified Atlanta

As visible above, our implementation of ResNet50 is clearly able to mark an airport in a given satellite image. It does so with an F1-score of 97.6 % and 58.9 %, for non-airport and airport pixels respectively, resulting in an average F1-score of 78.3 %.

More

If you wish to know more about the project and the product, we recommend checking out our report or our poster.

Conclusion

The conclusion that can be drawn from this project is that airports are a challenging type to classify, possibly because of their variety in appearance as well as similarity to other social structures. The experiment does however show that it is possible to train these models to find larger airports, but with an insufficient reliability. The best average F1-score for DeepLabV3+ achieved a result higher than the required of 0.80, while the ResNet50 implementation did not. To get a more accurate classification it is obvious that a much large set of training data is required, not at least when training a neural network that has millions of parameters.

Universitetsvägen 1

58431, Linköping

Email

jakgr008@student.liu.se

Phone

+46xxxxxxxxx