Awesome

UNet CNN

UNet model to segment building coverage in Boston using Remote sensing data

I implemented the UNet Convolutional Neural Network architecture for segmenting the building footprints from the image. Here I aimed at classifying each pixel representing building footprints. It is important to note that I specifically focused on the buildings and did not consider any other features like roads, bridges, etc. I used high resolution aerial image provided by the USGS.

Data preparation and methods

I used a cloud-free 2D- high resolution latest available aerial image to create the training dataset. One advantage of using 2D Aerial data over google images was in generating a binary building mask layer. The image was georeferenced with a constant scale so it would also help in precisely estimating the building coverage in metric units.

To create a binary raster (mask), I used reference vector building layer (ground truth) available on the ArcGIS online platform. Then I adjusted the image and vector layers to the same coordinate systems and converted original RGB layer to corresponding binary (Building:1, Other:0) data. I clipped the RGB and Masks images iteratively with (128,128) image dimensions. I used a threshold of 20% on the building coverage in the mask layer and considered only images with sufficient building coverage. After applying the threshold value, size of the data reduced to 792. I have split this data into a train (80%), test (10%) and validation(10%) sets. I then normalized, and shuffled the dataset.

This processing ended up in a very small dataset. I thought of using a substantial augmentation on this training dataset to maintain reasonable amount of images during training phase. Unfortunately, the training on augmented data did not progress well. So I ended up implementing the UNet model on this small dataset.

I trained the UNet model from scratch. I changed the dimensions of the input images to 128,128. As used in a similar study by Chorr et al, I modified the padding to the ‘same’, to avoid shrinking when doing convolutions. Also, I have added the batch normalization after each ReLU activation and used Adam Optimizer for faster convergence during training phase. I did training on smallbatches of 8 images with 50 epochs. I used Keras callback ReduceLROnPlateau to reduce the learning rate by a factor of 0.1.

Future work To address the problem of overfitting, the data augmentation and resampling procedures like kfold-cross-validation would help. Moreover transfer learning would speed up training and improve the performance of the model.

We can notice in the figure-3, the ground truth mask layer is not matching to the building features in the latest Aerial image. This mislabeling might have led to suboptimal performance during the test phase. It would be interesting to see the results with upto-date mask layer. It would be interesting to compare the performances of other states of the art image segmentation architectures like Mask R-CNN to accomplish the feature segmentation task.

In this analysis, I focused on data processing and basic semantic segmentation modeling. Once the model is tuned with a sufficient dataset then it can be used to estimate building footprint coverage in Boston.

References

Chhor, Guillaume, C. Bartolome Aramburu, and I. Bougdal-Lambert. "Satellite Image Segmentation for Building Detection using U-net." Web: http://cs229. stanford. edu/proj2017/final-reports/5243715. pdf (2017).