numbacruncha / roaddetection

Detection of roadways from satellite data

Road Detection

Detection of roadways within satellite data
Data provided: 45 training examples, 10 test images

Problem Statement

Derive a mask depicting roadways from each 3200 x 4800 satellite image tile.

Methodology

This is an image segmentation problem requiring pixel-wise binary classification of the input image resulting in a two-class output image. An extension of a fully connected convolutional network (FCN) was selected to solve this problem as it is suitable for use with a small number of training examples. A total of 8 false positives were removed from the dataset during preprocessing.

Solution Architecture:
U-net provides good segementation capability using a symmetrical downsampling and upsampling network of layers which is more efficient than an FCN. A pre-existing open-source implementation (tf_unet) was utilised for this purpose.

Optimisation:

• Input images were scaled from (3200, 4800) down to (400, 600) to circumvent resource constraints.
• The number of layers in the unet was set 4 with 48 features, which provided good performance.
• Batch normalization was added to the convolutional layer to scale and normalise the input arrays resulting in faster convergence.

Results

Training occured over 64 iterations across 50 epochs. The network was trained using an RTX2070 GPU in approximately 90 minutes resulting in an average accuracy of 98%.

Model training Accuracy:

Mask creation from the training set:

Creating a mask from the test set using the pre-trained model (no ground truth mask):

Improvements:
It would be desirable to avoid downsampling the training images in order to capture more details during learning, as well as to output a mask at the same resolution as the training images.

In order to increase the power of the network, it would be beneficial to train on multiple GPUs. Horovod - a C++ library developed by Uber for distributed tensorflow operations (https://github.com/uber/horovod) would be worth investigating to this end.

An implementation of U-net with iHorovod integration built-in would be a logical choice for testing alongside the proposed solution. https://github.com/ankurhanda/tf-unet

roaddetection's Issues

test_results

accuracy

training_result