preview

Essay

Decent Essays

Extensive review of literature helped to better understand MBDA’s requirement of exploring the use of deep learning in terrain classification. It is understood that there is proliferating research due to commercial potential of very high value. 1. Approach As Digital Surface Model dataset was chosen by MBDA for the task, image classification had to be ruled out. The dataset had semantic segmentation ground truth for DSM raster and the task boiled down to interpreting the class label from elevation values contained in each pixel. Though a few semantic segmentation architectures are available, U-net architecture was chosen for implementation. This was due to two reasons: (1) It is currently trending and widely accepted for its high accuracy …show more content…

Therefore, they had to be sampled to generate smaller sub-images (aka patches) with overlapping boundaries; this would avoid loss of spatial information in the boundary. PAGE 26 Figure6: Input down sampling: Besides the yellow bounded area, orange area was also sampled to avoid loss of information at the cropping boundary. 3. Most of the existing U-net implementation available for reference was for binary segmentation and the model training for multi-label segmentation wasn’t easy. Training was very hard due to the huge number of parameters that had to be trained. Therefore, a lot of customisation was needed to realise good results while training and testing. 4. Though the chosen deep learning framework – Tensorflow – is quite popular, it is still in its nascent stages and lacked a few features that made model training gruesome for some complex tasks. For e.g. lack of NumPy like indexing in Python, inadequate support for multi-label implementations etc. consumed a lot of time during model training to tide over the difficulties. 3. Model Implementation The U-net (Ronneberger et al., 2015) extends the typical sliding-window convolutional neural network architecture and aids pixel level classification. The original DSM was downsampled by using a sliding window approach to extract pixels of size 572 x 572 with a stride of 286. This

Get Access