Lab Report

Decent Essays
As mentioned earlier, classification of terrain objects was performed with U-net architecture that used deep convolutional network. The primary focus was to explore various deep learning techniques in classifying objects within a DSM. This section describes all the efforts towards training the U-net architecture for deep CNN.
1. Data processing: Encoding of class labels
The ground truth labels were available in three channels but they were not available in the usual form as a value ranging between [0,255]. Ground truth was contained as a value of [0 or 255] in three channels and the class label had to be inferred from the combination of values across the three channels. Initially, the ground truth
…show more content…
However in Set 2, the Red highlighted cells with value difference greater than Set1 belong to the same class. Though the ground truth label in Set3 seems to be fine, the difference between Red highlighted cells seem to be almost the same as the class label discriminating cells.
Figure 26: NumPy version of input and ground truth
This closeness of values posed a great difficulty in training and necessitated a lot of
‘forgetting of memory’ in the network (by means of dropout) to achieve good results.
In addition, as seen in below figure, there was considerable class imbalance in the DSMs due to disproportionate class distribution in the dataset. This is a typical machine learning problem that misleads accuracy metric. This caused the use of F-measure which considers precision and recall by way of their harmonic mean.
It thus necessitated oversampling for class label 4 which was the least represented class. The last label 5 was ignored in training as it was for clutter and hardly present in the training and test DSMs. Without the oversampling, class 4 classification rate was very poor with an accuracy of just under 10% while other class labels had an accuracy in the range of 40 t0 92 percent. Undersampling was not used as it would have led to discarding of the plentiful useful labels in the other classes.
Figure 27: Distribution of
Get Access