LBP Features:
For the local binary patterns, we used the ‘uniform’ method which is grayscale invariant and rotation invariant. We used a radius of 1 (i.e. we calculated the LBP for each pixel using only the 8-neighbourhood).
LBP Histograms: We flatten the LBP features and calculate the histogram on it using 10 bins.
Classical Models:
1. K-Means: (K=2, method='uniform')
The LBP histograms were used as the feature used to train the model.
Precision (UA)
Recall (PA)
F1-Score
Accuracy
OE
CE
Flooded
0.70
0.86
0.77
-
0.14
0.30
Non-Flooded
0.81
0.64
0.72
-
0.36
0.19
Macro Avg
0.76
0.75
0.74
0.75
0.25
0.24
2. SVM:
It used the LBP features from the training set as training data.