Figure 11. Confusion matrices of Te

Figure 11. Confusion matrices of Test image 4, Test image 5 and Test image 6 classification results by F-CNNs and SVM: Confusion matrix listed the accuracy measures.

The classification results show that our proposed model is able to detect flood pixels compared to SVM classifier. The accuracy measures in Figure 10a for Test-1 shows that the recall rate of flood water class is 81.7% which means that the classification model able to detect 81.7% flooded pixels accurately. Compared to F-CNNs, the conventional SVM classification only detects 23.8% (Figure 10b) flood pixels accurately. Much of the flooded pixels are classified as land or non-water by SVM classifier which lowers down the precision rate of non-water class to 49.6%.

Similarly, for Test-2 (Figure 10c) and Test-5 (Figure 11c) we have also observed that the F-CNNs model able to obtain 95.4% and 76.95% of recall rates respectively compared to only 0.10% and 26.36% of recall rates by the SVM classification method.

Both the classification model fails to detect the permanent water from Test-1 image. While F-CNNs model detect permanent-water pixels as flood water (Figure 9(C-1)), SVM classifier misclassifies a considerable amount of flood water pixels and permanent water pixels (Figure 9(D-1)) as non-water class.

The classification results (Figure 9(C-2,D-2)) of Test image-2 (Figure 9(A-2)) show that the F-CNNs model distinguishes between flood water and permanent water areas with 95.4% recall rate for flooded area detection (Figure 10c) while SVM classifier classifies the entire flooded areas as permanent water features and achieved as low as 0.10% recall rate (Figure 10d).

Both the classification methods on Test-5 achieved with an overall accuracy less than 50%. However, the overall accuracies obtained by F-CNNs model ( 45.14%) is higher than the overall accuracy obtained by SVM classifiers (10.60%).

However, the F-CNNs model does not able to achieve more than 70% overall accuracy level for every classification tasks, but it is clear from the results that the model is able to distinguish flood water from permanent-water features that the SVM classification method is not able to obtain as we observed in Figure 9(D-2) and Figure 9(D-6).

Accuracy level of non-water area detection from all test images for both the classification method are showing more than 50% accurately classified pixels except for Test-6 where the SVM classification results show (Figure 9(D-6) and Figure 11f) all the non-water pixels are misclassified as flood waters.

The overall classification performance also show that F-CNNs model achieves classification accuracy higher than SVM classifiers except in case of Test- 3 classification performance where overall accuracies of both the classifiers are more or less similar (overall accuracy 57.71% for F-CNNs classifier and 58.34% for SVM classifier).

Finally, the processing time of the SVM classifier is also another important factor which makes the SVM classifier lagging behind the F-CNNs classification model. The processing time of SVM-classification method for test-1 and test-2 images are 0.45 h and 2.86 h and F-CNNs took 1.05 min and 3 min respectively. The experimental results and accuracy measures therefore, indicates that the application of neighbouring information with fully convolutional neural networks approach can be applied on a more generalised basis compared to conventional pixel-based classification methods. The model also able to distinguish between flood water and permanent water if there is enough spectral variability exists between these two class types.