PSI - Issue 44

A. Casciato et al. / Procedia Structural Integrity 44 (2023) 1522–1529 A. Casciato et l./ S ructural Integrity Procedia 00 (2022) 0 0 – 000

1527

6

(a)

(b)

(c)

(d)

Fig. 4. Real and predicted distributions of building types in (a-b) Neuchatel and (c-d) Yverdon-Les-Bains.

3.3. Accuracy of RF model To evaluate the performance of the RF, the Confusion Matrix (CM), also known as error matrix, is obtained. The CM is a × matrix; where is the number of building types. A CM cell indicates the number of test samples for each combination of ground-truth building types ( ) and assigned building types ( ). The diagonal elements show the number of buildings that have been correctly predicted by the RF model. On the contrary, the off-diagonal elements show the number of buildings that have been incorrectly predicted by the RF model. It is important to mention that , the fewer building number for each building type, the harder it will be for the model to predict that building type, due to the scarcity of training examples. For this reason, more errors are expected for the M4 and RCF building types, which have the lowest contributions in the dataset. Fig. 5 shows the CMs obtained from the RF models trained and tested on the databases of (a) Neuchatel, (b) Yverdon-Les-Bains, and (c) Neuchatel and Yverdon-Les-Bains together. The results of the RF models are evaluated through the accuracy measure AM1, that is the overall accuracy of building types, which is based on the confusion matrix. AM1 is calculated as the ratio between the number of correctly classified buildings over the total number of buildings, as given by (1). 1 = ∑ ∑ ∑ (1)