FORECAST OF POST-DISASTER SCENARIO WITH MACHINE LEARNING APPLICATIONS: EXTENT OF DEBRIS PREDICTION AND FRAGILITY CURVES EVALUATION

In the last few years, the increasing advancement of technologies and Computer Science has permitted the adoption of Machine Learning (ML) techniques in various fields. This work proposes two applications of ML algorithms in civil engineering to forecast features of a post-disaster scenario of a district or a city. Python libraries have been used in both applications and performed on an Intel Xeon @ 2x2.2 CPU and a NVidia Tesla T4 GPU. The former application aims to forecast the extent of debris (EOD) produced by a structure after an earthquake. Eight different ML algorithms have been applied on an input dataset of 300 samples identified by 7 features: (i) structure material, (ii) number of stories, (iii) height of the building, (iv) year of construction, (v) earthquake magnitude, (vi) epicenter distance, and (vii) evaluation direction of debris extent. Dataset has been obtained from images collected into 25 earthquake reports, and the EOD is evaluated by graphic processing, comparing a well-known dimension in the image with debris. After the training and test phases on each algorithm, R-squared and Mean Squared Error values are extracted for comparison. Random Forest (RF) results the more accurate algorithm with the lower MSE value and without overfitting. The latter application aims to obtain structural features from images to evaluate fragility curves of a masonry building. Computer Image-vision methods have been applied for object detection in the images; in particular, openings in building facades are detected. The input dataset is composed of 2000 images collected from OpenImage Dataset. Seven different object detection methods have been applied and accuracy comparison is done by the mAP value. The two-stage ‘Faster R-CNN ResNext 101 FPN’ method gives the more accurate results on the validation set after 6000 iterations and it is used for the detection phase. Detected objects may be visualized on images by bounding-boxes and their location may be extracted as pixel coordinates in tensor format. Several post-processing functions, implemented in Python, are proposed for the evaluation of the opening ratio and the real location of openings. Outputs may be applied to several methods, present in literature, for the evaluation of fragility curve; in particular, a simplified Equivalent Frame Model (EFM) is reported and compared with a FEM investigation.