Institutional Repository [SANDBOX]
Technical University of Crete
EN  |  EL

Search

Browse

My Space

Learning from ground penetrating radar data to identify ancient buried structures

Manataki Meropi

Full record


URI: http://purl.tuc.gr/dl/dias/D6292817-745D-4379-9918-F4C6017DB1FA
Year 2021
Type of Item Doctoral Dissertation
License
Details
Bibliographic Citation Meropi Manataki, "Learning from ground penetrating radar data to identify ancient buried structures", Doctoral Dissertation, School of Mineral Resources Engineering, Technical University of Crete, Chania, Greece, 2021 https://doi.org/10.26233/heallink.tuc.89451
Appears in Collections

Summary

GPR data interpretation from archaeological prospection is a tedious and time-consuming process that requires skills and experience. The interpretation process is prone to mistakes, even by the more experienced users. The subsurface can create non-intuitive patterns, making the identification of the buried targets uncertain, requiring additional information from other methods and technologies. Archaeological remains may be bypassed or mistaken for other types of features. Further, residual noise can easily be mistaken as structural remains when in a stripe form that is quite common when surveying in rough terrains. Hence, a system capable of detecting archaeological remains from GPR data could be employed as a guide to assist their interpretation, saving time and reducing mistakes. Recent developments of Deep Learning (DL) and, in particular, Convolutional Neural Networks (CNN) have shown impressive results for similar tasks in other scientific domains like computer vision and medical image analysis. When it comes to GPR data, these methods and approaches have not yet been used to the same extent. The studies dealing with the automatic detection of buried antiquities using CNNs are very few, leaving an ample margin for investigation, and this research contributes towards this direction. In this study, AlexNet architecture is used to train CNN models for classifying GPR C-scans. The latter are 2D images derived from slicing pseudo-3D volumes that can be constructed when collecting data using survey grids. The data used have been collected from 52 archaeological sites located in Greece, Cyprus, and Sicily using a Noggin GPR system equipped with a 250MHz antenna. Data collection was conducted under the framework of research projects of the Laboratory of Geophysical - Satellite Remote Sensing and Archaeo-environment (GeoSat ReSeArch Lab) of the Foundation for Research and Technology Hellas (FORTH). The collected data were processed in MATLAB to export the C-scans. A preprocessing step is followed by applying an overlapping sliding window to crop square subregions of selected C-scans to increase the number of images used for training. Three classes were defined based on dominant features observed in the data: unidentified geophysical anomalies, structures, and noise of stripe form. In total, 18375 examples were selected, 6125 per class. Two datasets were constructed following two different splitting approaches to examine the generalization: a random and non-random one. The CNN implementation and training were performed in Python using the Tensorflow library and Keras API. Two optimizers were tested for each dataset and compared: The Stochastic Gradient Descent (SGD) with momentum and Adam. Tests to improve performance were also made by applying Batch Normalization (BN), Dropout, Image Augmentation, and tuning the learning rate and batch size using the Keras tuner library. Two final models were obtained, one for each dataset approach. The models were evaluated using 100 examples from two archaeological sites that were excluded from the training process.The results showed that the model obtained from the dataset with the random split performed better on the evaluation set, reaching a classification accuracy of 92% over 85%. However, it was observed that the predictions were lacking robustness on similar images. Hence more data and further improvements are required. Further, SGD with momentum performed better but required BN in all five convolutional layers to achieve learning. Dropout improved the results further, but not drastically. Against the expectations, Image Augmentation was not beneficial in any case. While Adam did not require BN for the models to learn, it performed poorer due to overfitting and showed no improvements when BN and dropout were used. The obtained results and good classification performance showed that this is a very promising direction, and the automatic detection of buried structures is a feasible task.

Available Files

Services

Statistics