Identifying Qanats and their Impact on the Formation of Zuzan City using Convolutional Neural Networks

Abstract
Water resource management in arid and semi-arid regions has been a paramount goal for humanity in shaping and sustaining urban and rural communities. Qanats, in these regions, are among the most significant systems for water supply and distribution. These underground systems are typically identifiable only by their shafts, which appear as circular depressions on the ground surface and in aerial and satellite imagery, revealing the path of the qanat. To save time, automated methods can be employed to identify these shafts and the qanat network, often involving neural networks and machine learning techniques. This paper provides empirical evidence for the application of deep learning techniques to automatically extract archaeological information from historical aerial images. The specific goal of this research is to accurately detect qanats in the historical landscape of the city of Zuzan using convolutional neural networks and the YOLO object detection algorithm, with aerial images from the 1940s and 1970s serving as input data. This study, employing a computational analysis of neural networks, aims to identify qanats in Zuzan to investigate the city’s formation and answer the following questions: 1) What is the process and accuracy of the convolutional network with the YOLO algorithm in detecting qanats in aerial images? 2) How have the qanats of the Zuzan plain, as the sole source of water supply and distribution, influenced the formation of the city of Zuzan. The proposed model was trained with 80% of the data for training and 20% for validation, with 200 epochs and a learning rate of 0.01. The results obtained from training the convolutional network with aerial images from the 1940s and 1970s and the YOLOv8 algorithm demonstrate the effectiveness of aerial data in deep learning for the automatic extraction and identification of four qanats leading to the city of Zuzan in different directions simultaneously, with an accuracy of 94%. In this region, these qanats were the primary factor in the formation and survival of the city of Zuzan during historical and Islamic periods, and they continue to operate today.
Keywords: Zuzan City, Qanat, Convolutional Neural Network (CNN), YOLO.
 
Introduction
Artificial intelligence has garnered increased attention in various scientific fields in recent years. Specifically, for archaeologists, AI enables them to access a vast array of archaeological data, thereby aiding in informed decision-making regarding field activities such as excavation, survey, and conservation. The identification of targets (objects) in satellite and aerial imagery using image processing techniques is a primary goal in the development of computer systems. In computational archaeology, machine learning-based methods for the automatic detection of archaeological evidence on remote sensing data have been a consistent focus in recent years. Image processing using machine learning has become one of the most widely used applications for archaeologists in recent years. One archaeological feature that has received significant attention is the qanat. This paper investigates the role of CNN² with the YOLO³ algorithm in identifying archaeological evidence, such as qanat channels, in the historical city of Zuzan. YOLO is a type of algorithm used to detect and classify objects within images and videos. This algorithm is renowned for its speed in object detection, making it particularly valuable in field archaeology where time is of the essence. The objective of this study is to identify qanat channels within the city of Zuzan by creating a database of historical aerial images, all of which feature qanat channels. Following this identification, the study will delve into the influence of these qanats on the formation of Zuzan city.
Research questions and Hypotheses: 1. How precise is the convolutional neural network employing the YOLO algorithm in identifying qanats and digitally extracting them from aerial imagery? 2. What role have the qanats of the Zuzan Plain, as the sole source and system of water distribution, played in the development of the city of Zuzan? By creating a database of qanats in old aerial images and implementing the YOLO version 8 algorithm in a convolutional neural network, the accuracy of detecting archaeological evidence will undoubtedly increase. Moreover, as the sole source of water distribution in a dry region lacking surface water, qanats were the primary factor in the formation of settlements in the Zuzan plain, including the city of Zuzan, during the Islamic period.
 
Identified Traces
Aerial images from the 40s and 70s with dimensions of 5315×5377 pixels were selected and homogenized. This was necessary to ensure consistent input dimensions for the convolutional neural network. For instance, all selected images from the National Cartographic Center were originally 5377x5315 pixels, but they were all downscaled to 800x800 pixels for standardization. If images were selected at their original pixel resolution, the image size would become excessively large, significantly increasing the algorithm’s execution time and rendering it computationally impractical. The archaeological evidence in these 800-pixel images remains unaltered. At this stage, the images were not cropped; instead, the Snagit software was used to preserve the original image quality.
The selected historical aerial images were labeled in the YOLO Label environment. The data was divided into training and testing sets. 80% of the data (57 images) was used for training, and 20% (15 images) was used for testing. In both the training and testing sets, no images of Zuzan were labeled. In this study, the features of interest were solely the shafts and channels of qanats, which were classified into a single class in the YOLO LABEL environment. 72 aerial images from various regions of Greater Khorasan, each with a resolution of 800x800 pixels, were trained on the YOLOv5 model with 200 epochs and a learning rate of 0.01 using Python 3.9. The optimal values for the number of epochs and the learning rate were determined experimentally based on the convergence of the error curve The Ultralytics library, built on PyTorch, was employed for this research. The processing system used was a CPU with five cores and 12GB of RAM on Google Colab. Data augmentation was also utilized to increase the training dataset.
 
Conclusion
Experimental results and metrics obtained from an 80% training set and 20% test set, with 200 epochs and a learning rate of 0.01 in a convolutional neural network, indicate a 94% accuracy in detecting qanats surrounding the city of Zuzan, the sole source of water for the city. Given the form of Zuzan city (despite leveling and intermingling of deposits) and its geographical location, it can be asserted that factors such as the sustainability of water resources have contributed to urban life and agricultural activities until now. The creation of a central axis in shaping the city, the formation of open spaces within the city, and most importantly, the lack of surface water are the main impacts of the identified qanats on the formation and transformations of Zuzan city during the Islamic period. In a study of Zuzan city, adopting a novel approach in the field of automated archaeology and neural networks, four qanat channels were identified. With the help of information obtained from local people, these qanats are named Zuzan, Qasem Abad, Khargardak, and Chub Deraz, which have been considered as the city’s water sources during different periods.