The UnderThesands project tested the applicability of MSRM (Multi-Scale Relief Model), SMTVI (Seasonal Multi-Temporal Vegetation Indices), TCT (Multi-Temporal Tasselled Cap Transformation), and Principal Component Analysis using Sentinel-2 and LANDSAT to identify paleochannels and traces of ancient canals within the chosen study areas. New algorithms were developed using Synthetic Aperture Radar Sentinel-1 data. The paleochannel network was then mapped, and the composite of SMTVI and MSRM was created, laying the groundwork for the project's upcoming phase involving machine/deep learning techniques.
UnderTheSands project developed the machine learning model for detecting levees in Google Earth Engine (GEE). The model, which combined SMTVI, TCT, and raw Sentinel-2 reflectance data with MSRM in the Murghab fan area, had insufficient applicability to other areas because of the limitations of the GEE platform. However, the team's extensive work on a deep learning U-Net model using training data from various locations resulted in the highest precision and recall of about 0.66 with nine selected Sentinel-2 bands or the SMTVI. To further enhance the model's performance, the team incorporated Swin UNETR and Attention UNET architectures based on Visual Transformers into the research, achieving a precision of 0.7 and recall of 0.78 with one of the selected areas. We plan to continue increasing the dataset's size and adjusting parameters to enhance the model's performance further.
The project focused on detecting the ancient underground water distribution systems called qanat (kārīz, kariz, or falaj), visible in satellite imagery as linear groups of holes. First, the Gorgan Plain region bordering Turkmenistan was selected as a training area. Several approaches to labelling, dataset splitting, deep learning techniques (Instant Segmentation and Object Detection), and architectures were tested (YOLOv5 and YOLOv7-9). A new approach to the dataset was created. This included marking individual shafts and their pairs. The training data was created to best represent qanats from Gorgan and Maywand areas. Part of the dataset consisted of augmented and artificial representations of qanats, which have been created. After completing the central core of the model, the post-processing and pre-processing techniques have been developed.
Two detailed study test areas, Murghab Fan and Serakhs / Geoksyur Oasis, have been selected for detailed archaeomorphological analysis of hydrological networks and autocorrelation techniques. The hypothetical dating of the paleochannels and remains of ancient canals (levees) has been proposed. The results were further integrated into a broader discussion of landscape and water management evolution in the selected studied areas. An intensive literature review has been conducted to cover all the analysed areas.
