Remote sensing dataset for machine learning in archaeological site recognition

The repository's data contains fragments of imagery used as input for machine learning. This data was developed from three types of imagery: Aerial photographs, Rao's Q index created from a DTM, SAR imagery. All three source layers were converted to grayscale images and then combined in specific proportions: 65% aerial photographs, 20% Rao's Q index, and 15% SAR imagery. These chosen proportions maximized the model's accuracy by mitigating noise in the SAR data while still incorporating subtle terrain variations. The conversion to grayscale was done to facilitate faster model training and to account for the varying colors of vegetation and soil features. The current dataset update includes the full weight matrix (weights) obtained from the machine learning process of the Mask R-CNN model. This file records the network state after completing a full cycle of 250 training epochs, implemented in the TensorFlow environment. Users can initialize the model architecture with these weights, allowing for immediate mask generation and archaeological anomaly detection on new RGB/SAR data without having to repeat the computationally expensive training process. The provided weights can serve as a starting point (pre-trained weights) for other researchers who would like to adapt the model to local specificities in other geographic regions, significantly accelerating the convergence of new models. Including the weights enables a thorough analysis of how the network interprets the fusion of Rao's Q and SAR intensity, which is essential for a reliable assessment of detection reliability in remote sensing archaeology. This research was funded by the National Science Centre, Poland under Grant no. 2024/08/X/ST10/00587