Airborne laser scanning (ALS) has, over the past two decades, substantially expanded the
possibilities of archaeological prospection, while at the same time exponentially increasing the
volume of data that needs to be reviewed, opening up space for deep learning methods in the
(partial) automation of archaeological feature recognition. This master's thesis tests the
performance of the open-source tool ADAF (Automatic Detection of Archaeological Features) in
the automatic detection of barrows (gomile), one of the most common yet morphologically
extremely heterogeneous archaeological structures in the Slovenian territory. Since the default
ADAF model was trained on Irish and English barrows, which are geomorphologically and
culturally different from Slovenian ones, a manual interpretation of ten selected Slovenian
barrow sites from different periods and regions was first carried out, in which 374 barrows were
recorded; three new models were then trained on locally acquired data (10,294 objects, of which
942 were barrows), and their results were compared with those of the default model. The
analysis showed that the default model detects only 7.75% of the manually recorded barrows,
whereas fine-tuning with local data nearly doubles the success rate (the best new model achieves
24.3% overlap), with barrow roundness being the key morphological factor for successful
detection, followed by the definition of its edges, while slope and area of the barrow have a
negligible effect. Based on the results, it is concluded that ADAF, in its current form, functions
primarily as a tool for the initial survey of areas with morphologically well-preserved barrows,
rather than as a substitute for expert field interpretation; substantially improving the
transferability and reliability of the models will, in the future, require a quality expansion and
field verification of the training datasets, as well as systematic testing of model transferability
across different regions and cultural contexts.
|
