My MSc dissertation, "Vision Transformers for Multimodal Archaeological Anomaly Detection", asks whether a Vision Transformer-style Masked Autoencoder can improve reconstruction-based archaeological anomaly detection compared with conventional convolutional approaches. Buried archaeological features are hard to spot in satellite imagery because the signal is subtle, fragmented, and easily confused with ordinary landscape variation — the aim was a clear, honest evaluation of whether modern deep learning actually helps.
I built a reproducible pipeline over a 2°×2° stretch of the Nile Valley, Egypt, combining atmospherically-corrected Sentinel-2 L2A imagery (via Google Earth Engine) with a Copernicus DEM (via OpenTopography) and ISRIC SoilGrids soil data, co-registering all three into consistent, terrain-aware tiles. Each tile feeds an eight-channel stack — six Sentinel-2 bands (blue, green, red, NIR, SWIR1, SWIR2) plus NDVI and NDRE vegetation indices — split into 512×512 patches. Training was fully unsupervised; a small ground-truth dataset, curated from UK sites with known archaeological excavations, was held back purely for evaluation.
Three architectures were tuned through a controlled architecture-search process and compared on 575 unique training tiles: a convolutional autoencoder baseline (CNN 3, ~779k parameters), a ViT-style Masked Autoencoder (MA 6×3 — 6 encoder / 3 decoder layers, 75% patch masking, ~764k parameters), and a ResNet50-based autoencoder. Notably, an untrained ResNet50 encoder beat every pretrained-ImageNet variant tested, suggesting natural-image features transfer poorly to 8-channel Sentinel-2 reconstruction — a ~32.2M-parameter model that ended up being outclassed by two much smaller ones.
The Masked Autoencoder came out on top, with better reconstruction metrics (MSE, MAE, PSNR) and marginally stronger precision/recall/F1 against the ground-truth set than the CNN and ResNet50 comparators. But the improvement was modest, not the step-change the literature suggested was possible, and false positives remained the dominant limitation across all three models — largely because a thresholded reconstruction-error map is only an indirect proxy for archaeology. I'd rather report that plainly than oversell it: the honest conclusion is that Terra-AID is currently better understood as a candidate-area triage tool that helps prioritise where archaeologists look next, not a definitive detector. Future work would centre on reducing false positives through better threshold calibration and spatial post-processing, and validating against a larger, independent case study such as HS2 Phase One.
Alongside the research, I built the whole workflow as a proper desktop application (Python, PySimpleGUI) rather than a stack of notebooks — dataset curation and preprocessing, config-driven training, evaluation, and prediction with visual outputs (heatmaps, cluster overlays, stitched anomaly maps) are all usable end-to-end without touching code.
