Deep Learning–Based Detection of Subsurface Features from Multispectral Imagery through Seasonally Variable Crop and Soil Marks

The detection of subsurface anthropogenic and geomorphological features in optical satellite imagery relies on weak and indirect surface expressions, called soil or crop marks, and vary with seasonal and environmental conditions. In agricultural landscapes, these signals appear through subtle soil and vegetation responses influenced by crop phenology, soil moisture, and land management, making consistent mapping challenging. This dissertation investigates the use of deep learning to detect crop and soil marks linked to subsurface features using multispectral Sentinel-2 imagery. It combines dataset design, spectral preprocessing, and tailored deep learning architectures to learn faint, seasonally dependent patterns. A key focus is the integration of multitemporal observations, treating seasonal variability as complementary information. By embedding this within a unified framework, the study improves segmentation continuity and generalisation, offering methodological insights for archaeological and geomorphological prospection in dynamic environments.