Multitemporal Multispectral Dataset for Palaeochannels Segmentation (MAPS)

Deep learning methods have significantly advanced the analysis of satellite imagery, greatly enhancing the efficiency of screening extensive datasets. However, progress in the automatic identification of elements under the surface (subsoil features), detectable only through proxies like soil and vegetation alterations, has been limited. This challenge is further compounded by the significant influence of seasonal variability and land use management on their visibility. Automatic identification often requires multitemporal observations to maximise the likelihood of obtaining favourable visibility conditions and to strengthen the reliability of each identification. Additionally, the limited availability of publicly accessible specialised datasets has impeded progress in applying deep learning approaches. This paper presents the first publicly available dataset of multitemporal Sentinel-2 imagery for identifying palaeochannels, a specific category of subsoil features. We established a baseline for semantic segmentation by evaluating three architectures with various parameter configurations, specifically designed to address the unique challenges of segmenting these elongated and branched features under diverse landscape conditions at different times of the year. The results offer preliminary insights into addressing the varying visibility of subsoil features by applying deep learning to analyse temporal series. The dataset is available on Hugging Face and IIT Dataverse repositories.