Agricultural Field Segmentation and Crop Type Mapping in Satellite Image Time Series Data by applying Graph Neural Network

Agriculture has evolved into a data-intensive domain in which satellite-based Earth observation plays a central role in supporting sustainable and informed decision-making. In particular, Satellite Image Time Series (SITS) can provide multispectral observations that capture the temporal evolution of crops throughout the growing season. However, exploiting these data remains challenging due to their high dimensionality, spatial complexity, and the need to model both spatial and temporal relationships effectively. This thesis proposes a novel deep learning framework based on Graph Neural Networks (GNNs) for the segmentation and crop classification of agricultural fields from Sentinel-2 time series. Unlike many existing approaches that rely exclusively on superpixel aggregation, the method introduced here preserves pixel-level resolution while using superpixels only as local structural units. This hybrid design allows to maintain fine spatial detail while reducing fragmentation. Moreover, temporal information is explicitly integrated into the graph node attributes, enabling the network to learn how spectral signatures evolve throughout the crop growth cycle. As a result, the model can effectively distinguish between crop types that may appear similar on a single acquisition date but exhibit distinct seasonal behavior. The proposed method was tested on a set of monthly Sentinel-2 images collected over an agricultural region in Austria. Experimental results demonstrate the ability of the framework to generate accurate and coherent segmentation maps, achieving strong performance in metrics such as mIoU (78,30%) and overall accuracy (87,61%). Visual evaluation further confirms a close correspondence between model predictions and reference data, highlighting the effectiveness of graph-based learning for capturing spatial and temporal relationships in SITS. In summary, the thesis demonstrates that GNNs offer a powerful and versatile paradigm for agricultural field segmentation from multispectral time series. By integrating temporal evolution, spatial context, and pixel-level precision within a unified architecture, the proposed approach contributes to the development of data-driven solutions for precision agriculture and environmental monitoring.