Field delineation from satellite images enables monitoring and measurement of agricultural land management, productivity indicators, and scale transitions in agriculture across the globe. Timely and repeat field delineations are hence listed as an Essential Agricultural Variable in light of the United Nation´s Sustainable  Development Goals. Recent advances in machine learning for computer vision have lifted state-of-the-art performance and field delineation has entered quasi-operational stage in consolidated agricultural settings. For heterogeneous and dynamic smallholder landscapes, however, current approaches are challenged by the scarcity of labeled reference data. Transfer learning – here defined as fine-tuning a pre-trained model for use in a different region - allows for resource-efficient transfer of field delineation models across heterogeneous geographies. This study explores opportunities for further reducing reference data requirements in transfer learning setups. We leverage pseudo-labels – i.e. labels predicted by a pre-trained model - for fine-tuning across geographies and sensor characteristics. The results of the study provide insights into the potential of pseudo-labels for supporting large-area field delineation in heterogeneous smallholder-dominated settings.