Research Information System
Advances in Space Research, vol.77, no.5, pp.5641-5663, 2026 (SCI-Expanded, Scopus)
Accurate identification of agricultural parcel boundaries is critical for sustainable agricultural management, land use planning, and digital agriculture applications. Especially in advanced analyses such as precision agriculture, crop monitoring and yield estimation, reliable parcel data constitute an essential analytical component. The Segment Anything Model (SAM), a recent advancement in visual semantic segmentation, demonstrates significant potential for the automatic extraction of agricultural parcels from high-resolution satellite imagery. However, several challenges remain regarding the model’s applicability across diverse agricultural contexts, including its parameterization, pre-processing requirements, and integration into operational workflows. The primary objective of this study is to integrate phenology-driven multi-temporal image selection with the zero-shot segmentation capability of SAM for the automatic delineation of agricultural parcel boundaries. Within the scope of the study, NDVI time series obtained from Sentinel-2 satellite data were overlaid with farmer-declared parcels to create characteristic spectral profiles for each crop. From these profiles, three dates (May 25, July 19 and September 22, 2022) that best represent the phenological stages of the crops were selected and PlanetScope images of these dates were used. Combinations of true color, false color and NDVI time series (NDVI TS) were generated from the images and provided as input to the SAM model. The results show that the NDVI TS approach outperforms the single date segmentation. The multi-temporal segmentation achieved an IoU of 0.89 and F1 score of 0.93, while the geometric segmentation errors remained low (GOSE: 0.12; GUSE: 0.13). In single date analyses, IoU values remained in the range of 0.73–0.81 and errors increased, especially during periods of poor vegetation cover. This study shows that phenologically-based multi-temporal image selection significantly improves segmentation performance and that zero-shot models such as SAM can perform highly accurate and operationally feasible boundary extraction without the need for large scale training data. Supported by high-resolution satellite data, this approach provides an efficient, cost-effective and directly applicable method for identifying cultivated land patterns and types, emphasizing that besides the model, the content characteristics and acquisition time of the imagery are also determining factors in segmentation success.