SarSAM: an integrated framework for agricultural field boundary delineation and mapping using high-resolution SAR (PAZ) imagery and the Segment Anything Model
Advances in Space Research, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Basım Tarihi: 2026
- Doi Numarası: 10.1016/j.asr.2026.08.100
- Dergi Adı: Advances in Space Research
- Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Artic & Antarctic Regions, Compendex, INSPEC, MEDLINE, Academic Search Ultimate (EBSCO), Engineering Source (EBSCO)
- Anahtar Kelimeler: Crop mapping, Field boundary, PAZ, SAM, SAR
- Hacettepe Üniversitesi Adresli: Evet
Özet
Agricultural field boundary delineation with high accuracy and reliability is of critical importance for sustainable agricultural management, land-use planning, precision farming applications, and parcel-based crop monitoring. However, conventional semantic segmentation approaches often remain insufficient for preserving boundary continuity and geometric integrity, particularly in agricultural landscapes where phenological variation and crop-pattern-dependent spectral/radiometric heterogeneity are pronounced. In this context, the present study introduces the SarSAM framework, which integrates agricultural field boundary delineation based on the Segment Anything Model (SAM) with parcel-based crop mapping using high-spatial-resolution PAZ X-band SAR imagery. In doing so, the study provides an original contribution to the still limited body of research on the integration of SAM and SAR for agricultural applications. The study employed three PAZ images acquired on 18 June 2020, 10 July 2020, and 1 August 2020. These acquisition dates were not selected through an experimental optimization process; rather, they were determined by the temporal distribution of images available in the PAZ archive, enabling the construction of a multi-temporal SAR composite representing an approximately 40-day period. This strategy aimed to relatively mitigate speckle effects and radiometric fluctuations that may dominate single-date SAR imagery. Subsequently, edge prominence was enhanced through histogram equalization and linear contrast stretching, and automatic segment generation was performed using the ViT-H architecture of SAM with the optimal parameter settings of points_per_side = 64 and crop_n_layers = 2. Finally, post-processing steps were applied to remove segments that did not represent agricultural production areas, thereby retaining only the relevant agricultural parcels for analysis. Quantitative evaluation demonstrated that the proposed approach achieved high performance in agricultural field boundary delineation. The Precision, Recall, IoU, and DICE values were calculated as 0.96, 0.89, 0.88, and 0.92, respectively. The GOSE and GUSE values of 0.12 and 0.17 indicated that the remaining errors were predominantly associated with under-segmentation, mainly in the form of adjacent parcels being merged into a single segment. In addition, the SAM-derived segments were used as object units for parcel-based crop mapping using Sentinel-2 NDVI time series and a Random Forest classifier, resulting in an overall accuracy of 93.12% ± 2.18%. These findings demonstrate that the SarSAM framework provides a promising and operationally practical approach for agricultural parcel delineation and object-based crop mapping through its cloud-independent SAR observations and foundation-model structure that does not require retraining. Moreover, the results indicate that the derived segments can serve not only for boundary delineation but also as reliable spatial units for parcel-based agricultural analyses.