Assessing the Impact of Spatial Resolution and Hyperparameters on Automatic Agricultural Parcel Delineation Using the Segment Anything Model With Multi‐Resolution and Super‐Resolved Satellite Imagery
TRANSACTIONS IN GIS, cilt.39, sa.4, ss.1-25, 2026 (SSCI, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 39 Sayı: 4
- Basım Tarihi: 2026
- Doi Numarası: 10.1111/tgis.70306
- Dergi Adı: TRANSACTIONS IN GIS
- Derginin Tarandığı İndeksler: Academic Search Ultimate (EBSCO), Natural Science Collection (ProQuest), Business Source Ultimate (EBSCO), Earth, Atmospheric, & Aquatic Science Collection (ProQuest), Scopus, Sociology Source Ultimate (EBSCO), Social Sciences Citation Index (SSCI), ABI/INFORM, Compendex, Environment Index, Geobase, INSPEC
- Sayfa Sayıları: ss.1-25
- Hacettepe Üniversitesi Adresli: Evet
Özet
Accurate and automatic determination of the boundaries of agricultural parcels is essential for planning sustainable agricul-tural policies, managing agricultural production, tracking products and enabling precision farming applications. In recent years,pre-trained visual base models with zero-shot segmentation capabilities, such as the Segment Anything Model (SAM), have in-troduced a new paradigm to the field. Integrating such models with satellite images of varying characteristics opens up new pos-sibilities in terms of both accuracy and operational efficiency. While traditional satellite systems such as Sentinel-2, PlanetScope,and Pleiades provide data at specific resolution ranges, super-resolution (SR) like S2DR3 and Sen2SR, which have been developedin recent years, can enhance the spatial detail of existing data for analysis. In this study, we tested satellite images with differentspatial resolutions for the automatic extraction of agricultural parcel boundaries using the SAM model. We also evaluated the per-formance and contribution of super-resolution datasets like S2DR3 and Sen2SR. Additionally, the impact of the model's hyperpa-rameter configurations on segmentation accuracy was analyzed, and pre- and post-processing steps were applied to ensure thatonly agricultural parcel areas were obtained. The findings revealed that resolution directly impacts segmentation performance,though not in a linear manner. Although Sen2SR images have a higher spatial resolution than PlanetScope data, the former pro-duced weaker results, particularly with regard to geometric error metrics (GOSE: 0.318, GUSE: 0.373). Although Pleiades imagesachieved the highest numerical accuracy values (intersection over union (IoU): 0.942; Dice coefficient: 0.971), they were found tobe limited in practical applications due to issues such as over segmentation, processing time, and cost. In contrast, S2DR3 dataprovided balanced segmentation results, both visually and numerically (IoU: 0.920; DICE: 0.958), and was found to be the mostoperationally applicable solution due to its relatively low computational demand and short processing time. Hyperparameteranalyses showed that accuracy increased by up to 40% in low-resolution Sentinel-2 images. However, improvements beyond thedefault configuration were limited in high-resolution datasets (points_per_side: 64; crop_n_layers: 2). Considering the trade-offbetween processing time and accuracy, the default configurations (points_per_side = 32 and crop_n_layers = 0) were found toprovide sufficient accuracy for many scenarios. In conclusion, this study demonstrates that powerful base models such as SAMcan be effectively applied to satellite images enhanced with super-resolution techniques, and that synthetic data such as S2DR3offer the most operationally feasible solutions when criteria such as cost, hardware, processing time, and accuracy are consid-ered. The study provides practical recommendations for practitioners regarding the resolution–hyperparameter relationship andmakes practical contributions to large-scale or resource-constrained agricultural monitoring scenarios.