Systematic evaluation of SAM prompting strategies for agricultural land extraction with GeoSAM
Year: 2026
Type: article
Abstract: This study presents GeoSAM, a QGIS plugin that integrates five SAM prompting strategies (Text, Box, Point, Detection-guided Box and Detection-guided Point) for agricultural land extraction. Using 25 cm-resolution aerial imagery, Point Prompt achieved the highest F1-Score (0.9785) but is limited for large-scale use. Text Prompt scored 0.9406, with generic terms (e.g. ‘farm plot’) outperforming crop-specific terms. Adding text to Box Prompt improved its F1-Score from 0.5417 to 0.9220 (~70% gain), highlighting the value of semantic information. Detection-guided Point (0.8861) slightly outperformed Detection-guided Box+Text (0.8785). Compared to 97.05% F1 with 8 cm drone imagery, the performance dropped to 88.62% at 25 cm resolution, underscoring resolution's critical impact. GeoSAM helps practitioners navigate the accuracy–automation trade-off by offering context-appropriate segmentation strategies within a unified tool.
Source: Geocarto International
Authors: Jeongbae Jeon, Sangwan Lee, Solhee Kim, Taegon Kim
Language: English
Subfield: Media Technology
Field: Engineering
Domain: Physical Sciences
SDG: Zero hunger
Open Access status: gold
Estimated APC paid: $2,390
Funder: Rural Development Administration
Award: G6694346003
OpenAlex ID: W7203627670