Abstract

Accurate plant emergence information is essential for evaluating crop stands and supporting early-season resource management. However, conventional emergence estimation relies on an extensive manual process. Existing automated stand-counting methods from commercial vendors focus heavily on major commodity crops like corn and soybeans. This study developed an open-source, QGIS-based workflow for UAV-derived stand counting and spatial characterization in horseradish (Armoracia rusticana). Horseradish is a specialty crop for which automated stand-counting methods remain largely unexplored. High-resolution RGB images were collected using a DJI Mavic 3M in a commercial horseradish field in Illinois, USA. Images covered a 3-acre field and were evaluated using 20 randomly distributed 20 × 35 m (700 m2) sampling areas. Manual image-based counting resulted in a mean stand count of 1,138 ± 98 plants per 700 m2, or approximately 65,777 plants per acre. This value is 83% of the theoretical population expected from the planting geometry. Two automated approaches were evaluated against the manual reference: a traditional computer vision (CV) approach and a GeoAI-based workflow using Mask R-CNN instance segmentation. The traditional CV approach achieved a mean absolute error (MAE) of 67 plants and a mean absolute percentage error (MAPE) of 6.1%. The GeoAI approach produced an MAE of 144 plants and a MAPE of 13%. Both approaches overestimated the horseradish population because weeds were occasionally identified as individual plants, a limitation of this study. Beyond stand count, the study also examined individual canopy area extraction and crop-row geometry. The GeoAI row-detection approach correctly identified all rows in an independent test image. These results demonstrate that a computationally simple, open-source CV approach can provide an accurate stand-count estimate under early-season conditions.

Keywords

Horseradish, Stand count, Unmanned Aerial Vehicle (UAV), QGIS, GeoAI, Computer Vision, Mask R-CNN

DOI

10.5703/1288284318728

Share

COinS
 

Open-Source QGIS framework for UAV Stand Counting in Horseradish

Accurate plant emergence information is essential for evaluating crop stands and supporting early-season resource management. However, conventional emergence estimation relies on an extensive manual process. Existing automated stand-counting methods from commercial vendors focus heavily on major commodity crops like corn and soybeans. This study developed an open-source, QGIS-based workflow for UAV-derived stand counting and spatial characterization in horseradish (Armoracia rusticana). Horseradish is a specialty crop for which automated stand-counting methods remain largely unexplored. High-resolution RGB images were collected using a DJI Mavic 3M in a commercial horseradish field in Illinois, USA. Images covered a 3-acre field and were evaluated using 20 randomly distributed 20 × 35 m (700 m2) sampling areas. Manual image-based counting resulted in a mean stand count of 1,138 ± 98 plants per 700 m2, or approximately 65,777 plants per acre. This value is 83% of the theoretical population expected from the planting geometry. Two automated approaches were evaluated against the manual reference: a traditional computer vision (CV) approach and a GeoAI-based workflow using Mask R-CNN instance segmentation. The traditional CV approach achieved a mean absolute error (MAE) of 67 plants and a mean absolute percentage error (MAPE) of 6.1%. The GeoAI approach produced an MAE of 144 plants and a MAPE of 13%. Both approaches overestimated the horseradish population because weeds were occasionally identified as individual plants, a limitation of this study. Beyond stand count, the study also examined individual canopy area extraction and crop-row geometry. The GeoAI row-detection approach correctly identified all rows in an independent test image. These results demonstrate that a computationally simple, open-source CV approach can provide an accurate stand-count estimate under early-season conditions.