Validação da plantabilidade em soja com uso de visão computacional

Abstract

This study presents a workflow for assessing soybean planting quality from point-based imagery acquired by a remotely piloted aircraft, integrating computer vision, spatial analysis, and geostatistics. The study was conducted at Mandaguari Farm (Indianópolis, Minas Gerais, Brazil), center pivot 3-1 (Field 1), covering 14.13 ha. Data acquisition, performed on October 30, 2025, followed two stages: (i) a flight at approximately 120 m AGL to generate an orthomosaic and digital elevation and terrain models; and (ii) a low-altitude point-based flight at approximately 4 m AGL, with one image per sampling point at the VC growth stage. Point images were processed with a YOLOv8m instance-segmentation model trained on 640 × 640- pixel tiles using a point-level dataset split. Validation reached mAP50(box)=0.963, mAP50(mask)=0.960, mAP50–95(box)=0.782, mAP50–95(mask)=0.715, precision(box)=0.918, and recall(box)=0.938. Segmented masks were converted to centroids in QGIS, and a Python script estimated the dominant planting direction, grouped crop rows, and calculated within-row plant spacing, classifying skips and doubles relative to the expected spacing within each image. As a quality-control criterion for centroid linking, points with drop_pct ≥ 25% were treated as severe linking failures and removed, yielding n=39 for descriptive and spatial analyses. After filtering, median values were %CV=65.20%, %doubles=8.93%, and %skips=18.68%. Ordinary kriging was used to generate thematic maps and variograms; however, Moran's I did not indicate significant spatial autocorrelation at the 5% level, and cross-validation yielded low R² values, indicating weak spatial structure and limited predictive performance under the evaluated conditions. The workflow was reproducible and suitable for quantifying point-level planting-quality indicators, whereas stronger spatial products will require higher sampling density and/or additional covariates.