A sequential YOLOv11n–CNN hybrid framework for agricultural field delineation and crop classification from Sentinel-2 imagery

Accurate agricultural field delineation and target-class classification from satellite imagery are important for large-scale agricultural monitoring and precision agriculture. This study proposes a sequential YOLOv11n–CNN framework that integrates YOLOv11n-based instance segmentation with a transfer learning-based MobileNetV3Small classifier for field-level mapping from Sentinel-2 imagery. In the proposed pipeline, YOLOv11n first localizes individual agricultural fields and generates instance masks. The predicted masks are then used to construct mask-filtered regions of interest (ROIs), which are classified by MobileNetV3Small into ten target crop and land-cover classes. This sequential design explicitly separates spatial prediction from downstream classification, enabling the contribution of each stage to be evaluated independently. On the held-out Dataset-B test set, the YOLOv11n spatial stage achieved a class-agnostic localization precision of 0.769, recall of 0.842, F1-score of 0.804, mask [email protected] of 0.881, mask [email protected]:0.95 of 0.654, and a mean matched-mask IoU of 0.772. On 176 class-agnostically matched field instances, MobileNetV3Small achieved a conditional overall accuracy of 0.847, Cohen’s Kappa of 0.829, and macro-F1 of 0.847, compared with 0.790, 0.766, and 0.792, respectively, for the native YOLOv11n class prediction. Across the complete Dataset-B test set, end-to-end class-aware instance-level precision, recall, and F1-score were 0.651, 0.713, and 0.680, respectively. While the field-level exact McNemar test returned p = 0.041, image-cluster-aware paired inference did not establish a statistically significant OA improvement (ΔOA = 0.057; 95% CI: −0.009 to +0.115; p = 0.095). Ablation analysis further showed that mask-filtered ROIs improved overall accuracy from 0.818 to 0.847 compared with conventional bounding-box ROIs. In a separate controlled Dataset-A classification benchmark, MobileNetV3Small achieved an overall accuracy and macro-F1 of 0.924. Overall, the results support the effectiveness of separating instance-level field delineation from dedicated downstream classification and demonstrate the potential of the proposed framework for field-level agricultural mapping from Sentinel-2 imagery.