# Detection of Taiqiu sweet persimmons during the color-transition period with an improved YOLO11-FC2T model and causal analysis

## TL;DR

This paper introduces an improved YOLO11 model for detecting Taiqiu sweet persimmons during the color-transition period, enhancing accuracy and reliability in orchard conditions.

## Contribution

The novel YOLO11-FC2T model with architectural enhancements and a new data-augmentation method improves fruit detection in complex agricultural settings.

## Key findings

- YOLO11-FC2T achieves 91.7% precision and 86.7% recall, outperforming the baseline YOLO11 model.
- The model reduces false detection rates by 45.2% on a challenging tail-case dataset.
- Architectural components and DiffuseMix augmentation were validated for their individual and combined contributions using causal-effect analysis.

## Abstract

Accurate detection of Taiqiu sweet persimmon in orchards is essential for estimating yield, planning harvest operations, and supporting intelligent management in precision agriculture. However, current fruit-detection approaches for this cultivar, especially during the color-transition period, suffer from highly subjective and inefficient manual inspection and from poor adaptability of existing deep-learning models to complex field scenes.

In this study, we propose an improved YOLO11-based detector, YOLO11-FC2T, for robust detection under conditions with strong color–background coupling, small or adherent fruits, and uneven illumination. YOLO11-FC2T introduces four key architectural modifications: (1) a C3k2_FasterBlock to improve gradient-efficient feature learning; (2) a C2PSA_CGA module to enhance channel–spatial focus via coordinate-guided aggregation; (3) a three-layer Dysample-T structure to strengthen multi-scale representation; and (4) a cross-scale attention fusion module, CAFMAttention, to better decouple fruits from cluttered backgrounds. To further enhance generalization in complex orchard scenes without additional labeling cost, we introduced the DiffuseMix data-augmentation method and apply it to color-transition images.

Experiments show that YOLO11-FC2T clearly outperforms the YOLO11 baseline. The model achieves a precision of 91.7% (+1.0%), recall of 86.7% (+2.8%), mAP@0.5 of 94.8% (+1.6%), and mAP@0.5-0.95 of 81.2% (+4.0%), where mAP@0.5 uses an IoU threshold of 0.50. On a challenging tail-case set of 537 images, the false detection rate is 1.30%, with a 45.2% reduction in errors relative to YOLO11. In the performance evaluation stage, we first perform causal-effect analysis based on the Average Treatment Effect (ATE) to quantify the independent and joint contributions of each architectural component and of DiffuseMix; at the same time, the efficiency of the model is analyzed by the number of parameters (Params, M) and per-image inference latency (ms). in addition, we construct and use a dedicated tail-case dataset as a supplementary experiment to further verify the robustness and effectiveness of these improvements in the most difficult scenes. Finally, we introduced cross-condition test set to further validate the generalization capability of YOLO11-FC2T. The above results indicate that YOLO11-FC2T not only improves the indicators, but also possesses reliable generalization ability and stability.

Overall, YOLO11-FC2T addresses key detection challenges during the color-transition period and provides a practical, portable solution for automated fruit identification and counting in precision agriculture. The above results indicate that YOLO11-FC2T not only improves the indicators, but also possesses reliable generalization ability and stability.