From Detection to Interpretation: Generating Human-Readable Cabbage Disease Reports Using YOLOv8s and LLaMA 3.1-13B-Instruct

Abstract

Background and Objectives: Recent advances in deep learning have significantly improved plant disease detection, particularly through convolutional neural networks (CNNs) and YOLO-based object detection architectures. In precision agriculture, these approaches have demonstrated strong capabilities for identifying and localizing disease symptoms with high accuracy. However, most existing systems primarily produce technical outputs such as bounding boxes, class labels, and confidence scores, which are difficult for non-expert agricultural users to interpret and utilize for practical decision-making. Consequently, a gap remains between detection performance and real-world usability. Farmers and agricultural practitioners often require contextual explanations, severity interpretation, and actionable recommendations rather than raw detection outputs. At the same time, recent developments in large language models (LLMs) have shown strong potential for transforming structured information into coherent and human-readable narratives. Nevertheless, the integration of object detection and language generation for agricultural disease reporting remains largely unexplored. To address this limitation, this study proposes an end-to-end detection-to-interpretation framework that integrates YOLOv8s and LLaMA 3.1-13B-Instruct to automatically transform cabbage disease detection outputs into human-readable and actionable agricultural reports. The framework aims to bridge the gap between visual disease detection and user-centered interpretation, thereby improving the interpretability, accessibility, and practical usability of artificial intelligence (AI) systems in precision agriculture.

Methodology: The proposed framework consists of three main components: disease detection, structured representation transformation, and natural language generation (NLG). A publicly available cabbage disease dataset containing Black Rot, Downy Mildew, and Healthy classes was adapted for object detection through manual annotation and expert validation. The final dataset comprised 1,500 images and was divided into training, validation, and testing subsets using a 70:15:15 ratio. YOLOv8s was employed as the disease detection model due to its favorable balance between detection accuracy and computational efficiency. The model was trained using resized images of 640 × 640 pixels with data augmentation techniques including rotation, scaling, flipping, and brightness adjustment. Detection outputs consisting of bounding boxes, disease labels, and confidence scores were subsequently transformed into a structured representation containing disease type, confidence level, spatial location, and estimated severity. This intermediate representation served as a semantic bridge between computer vision outputs and language generation. LLaMA 3.1-13B-Instruct was then used to generate natural language explanations through prompt-based structured-to-text generation. Few-shot prompting, controlled decoding parameters, and post-processing mechanisms were applied to improve consistency and readability. The framework was evaluated through multiple perspectives, including object detection metrics, text generation metrics, baseline comparisons, ablation analysis, and qualitative expert assessment.

Main Results: Experimental results demonstrate that the proposed framework achieves strong performance across both detection and interpretation tasks. The YOLOv8s detection module achieved a precision of 0.85, recall of 0.83, and mAP@0.5 of 0.87, indicating reliable disease localization and classification performance. Compared with the Faster R-CNN baseline, which achieved an mAP@0.5 of 0.81, YOLOv8s provided superior detection accuracy while maintaining computational efficiency suitable for real-time agricultural deployment. Class-level analysis showed that Black Rot achieved the highest detection performance with an mAP@0.5 of 0.89, while Downy Mildew achieved 0.85 due to the more subtle visual characteristics of its symptoms. For the language generation component, the generated explanations achieved a BLEU score of 0.48, ROUGE-L of 0.61, METEOR of 0.54, and BERTScore of 0.89, demonstrating strong semantic alignment with expert-written references. Baseline comparisons further revealed that the proposed framework outperformed both template-based reporting and a lightweight LLaMA 3.2-3B-Instruct model across all evaluation metrics. Ablation analysis confirmed the importance of the structured representation and prompt engineering components. Removing the structured representation reduced BERTScore from 0.89 to 0.83, while removing few-shot prompting reduced BLEU from 0.48 to 0.42 and resulted in more generic and repetitive outputs. Qualitative evaluation involving three agricultural experts demonstrated that 88% of the generated explanations were rated as accurate, clear, and actionable. Furthermore, the evaluation achieved strong inter-rater agreement with a Cohen’s Kappa coefficient of κ = 0.81, indicating high reliability and consistency among expert assessments.

Conclusions: The proposed YOLOv8s–LLaMA 3.1-13B-Instruct framework effectively bridges visual disease detection and agricultural interpretation by transforming technical detection outputs into understandable, context-aware, and actionable disease reports. The results demonstrate that integrating object detection with LLMs can significantly improve the interpretability and practical value of AI-based plant disease diagnosis systems. Beyond achieving strong detection accuracy and semantic generation quality, the framework supports user-centered decision-making by providing explanations that are accessible to non-expert agricultural users. The study contributes a novel detection-to-interpretation pipeline, an expert-validated cabbage disease detection dataset, and a structured representation mechanism that connects computer vision and NLG. Nevertheless, the framework currently relies on heuristic-based severity estimation and is limited to a small number of disease categories under constrained environmental conditions. Future research should focus on incorporating lesion segmentation, vision transformer-based architectures, multimodal environmental information, domain-specific language model adaptation, and large-scale field validation to improve robustness, scalability, and applicability in real-world precision agriculture environments.