A Hierarchical Deep Learning Framework for Pineapple Disease Detection, Typification, and Spatial Disease Visualization in Tropical Agricultural Environments

Technical Summary: A Hierarchical Deep Learning Framework for Pineapple Disease Detection, Typification, and Spatial Disease Visualization

Problem Statement
Early and reliable diagnosis of pineapple diseases in tropical agricultural environments remains a significant challenge for precision agriculture. Current manual inspection methods are labor-intensive, subjective, and difficult to scale. Automated solutions face substantial hurdles due to variable illumination, fruit overlap, occlusions, heterogeneous backgrounds, and the high visual similarity among different disease symptoms. Furthermore, existing deep learning approaches often rely on single-stage inference pipelines evaluated on controlled datasets (e.g., isolated leaves), which fail to generalize to the complexity of real-field tropical conditions. There is a specific gap in frameworks that integrate disease screening, typification, multi-instance analysis, and spatial visualization within a unified architecture suitable for resource-constrained family farming and UAV-assisted monitoring.

Methodology
The authors propose a hierarchical deep learning framework designed for real-field tropical conditions. The system employs a two-stage inference strategy:

1. Stage 1: Binary Phytosanitary Screening: A YOLOv8 detector is used to localize pineapple fruits and classify them as either "healthy" or "diseased." This stage filters out healthy fruits to reduce computational load in subsequent steps.
2. Stage 2: Disease Typification: Symptomatic fruits identified in Stage 1 undergo Region of Interest (ROI) extraction with padding to preserve contextual information. Two alternative refinement strategies are evaluated for this stage:
   • Detector-based: A second YOLOv8 model performs simultaneous localization refinement and classification.
   • Classifier-based: A DenseNet121 classifier operates on the cropped ROIs to perform image-level disease classification.

Key Technical Components:

• Symptom-Oriented Grouping: To mitigate inter-class ambiguity, diseases are organized into four groups based on shared visual symptoms and agronomic relevance rather than strict pathological categories:
  • Group 1: Black Spot / Fusarium Guttiforme / Fruit Mealybug.
  • Group 2: Fruit Rot / Fruit Borer / Crown Rot.
  • Group 3: Mealybug Wilt / Root Rot.
  • Group 4: Multiple Crown / Fruit Fasciation.
• Multi-Instance Propagation: The framework supports two propagation modes: selecting only the highest-confidence symptomatic detection (for efficiency) or propagating all symptomatic detections (for simultaneous multi-fruit analysis).
• Spatial Visualization: The system integrates confidence-aware and class-oriented visualization, assigning specific RGB colors to disease groups and mapping prediction confidence to color intensity, facilitating spatial disease mapping.

Dataset and Experimental Setup
The study utilized a dataset comprising 68.86% real-field images and 31.14% publicly available images, totaling 2,383 training, 677 validation, and 338 test images. The dataset reflects tropical variability in illumination, scale, and occlusion. An independent blind-test evaluation was conducted using 50 previously unseen field-acquired images (10 per class) to assess generalization. Experiments were performed on an NVIDIA Tesla T4 GPU using PyTorch and Ultralytics frameworks.

Key Results

• Stage 2 Performance:
  • The YOLOv8 typification model achieved 96.1% precision, 96.0% recall, 97.7% mAP@50, and 71.3% mAP@50–95 on the independent test set.
  • The DenseNet121 classifier achieved 98.82% accuracy on cropped disease images, outperforming YOLOv8 in pure classification metrics due to the absence of localization requirements and reduced background interference.
• Hierarchical Framework Evaluation (Blind Test):
  • YOLOv8 + YOLOv8: Achieved 78.0% image-level accuracy with an average prediction confidence of 85.0%. Inference time was 19.0 ms per image.
  • YOLOv8 + DenseNet121: Achieved 84.0% image-level accuracy with an average prediction confidence of 96.4%. Inference time increased to 35.8 ms per image.
• Confidence Refinement: The hierarchical approach significantly improved prediction confidence. The YOLOv8+DenseNet121 configuration increased average confidence from 82.0% (Stage 1) to 96.4% (Stage 2), a gain of 14.4 percentage points, compared to a 3.0 percentage point gain for the YOLOv8+YOLOv8 configuration.

Significance and Claims
The paper claims that the proposed framework offers a practical foundation for intelligent phytosanitary monitoring in tropical agriculture. Its primary significance lies in:

1. Robustness to Field Conditions: By using a dataset dominated by real-field images and symptom-oriented grouping, the framework addresses the visual variability and ambiguity common in tropical environments.
2. Trade-off Analysis: The study provides a comparative evaluation of detector-based versus classifier-based refinement, demonstrating that while the YOLOv8+YOLOv8 configuration is more computationally efficient (19.0 ms), the YOLOv8+DenseNet121 configuration offers superior diagnostic reliability and confidence calibration (84.0% accuracy vs. 78.0%), which is critical for decision support.
3. Spatial Interpretability: The framework enables multi-instance analysis and class-oriented spatial visualization, supporting future integration with georeferenced UAV-assisted monitoring systems for localized phytosanitary inspection.

The authors conclude that combining object detection with image classification constitutes an effective strategy for improving automated pineapple disease monitoring, providing a scalable solution adaptable to the resource constraints of family farming systems.