GFDD24: Guava Fruit Disease Dataset 2024

• GFDD24 is a curated image dataset featuring expert-validated, augmented RGB images for detecting guava diseases (Healthy, Fruit Flies, Anthracnose).
• It employs a robust pipeline integrating EfficientNet-B0 for feature extraction and a cascade of gradient boosting models for near-perfect classification.
• The dataset supports real-time deployment on edge devices, enhancing precision agriculture and crop protection in Bangladesh.

Guava Fruit Disease Dataset 2024 (GFDD24) is a curated image dataset designed for automated detection of guava fruit diseases, particularly tailored to the agricultural landscape of Bangladesh. It comprises expert-validated RGB images of guava fruits, categorized according to their health status—Healthy, Fruit Flies infected, and Anthracnose infected—sourced from plantations in Rajshahi and Pabna regions. GFDD24 provides the foundational benchmark for developing and evaluating computer vision models aimed at real-time monitoring and disease management in guava cultivation (Rijon et al., 23 Dec 2025).

1. Dataset Composition and Acquisition Protocol

GFDD24 consists of 3784 processed images, derived from 473 raw RGB photographs (Healthy: 161, Fruit Flies: 173, Anthracnose: 139). Dataset augmentation was performed at an 8× rate within each class, yielding final class sizes of 1288, 1384, and 1112 for Healthy, Fruit Flies, and Anthracnose, respectively. Image acquisition was conducted under natural daylight using DSLR cameras at an original resolution of 512×512.

Preprocessing included resizing, per-channel normalization, Contrast Limited Adaptive Histogram Equalization (CLAHE), unsharp masking for edge enhancement, and geometric augmentations (rotations, flips, translations) to simulate real-world variance. For subsequent model training, random undersampling balanced all classes at 1112 images and standardized the input to 224×224 pixels, with pixel values rescaled by 1/255.

Class	Raw Images	Augmented	Balanced Final
Healthy	161	1288	1112
Fruit Flies	173	1384	1112
Anthracnose	139	1112	1112

2. Feature Extraction with EfficientNet-B0 CNN

GFDD24's benchmark pipeline leverages the EfficientNet-B0 architecture as the primary convolutional neural network (CNN) feature extractor. The input tensor size for the CNN is $224 \times 224 \times 3$. Initial processing involves a Conv3×3 stem (32 filters, stride=2, padding=same), BatchNorm, and Swish activation. This is succeeded by sequential MBConv blocks, each with expand-conv-project and squeeze-and-excitation mechanisms, culminating in a single Conv1×1 head with 1280 output channels.

Feature extraction is finalized by global average pooling, producing a feature vector $z\in\mathbb{R}^{1280}$ per image, subsequently used for downstream classification. No dropout was applied during fine-tuning (original drop rate for EfficientNet-B0 is 0.2).

3. Gradient Boosting Ensemble: XGBoost, LightGBM, CatBoost

Following CNN-based feature generation, GFDD24 employs gradient boosting machines (GBMs) including XGBoost, LightGBM, and CatBoost for classification. Each GBM ingests the 1280-dimensional feature vector. Hyperparameters were largely default, with learning_rate=0.1, n_estimators=100, max_depth=3 for XGBoost/LightGBM, and CatBoost configured for 1000 iterations and depth=6.

The overall ensemble prediction computes:

$$\hat y = \sum_{m=1}^{M} \gamma_m\,h_m(x), \quad \gamma_m\in\mathbb R,\; h_m:\mathbb R^{1280}\to\mathbb R^3$$

where $M$ is the total number of base learners, each mapping the feature vector to class logits.

4. CNN–GBM Cascade Decision Rule

GFDD24 introduces a cascade classification framework for robust decision-making. Raw images are resized and rescaled for CNN feature extraction. The primary classifier (e.g., AdaBoost→LightGBM) outputs class probabilities $p_{\text{base}}(y|z)$, with confidence $\gamma(z)=\max_c p_{\text{base}}(y=c|z)$. If $\gamma(z)\ge0.8$, the base prediction is accepted. Otherwise, CatBoost's output $p_{\text{cat}}(y|z)$ determines the label:

$$\hat y(z)= \begin{cases} \arg\max_c\,p_{\text{base}}(y=c|z), & \gamma(z)\ge\tau \ \arg\max_c\,p_{\text{cat}}(y=c|z), & \gamma(z)<\tau \end{cases}$$

with $\tau=0.8$. This two-stage approach reduces uncertainty and leverages classifier complementarity for edge-case resolution.

5. Training Methodology and Performance Analysis

The training protocol for GFDD24 consists of an 80%/20% train-validation split, with test data held out for final assessment. Optimization utilizes Adam with a fixed learning rate of $1\text{e}{-3}$, mini-batch size 32, and categorical cross-entropy loss:

$$L = -\frac1N\sum_{i=1}^N \sum_{k=1}^3 y_{i,k}\,\log\hat p_{i,k}$$

Performance metrics are reported on the balanced test set using AdaBoost+LightGBM, showing class-wise and aggregate precision, recall, $F_1$-score, and accuracy at 0.9999. The confusion matrix is nearly perfectly diagonal (99.99% mass), indicating minimal misclassification.

Class	Precision	Recall	F1-score	Support
Healthy	0.9999	0.9999	0.9999	1112
Fruit Flies	0.9999	0.9999	0.9999	1112
Anthracnose	0.9999	0.9999	0.9999	1112
Accuracy			0.9999

Ablation comparisons highlight the superiority of balanced training: RF+LGBM increases from 96.23% to 96.42% accuracy, while AdaBoost+LightGBM achieves stable performance at ~99.99%.

6. Deployment: Real-Time, Edge, and Application Scenarios

The total model footprint is approximately 21 MB (EfficientNet-B0: 20 MB, LightGBM: 1 MB), with possibility of compression (8-bit quantization) to ≈8 MB for CPU/mobile inference. Inference latency per image is reported as ≈11 ms (GPU, T4) and ≈65 ms (4-core CPU). System throughput can be further optimized by channel pruning or TensorRT/ONNX conversion, targeting smartphone, tractor-mounted, or batch-processing deployment modes.

Recommended deployment practices include REST API integration, TensorFlow Lite or ONNX conversion for edge computing, and adapting inference strategy for either batch sorting or in-field individual analysis.

7. Context, Implications, and Future Research Directions

GFDD24 establishes a standardized, high-resolution benchmark for guava disease detection in Bangladesh, enabling the development of real-time monitoring systems crucial for crop protection against anthracnose and fruit fly infestations. The two-stage CNN–GBM cascade and rigorous data balancing result in near-perfect precision, recall, and $F_1$-scores on held-out data.

A plausible implication is that a similar framework could be adapted for other regional agriculturally critical crops, contingent on dataset availability and disease typology. This suggests ongoing research potential in cross-crop transfer learning and scalable deployment for precision agriculture in resource-constrained environments (Rijon et al., 23 Dec 2025).