基于影像组学与可解释机器学习的香蕉病害快速识别

Abstract

Introduction:

Banana production is increasingly threatened by fungal diseases such as Fusarium wilt and Black Sigatoka, posing severe risks to food security and agricultural economies. Recent image-based approaches using deep learning have shown high predictive capacity for plant disease recognition; however, their limited transparency, calibration uncertainty, and sensitivity to domain shifts can restrict their use in decision-support workflows that require auditability.

Methods:

This study proposes an interpretable and calibrated Artificial Intelligence framework for multiclass banana disease-pattern characterization based on radiomic feature analysis of RGB leaf images. Radiomic features were extracted from HSV-segmented banana leaf regions, resulting in a dataset of 14,763 samples characterized by 103 quantitative descriptors and labeled as Healthy, Sigatoka, or Fusarium wilt race 1.

Results:

Among the evaluated radiomics classifiers, the calibrated Random Forest achieved accuracy = 0.85, balanced accuracy = 0.84, macro-F1 = 0.84, and macro ROC-AUC OvR = 0.95 on the held-out test set. Bootstrap analysis yielded 95% confidence intervals of [0.8406, 0.8691] for accuracy and [0.8306, 0.8606] for balanced accuracy. Three deep learning baselines trained on the same partition achieved higher predictive performance: MobileNetV3 with accuracy = 0.96, macro-F1 = 0.95, and macro ROC-AUC OvR = 0.97; EfficientNet with accuracy = 0.96, macro-F1 = 0.97, and macro ROC-AUC OvR = 0.96; and ResNet-18 with accuracy = 0.96, macro-F1 = 0.96, and macro ROC-AUC OvR = 0.96.

Discussion:

The CNNs produced strong classification performance on the evaluated repositories, and the radiomics approach demonstrated to be a complementary interpretable and explainable calibrated reference model. SHAP, LIME, permutation importance, accumulated local effects, calibration curves, and Brier score decomposition supported feature-level inspection of the final model.

1 Introduction

Banana (Musa spp.) is one of the most important staple and export crops worldwide, playing a critical role in food security and socioeconomic stability across tropical and subtropical regions (). Ecuador is one of the world's leading exporters of bananas, and the crop represents a strategic pillar of its national agricultural economy (). However, banana production is increasingly threatened by fungal diseases that compromise yield, fruit quality, and plantation sustainability, most notably Fusarium wilt and Black Sigatoka (). The growing pressure exerted by these diseases has intensified the need for rapid, early, accurate, scalable, and interpretable diagnostic tools capable of supporting timely phytosanitary interventions.

Fusarium wilt of banana, caused by Fusarium oxysporum f. sp. cubense, is considered one of the most destructive plant diseases globally because of its soilborne nature, long persistence, and the lack of effective chemical control (). Of particular concern is Tropical Race 4 (TR4), which has spread across Asia, Africa, and Latin America in recent years (). In Ecuador, recent phytosanitary alerts and institutional containment actions have reinforced the need for surveillance technologies that support risk prioritization, field inspection, and confirmatory laboratory workflows as stated by the Ecuadorian agricultural authority . Although these technologies are not intended to replace official diagnosis, they can help to enhance disease monitoring, the recognition of disease patterns, and prioritize suspect cases for expert inspection and laboratory testing (; ).

Black Sigatoka, caused by Pseudocercospora fijiensis, represents another major fungal threat, leading to leaf necrosis, reduced photosynthetic capacity, and yield losses if not properly managed (). Unlike Fusarium wilt, Sigatoka symptoms are primarily foliar and visually detectable; however, symptom progression varies with environmental conditions, disease stage, cultivar, and management practice, making manual recognition subjective and error prone. In addition, visual overlap among disease-related leaf alterations may lead to delayed response or incorrect operational actions (; ).

Traditional disease identification in banana plantations relies heavily on expert visual inspection and laboratory-based confirmation. While effective, these approaches are time consuming, costly, and difficult to deploy continuously across large agricultural areas, particularly when diseases such as Foc TR4 require coordinated biosecurity, surveillance, and confirmatory response workflows (; ). Artificial Intelligence (AI) and Machine Learning (ML) have therefore emerged as tools for automated analysis of visual and spectral data in plant disease detection (; ). Deep learning (DL) architectures, including convolutional neural networks (CNNs), MobileNets, Vision Transformers, and emerging foundation or vision-language models, can achieve strong performance in plant disease recognition from RGB, UAV, multispectral, or hyperspectral imagery (; ; ). Nevertheless, high predictive accuracy alone is not sufficient for phytosanitary decision support: transparency, probability calibration, leakage-aware validation, interpretability, explainability and external testing are also required (; ; ).

Radiomics is an approach for extracting quantitative descriptors that characterize intensity, shape, and texture patterns from images (). Originally developed in medical imaging, radiomics is conceptually aligned with image-based plant phenotyping because both aim to convert image patterns into quantitative, analyzable traits (). Texture descriptors derived from gray-level co-occurrence matrices (GLCM), gray-level run-length matrices (GLRLM), gray-level size-zone matrices (GLSZM), gray-level dependence matrices (GLDM), and neighboring gray-tone difference matrices (NGTDM) can capture lesion heterogeneity, tissue irregularity, and spatial organization; these feature families are widely used in standardized radiomics implementations such as PyRadiomics and follow IBSI-oriented definitions (; ). When combined with classical ML classifiers, radiomics enables models whose predictions can be inspected at the feature level using permutation importance, SHAP, LIME, and accumulated local effects (; ).

The methodological gap addressed in this study is not the absence of high-accuracy image classifiers. Rather, the gap lies in the limited availability of calibrated, leakage-aware, and feature-level interpretable baselines for multiclass banana disease-pattern characterization involving H ealthy, S igatoka, and F usarium wilt categories. This study evaluates a radiomics-based ML framework together with three CNN baselines (MobileNetV3, EfficientNet, and ResNet-18), to quantify the accuracy–interpretability trade-off under the same data partition. The radiomics model is positioned as a complementary, interpretable and auditable reference model, whereas the CNN baselines establish strong performance image-based benchmarks.

The remainder of this paper is organized as follows. Section 2 reviews recent AI-based approaches for plant disease characterization, including DL, interpretable models, calibration, and leakage-aware validation. Section 3 describes the data sources, leakage-aware partitioning, HSV-based segmentation, radiomic feature extraction, feature preprocessing, classical classifiers, CNN baselines, probability calibration, and statistical analyses. Section 4 presents the comparative results, bootstrap confidence intervals, McNemar testing, CNN baseline comparison, calibration assessment, and explainability analyses. Section 5 discusses the biological plausibility, operational implications, domain-shift limitations, and deployment constraints. Section 6 concludes with future directions for Ecuadorian field validation and hybrid interpretable DL systems.

2 Related work

2.1 AI-based approaches for plant disease characterization

Fusarium wilt, in particular TR4, has attracted increasing research attention because of its rapid spread and impact on banana production (). Existing characterization strategies focus primarily on molecular diagnostics, soil assays, and expert visual inspection (). While these methods are reliable, they are not easily scalable for continuous field monitoring. Recent work on Fusarium TR4 in Latin America and the Caribbean has emphasized that effective management requires surveillance, biosecurity, containment, resistant cultivars, and coordinated regional response strategies (). In addition, the socioeconomic effects of TR4 on banana producers in Latin America further reinforce the need for scalable monitoring and early-warning tools (). AI-driven approaches for Fusarium wilt characterization remain more limited than those for foliar diseases, and many image-based studies do not explicitly address differentiation between Fusarium wilt and other diseases with overlapping visual characteristics.

Black Sigatoka has been more extensively studied using image-based techniques, but many studies address binary classification or disease severity estimation rather than multiclass differentiation among healthy leaves and multiple disease types. Recent studies using UAV-based multispectral imagery and DL have demonstrated the feasibility of plantation-level monitoring; however, these approaches remain sensitive to dataset scale, geographic variability, acquisition devices, and environmental conditions (). These limitations are relevant for banana disease recognition because cultivar diversity, field management, illumination, canopy structure, and regional disease pressure can affect the visual appearance of symptoms and the generalization of image-based models.

AI techniques have been explored for automating the recognition of disease patterns, mainly based on computer vision and using ML techniques. Early approaches focused on handcrafted features combined with classical classifiers, and recent reviews of plant disease detection emphasize that DL has become a dominant modeling paradigm because of its capacity to learn discriminative visual features directly from images (; ). Compact CNNs such as MobileNet variants are particularly relevant for agricultural applications because they reduce parameter count and computational requirements relative to larger architectures, whereas EfficientNet and ResNet families remain common image-classification benchmarks. At the same time, recent literature highlights continuing challenges in cross-domain generalization, benchmark design, data leakage, annotation quality, and interpretability (; ).

CNN-based approaches have been applied to detect foliar diseases such as Black Sigatoka, using leaf images captured under natural illumination. Previous studies (; ) have reported high classification accuracy using architectures such as VGG, ResNet, MobileNet, and related CNN variants, showing the potential of DL for large-scale disease monitoring (; ). UAV-based and multispectral imaging combined with deep neural networks have further extended these capabilities to plantation-level analysis, including Black Sigatoka monitoring in banana crops (; ). However, these models often require sufficiently large and well-annotated datasets, and their performance may be sensitive to domain shifts caused by illumination, background variability, acquisition devices, geographic conditions, and canopy structure. In addition, many DL models operate through latent feature representations, which can limit direct feature-level interpretability and complicate adoption in decision-support or regulatory contexts (; ).

Parallel to DL developments, classical ML methods using handcrafted or radiomic descriptors remain relevant, particularly in scenarios where calibration, computational constraints, auditability, or explainability are important. Texture-based descriptors, including Haralick features derived from gray-level co-occurrence matrices (GLCM), have been widely used to characterize plant disease symptoms because they capture spatial heterogeneity and lesion structure (; ; ). These approaches are generally more transparent at the feature level than end-to-end DL models, although they may provide lower raw predictive accuracy. This motivates a direct comparison between calibrated radiomics models and CNN baselines under the same data partition, allowing the accuracy–interpretability trade-off to be quantified empirically.

2.2 Radiomics, calibration, and explainability in agricultural AI

Radiomics extends traditional texture analysis by extracting a large set of quantitative descriptors that characterize intensity distributions, shape, spatial heterogeneity, and multiscale texture (). Originally developed in medical imaging, radiomics has recently been adopted in agricultural and plant phenotyping contexts, where subtle morphological and textural changes may reflect disease progression, stress responses, or tissue degradation (; ). These descriptors provide an intermediate representation between raw image pixels and model predictions, enabling quantitative characterization of visual patterns that may otherwise be difficult to inspect directly.

Recent research has demonstrated the utility of radiomic and texture-based features, such as first-order statistics, shape descriptors, GLCM, GLRLM, GLSZM, GLDM, NGTDM, and wavelet-derived descriptors, for characterizing heterogeneous image patterns associated with plant stress and disease. Foundational texture descriptors (; ), including Haralick features derived from gray-level co-occurrence matrices, remain relevant for methodological definitions because they capture spatial heterogeneity and local gray-level relationships. In plant pathology, these feature families are useful because fungal infections can induce lesion coalescence, tissue irregularity, necrotic patterns, and non-uniform spatial organization, all of which can be represented through quantitative shape and texture descriptors.

Beyond predictive accuracy, recent research also emphasizes the importance of model calibration and interpretability for deploying AI systems in high-stakes or decision-support domains, including agriculture (; ). Poorly calibrated models may produce overconfident predictions, leading to inappropriate prioritization, inspection, or management decisions. Techniques such as Isotonic Regression () and Platt scaling () have been proposed to improve probability reliability when predicted probabilities are used for risk-aware decision support (; ). In phytosanitary workflows, calibration is relevant because model outputs may be used to define inspection thresholds, prioritize uncertain cases, or communicate confidence to agronomists and plant-health authorities.

Explainable AI (XAI) methods, including Permutation Importance, SHAP, LIME, and Accumulated Local Effects (ALE), are increasingly used to interpret model behavior and identify features associated with model predictions (; ; ). Recent agricultural AI literature has further emphasized the role of XAI in improving transparency, interpretability, and trust in AI-driven decision-making systems (; ). These techniques are particularly valuable in radiomics-based models because explanations can be linked to explicit image descriptors such as perimeter, sphericity, intensity percentiles, and texture non-uniformity. This feature-level interpretation distinguishes radiomics from end-to-end CNN models, where explanations often require post-hoc saliency or attribution methods over latent representations.

Recent work on data splitting in biological ML has also shown that random or naïve partitions may inflate performance when related biological entities, duplicated observations, or highly similar samples appear in both training and test sets. () introduced DataSAIL as a leakage-reduced splitting strategy intended to enable more realistic evaluation of ML models in biological data, particularly under out-of-distribution settings. This issue is relevant for agricultural image datasets because images may share acquisition conditions, plant identity, plot identity, or visually similar backgrounds. Therefore, leakage-aware validation is necessary to avoid overly optimistic performance estimates and to better approximate deployment conditions.

While DL approaches dominate recent studies on plant disease recognition, there remains a need for calibrated, leakage-aware, and feature-level interpretable frameworks tailored to multiclass banana disease-pattern characterization. Radiomics-based ML approaches provide a complementary reference model for this purpose as an auditable framework that allows probability calibration and inspection of quantitative descriptors associated with disease-related image patterns. In this study a joint evaluation of a calibrated radiomics model and CNN baselines under the same data partition is proposed.

3 Materials and methods

3.1 Data sources and leakage-aware partitions

Two publicly available banana leaf image repositories were used to construct a three-class dataset comprising H ealthy, S igatoka, and F usarium wilt race 1 (F usarium1) samples. Images corresponding to Fusarium wilt and a subset of Healthy leaves were obtained from the Banana Leaves Imagery Dataset collected in Tanzania (; ). Additional Healthy samples and Sigatoka images were sourced from the Banana Leaf Disease Dataset Bangladesh (). All images were RGB and organized in class-specific directories prior to preprocessing (Figure 1).

Figure 1

Following image segmentation and radiomic feature extraction, the resulting dataset consisted of N = 14, 763 radiomic samples. The class distribution after radiomic extraction was 4,866 samples for the H ealthy class, 5,767 samples for the S igatoka class, and 4,130 samples for the F usarium1 class (Figure 2). This distribution reflects moderate class imbalance, motivating the use of balanced accuracy and macro-averaged F1-score throughout the study.

Figure 2

To reduce the risk of information leakage, the experimental protocol used a leakage-aware partitioning strategy based on original image identity. A group identifier was constructed from the original image path/stem, and samples derived from the same source image identity were assigned to only one of the training, validation, or test subsets. The dataset was partitioned into training, validation, and test subsets with proportions of 70%, 15%, and 15%, respectively. The test subset was held out and remained frozen during model selection, hyperparameter tuning, feature preprocessing, and probability calibration.

The public repositories used in this study do not provide verified plant-level, plot-level, or farm-level identifiers. Therefore, the present validation should not be interpreted as a plant-level or plot-level external validation. A stricter plant/plot-level or similarity-aware split, such as DataSAIL-like splitting (), would be preferable when the corresponding metadata or image-similarity matrices become available.

3.1.1 Definition of experimental unit and radiomic samples

In this study, the experimental unit was defined at the image level. Each RGB image obtained from the public banana leaf repositories was associated with one class label according to the original dataset annotation: H ealthy, S igatoka, or F usarium1. After HSV-based segmentation, a leaf-level region of interest (ROI) was extracted from each image and used for radiomic feature computation. Thus, each radiomic sample corresponds to one image-derived observation represented by a 103-dimensional radiomic feature vector.

The final dataset contained 14,763 radiomic samples, corresponding to