Mulberry Leaf Disease Detection Using Machine Learning and Deep Learning: A Systematic Literature Review

Abstract

Sericulture—the industry of rearing silkworms to harvest silk—depends almost entirely on the sustained health of the mulberry plant (Morus spp.), since mulberry foliage constitutes the sole feed for the domesticated silkworm Bombyx mori. When foliar diseases such as leaf rust, powdery mildew, anthracnose, leaf spot, and leaf blight take hold, the consequences propagate directly into silk yield and farmer income—losses that the Food and Agriculture Organization has estimated at 10–30% annually in heavily affected regions [7]. Conventional diagnosis, which depends on trained agronomists conducting on-site visual inspections, cannot meet the speed and scale demands of modern plantation management. Over the past four years, researchers have responded by adapting an expanding suite of machine learning (ML) and deep learning (DL) tools to this specific problem. The present work synthesizes fifteen peer-reviewed publications from 2021 to 2025, selected through a PRISMA-adapted protocol, and evaluates them across six analytical dimensions: model architecture, training data, performance benchmarks, interpretability mechanisms, deployment readiness, and stated limitations. Read collectively, the corpus traces an unmistakable methodological progression—from support vector machine (SVM) hybrids anchored at 77.5% classification accuracy, through CNN transfer learning and attention-augmented networks occupying the 90–98% band, to purpose-engineered architectures such as BerryNet that reach 99.09%. Alongside this performance arc, explainability has transitioned from an optional enhancement to a near-standard requirement: Grad-CAM, SHAP, and LIME appear individually or in combination across eight of the fifteen reviewed papers. Three structural deficits persist across the entire body of work—the lack of a geographically representative benchmark dataset, the near-absence of validation under genuine field conditions, and the failure to connect detection outputs to practical management guidance. These gaps, rather than architectural limitations, now represent the primary barriers to real-world impact.