A Comprehensive Review of Hyperparameter Tuning Techniques with Special Reference to Mango Leaf Disease Detection

Abstract

Mango is among the most economically significant fruit crops cultivated across South and Southeast Asia, and its yield is mostly affected by diseases such as anthracnose, powdery mildew, bacterial canker, gall midge infestation, and sooty mould. Convolutional Neural Networks (CNNs) and deep learning architectures are the most popular for automating the diagnosis of such diseases directly from leaf images, but the reported accuracy of these models is highly sensitive to the choice of hyperparameters, including learning rate, batch size, optimizer type, number of epochs, network depth, dropout rate, and kernel configuration. This paper presents a comprehensive review of hyperparameter tuning (also called hyperparameter optimization, HPO) techniques and evaluates their applicability, strengths, and limitations for the specific problem of mango leaf disease detection. Manual tuning, grid search, and random search are examined alongside more advanced strategies including Bayesian optimization, evolutionary and genetic algorithms, swarm-intelligence-based methods (particle swarm optimization, African buffalo optimization), and reinforcement-learning-driven neural architecture search. A structured review of recent mango leaf disease detection studies is synthesized to map which HPO strategies have been used, on which architectures, and with what reported performance gains. Based on this synthesis, several research gaps are identified, most notably the scarcity of systematic, architecture-agnostic comparisons of HPO methods on mango-specific datasets, the limited use of computationally efficient search strategies suited to resource-constrained agricultural deployment, and the general absence of standardized benchmarking protocols.