A Systematic Review of the Performance of Machine Learning Algorithms in Detecting Black Sigatoka in Banana Crops (#416)

Authors

Castañeda-Cubas, Gustavo

Hoyos-Obando, Fabrizio Raúl

Dios-Castillo, Christian Abraham

Abstract

The banana crop faces critical phytosanitary challenges, notably Black Sigatoka, which significantly reduces productivity. Traditional visual inspection methods are slow and subjective, prompting the use of artificial intelligence (AI) to enhance detection efficiency. This study presents a systematic literature review focused on evaluating the performance of AI models in detecting diseases and assessing banana ripeness, based on experimental studies published between 2020 and 2025. The review applied the PRISMA protocol and included 55 articles indexed in Scopus and Web of Science, selected using strict inclusion criteria involving real image datasets and performance metrics. The results indicate that deep learning models, particularly convolutional neural networks (CNN), outperform traditional machine learning approaches. Among them, SqueezeNet emerged as the most effective model for detecting Black Sigatoka, with precision values exceeding 97% in studies specifically focused on this disease. Other high-performing models include ShuffleNetv2 and MobileNetv3Small. While ResNet50 exhibited excellent metrics in broader disease classification contexts, the data reveal that lightweight architecture such as SqueezeNet offers the highest consistency and performance in field conditions. The study concludes that SqueezeNet is the most suitable architecture for detecting Black Sigatoka, due to its high precision, lightweight structure, and consistent performance across evaluations. These findings support the adoption of efficient deep learning architecture in precision agriculture scenarios.