Mango-Mamba and VN-MangoLeaf: A lightweight Mamba model and New Dataset for Mango leaf disease classification

Mango leaf disease represents a significant threat to fruit quality and yield, necessitating highly accurate, real-time detection systems. However, existing Deep Learning approaches, particularly Transformer-based models, often suffer from prohibitive computational complexity (quadratic scaling), limiting their deployment on resource-constrained edge devices. To address this challenge, this study introduces MangoMamba, a novel lightweight hybrid architecture specifically optimized for mobile deployment. The proposed model integrates Multi-Scale Mamba Mixers with Large-Kernel Attention mechanisms within a hierarchical four-stage framework, enabling linear computational complexity while preserving global receptive fields. Experimental evaluations were conducted on the MangoLeafBD dataset and the newly curated VN-MangoLeaf dataset, which comprises 7000 images of Vietnamese mango varieties. Results demonstrate that MangoMamba achieves competitive classification accuracies of 99.75% and 98.71% on the respective datasets. Crucially, the model exhibits exceptional efficiency with only 5.8 million parameters and an inference latency of 1.46 ms per image on T4 GPU, approximately 80 times faster than recent ViX-MangoEFormer architectures. Furthermore, the practical feasibility of the proposed approach is validated through a functional Android application capable of offline inference (100–300 ms latency) on standard smartphones. These findings confirm that MangoMamba establishes a new competitive trade-off between accuracy and efficiency for smart agriculture applications. • MangoMamba achieves ≤ 1.52 ms GPU inference with competitive accuracy. • Hybrid Mamba-Attention model offers linear complexity for disease detection. • VN-MangoLeaf dataset introduced: 7000 images from Vietnam with Red Rust. • Three-phase curriculum learning enables cross-geographical knowledge transfer. • Lightweight ≤ 25 MB model enables 100-300 ms smartphone-based detection.