GhostNetV2–PVT Fusion with LightGBM for Accurate Maize Leaf Disease Classification
Annotatsiya
This study presents a convolution–transformer–LightGBM framework for automatic maize leaf disease recognition. Local appearance is encoded by GhostNetV2, and long-range context is modeled with a Pyramid Vision Transformer. The fused descriptors are compressed through principal component analysis before classification with LightGBM. On the maize leaf dataset, which includes Blight, Common Rust, Gray Leaf Spot, and Healthy classes, the system reaches 93.45% overall accuracy, 93.45% recall, 93.99% precision, and a Matthews Correlation Coefficient of 91.23%. Rather than optimising a single end-to-end network for latency or footprint, the proposed multi-stage architecture prioritises diagnostic reliability: CNN and transformer backbones supply complementary spatial and contextual cues, PCA removes redundancy while preserving variance, and LightGBM produces calibrated, interpretable probabilities. Comprehensive statistical analyses—class-wise scores, confusion matrices, precision–recall curves, and ROC characteristics demonstrate consistently balanced behaviour across all four disease categories, confirming the model's suitability for real-world agronomic decision support.