A Hybrid VGG16–PCA–Fuzzy Based Framework for the Classification of Morphologically Similar Maize Pests

Abstract Growing food demand continues to place significant pressure on African agriculture, where maize remains a critical staple crop. However, maize production is threatened by pests such as maize stalk borer (MSB), African armyworm (AAW), and fall armyworm (FAW), whose overlapping morphological characteristics make accurate identification challenging. Current existing pest monitoring methods are laborintensive and prone to human error, while standalone convolutional neural networks (CNNs) often struggle with ambiguous decision boundaries between morphologically similar pest classes. To address these limitations, this study proposes a hybrid VGG16–PCA–Fuzzy framework for fine-grained maize pest classification. The proposed framework introduces a Gaussian membership-based uncertainty modelling mechanism within the latent feature space to model overlapping inter-class and intraclass pest feature distributions. The convolutional layers of the pretrained VGG16 network extract deep visual features from pest images, which are subsequently projected into a lower-dimensional PCA component space. The resulting PCA components are then fuzzified using either triangular or Gaussian membership functions before softmax classification. Four architectures were evaluated: standalone VGG16, VGG16–PCA, VGG16–PCA–TMF, and VGG16–PCA–GMF, using 10 independent experimental runs on a dataset comprising 7,610 maize pest images from larval and adult pest stages. Experimental results demonstrate that the proposed VGG16–PCA–GMF framework achieved the best overall performance, attaining 97.17 ± 0.41% accuracy and 97.15 ± 0.41 F1-score, compared to 92.24 ± 0.74% accuracy and 92.16 ± 0.78 F1-score achieved by the standalone VGG16 baseline. Statistical analysis further confirmed the consistency and stability of the proposed framework across independent experimental runs. The findings demonstrate the effectiveness of the proposed hybrid framework, particularly the Gaussian membershipbased variant, for fine-grained discrimination of morphologically similar agricultural pests under visually overlapping conditions.