Sequential Transfer Learning for Plant Disease Recognition Across Perennial Crop Domains Using MobileNetV2-CBAM
Purpose: This study investigates whether a learned disease-recognition representation can be progressively reused across distinct perennial crop domains and how its predictive behaviour changes as target-domain heterogeneity increases. Design/methodology/approach: A lightweight MobileNetV2 framework enhanced with the Convolutional Block Attention Module (CBAM) was evaluated through sequential transfer from Walnut to Cashew, Coffee and Mango, followed by validation on the heterogeneous ABCGMP six-crop benchmark. Each stage used the preceding model representation and was evaluated on an independent test set using complementary classification measures. Findings: Test accuracy increased from 75.76% for the Walnut baseline to 84.85% after CBAM integration, followed by 90.67% on Cashew, 100.00% on Coffee and 99.50% on Mango. The final ABCGMP evaluation achieved 88.44% accuracy, 81.49% Balanced Accuracy and 0.9949 ROC-AUC. The results show that substantial predictive capability was retained through successive crop transfers, while broader multi-crop heterogeneity introduced a measurable performance constraint. Originality: The study provides an empirical characterisation of sequential knowledge reuse across multiple perennial crop domains using a common lightweight attention-enhanced representation, extending evaluation beyond isolated source-to-target transfer.