Reducing catastrophic forgetting in CNNs for plant stress classification using continual learning

Plants face a range of biotic and abiotic stresses that reduce yield, and in agricultural countries like Bangladesh, manual detection of these stresses remains slow and error-prone. Convolutional Neural Networks (CNNs) classify plant leaves accurately, but they suffer from catastrophic forgetting when trained on tasks sequentially. This is a major obstacle in real agricultural settings, where new crops and stress conditions arrive over time. Existing continual learning (CL) studies in this domain rely on relatively heavy backbones, leaving open the question of whether lightweight CL pipelines can retain prior-task knowledge under realistic resource constraints. We investigate this question by integrating two established CL methods, Elastic Weight Consolidation (EWC) and Learning without Forgetting (LwF), into EfficientNet-B0, a backbone with only 5.3M parameters. The setup is evaluated on the Nutrispace cucurbit nutritional deficiency dataset, where three plant species (ash gourd, bitter gourd, and snake gourd) are treated as three sequential tasks, each with the same three classes: healthy, nitrogen deficiency, and potassium deficiency. Without continual learning, accuracy on the earliest task collapses to 30% by the end of training. EWC preserves over 61% accuracy on prior tasks while reaching 98% on the final task, and LwF reaches 98% on the final task with slightly lower retention on earlier ones. Pairwise Welch's t-tests confirm that both methods significantly outperform the baseline ([Formula: see text]) and that EWC retains prior-task knowledge significantly better than LwF ([Formula: see text]). These results show that lightweight CNNs paired with established CL techniques offer a workable path for plant stress classification in resource-constrained agricultural AI.