Designing a squeeze-and-excitation-capsule BiLSTM transformer for plant leaf disease recognition

Abstract

Deep learning (DL) is critical in plant disease recognition and classification with precision like those of expert human evaluators. However, development of effective systems is often disrupted due to the complexity and variability of disease pathogenesis. To address these challenges, this research applies to a hybrid DL architecture that integrates spatial encoding, sequential modelling, and attention for visual recognition. This proposed model can incorporate squeeze-and-excitation (SE) with residual blocks, capsule network (CapsNet), bidirectional long short-term memory (BiLSTM), and transformer network (TransNet)-based attention to realize spatial relationships and long-range dependencies for improving recognition accuracy. The proposed model is assessed on the corn leaf disease dataset (CLDD) and rice leaf diseases dataset (RLDD), and its performance is compared to leading-edge models. CLDD and RLDD achieved 99.88 and 99.10% training accuracy respectively. The area under the curve (AUC) reached almost ceiling recognition on CLDD, with 99.73, 99.96, 99.96, and 99.98% for blight (BL), common rust (CR), gray leaf spot (GL), and healthy (HE) result. RLDD results were also high, with 94.98, 93.70, 97.66, 84.57, 99.58, and 98.85% for bacterial leaf blight (BLB), brown spot (BS), HE, leaf blast (LB), leaf scald (LS), and narrow brown spot (NBS), respectively. The results of these tests show the remarkable promise and performance of the proposed model in plant disease recognition applications.