This study presents an innovative approach for enhancing the accuracy and efficiency of plant disease detection through a feature selection-based deep learning mechanism. The proposed methodology consists of several key stages, including image collection, denoising, quality enhancement, segmentation, feature extraction with selection, and classification. Initially, input samples are collected from a publicly available dataset, followed by denoising using an Extended Guided Filtering (EGF) approach and quality enhancement employing a Flexible Mean Adjustment (FMA) method. Precise segmentation of disease-affected portions is achieved through a new Extended Deep-view Fuzzy C-Means (ED-FCM) approach. Subsequently, a Modified Lyrebird assisted DenseNet-201 (ML_D201) model is employed for feature extraction, and optimal feature sets are selected using a Modified Lyrebird Optimization (MLO) approach. Disease classification is performed using a novel Stacked Capsule Squeeze Excitation based Bidirectional Long Short-Term Memory (SCSE_BiLSTM) model, with parameters fine-tuned using a Chaotic Artificial Hummingbird (CAH) approach. The proposed methodology effectively addresses the challenges of dimensionality and processing time, resulting in improved plant disease classification accuracy.