A Hybrid Enhancing And Optimizing Crops Disease And Land Cover Classification Using Adaptive Recurrent FusionNet Framework

Automated detection of crop leaf diseases and classification of remote sensing land cover categories remain challenging owing to complex backgrounds, illumination vari-ability, spectral distortions, and high intra-class visual similarity. Existing frameworks provide strong baselines but commonly suf-fer from scale-sensitive segmentation, redundant feature fusion, limited contextual representation, and slow convergence in high-dimensional feature spaces. This paper proposes the Adaptive Recurrent FusionNet (ARFusionNet) framework — a Flask-based web application that integrates four coordinated inno-vations: Multi-Scale Adaptive Contrast Normalisation (MACN) for illumination-robust preprocessing; Graph-Based Superpixel Attention Segmentation (GSAS) for adaptive Region-of-Interest extraction; Bidirectional Gated Recurrent Units (BiGRU) embed-ded within Residual Efficient Convolution Blocks with Adaptive Weighted Feature Aggregation (AWFA); and Hybrid Binary Differential Evolution controlled Particle Swarm Optimisation (BDE-PSO) for efficient feature selection. DenseNet121 serves as the backbone feature extractor. We validate the system on the Plant Pathology 2020 dataset (1,821 high-resolution apple leaf im-ages; four disease classes). ARFusionNet achieves 98.2% classifi-cation accuracy, surpassing the state-of-the-art baseline (97.6%), while reducing training time by approximately 78 seconds and remaining fully executable on a standard CPU laptop without GPU dependency. The accompanying web application exposes eight interactive diagnostic modules including leaf visualisation, Canny edge display, convolved feature maps, neural network architecture visualisation, and real-time per-image prediction.