An Automated Framework of Superpixels-Saliency Map and Gated Recurrent Unit Deep Convolutional Neural Network for Land Cover and Crops Disease Classification

Abstract

In this work, we proposed an automated deep learning and saliency map architecture for the segmentation of crops, leaf disease segmentation, and land cover classification. The proposed framework is based on two embedded steps. In the first step, crop leaf disease segmentation was performed using superpixel clustering-based saliency maps and Bayesian optimization. The contrast enhancement technique is designed in the first segmentation phase and is passed to the saliency technique for the disease segmentation. In the second phase, EfficientNet-b0 architecture is fine-tuned with hyperparameters optimized via Bayesian Optimization. Also, the fine-tuned model is embedded with a single self-attention residual block fused with an efficient average pool layer. Training has been performed on segmented and contrast-enhanced images that were later fused using a serial-embedded approach. The extracted features in the testing phase are further optimized using the modified moth flame-controlled bisection (MFcB) technique. Finally, the extracted features are classified using machine learning classifiers for the final classification. Experiments are performed on the publically available cucumber leaf dataset and Remote sensing dataset with an improved accuracy of 97.6% and 92.90%, respectively. A comparison with state-of-the-art techniques shows that the proposed architecture has improved performance.