FuzzyShallow: A framework of deep shallow neural networks and modified tree growth optimization for agriculture land cover and fruit disease recognition from remote sensing and digital imaging.

• A data augmentation process based on contrast enhancement techniques is proposed. • A EfficientNet-b0 via Self-attention, is developed with additional layers and minimization weights. • The model is trained using optimized hyperparameters through Bayesian Optimization. • Fused features and selected best features using an improved tree growth optimization algorithm and fuzzy function. One of the most significant issues in the agricultural and remote (RS) sensing industry is detecting infected plants at the early stages. Deep learning-based methods for classifying and detecting agricultural diseases have shown remarkable achievements in technologically sophisticated horticultural research in recent decades. The shortage of imaging data for training a deep neural network model is challenging for accurately classifying fruit disease. Moreover, the remote sensing imaging data includes complex patterns of plant areas that are difficult to identify using conventional techniques. Remote sensing technology efficiently gathers fruit health data, detects disease signs, and monitors citrus orchards for early diagnosis and prevention. The fuzzy deep neural network techniques performed better for classifying remote sensing and digital imaging data for agriculture. This paper proposes a fuzzy deep learning and optimization-based novel framework for citrus fruit disease and agriculture land cover recognition. The Mendeley dataset and NWPU-RESISC45 are employed in this work for the experimental process. The challenge is that these datasets contain a limited number of images. Also, the classes are imbalanced, degrading the learning capability of the proposed models. Therefore, in the first step, we proposed a contrast enhancement technique based on brightness preserving histogram and entropy that generated the improved images that later merged with original data as an augmentation step. We modified the EfficientNet-b0 model in the next step by adding a few convolutional and self-attention layers. Bayesian Optimization has initialized hyperparameter values such as learning rate and momentum. The modified model is trained separately on original and enhanced images to keep distinct fuzzy information at the output layer. After that, deep features are extracted and fused using an Entropy-Serial approach called improved serial fusion. The fused features set observed a few irrelevant information that was further optimized using an improved tree growth optimization algorithm with a fuzzy function. The selected features are finally classified using shallow neural networks and machine learning classifiers. The experimental process obtained an improved average accuracy of 98% and 96.5% on the Mendeley and NWPU datasets, respectively. A t -test is also conducted to check the means of two classifiers (best and worst). In addition, a comparison is performed with recent techniques, showing improved precision and recall rate. [ABSTRACT FROM AUTHOR]