Conv2D Xception Adadelta Gradient Descent Learning Rate Deep learning Optimizer for Plant Species Classification

Almost every element of living beings is undergoing extraordinary transformation as a result of technological advances. While many areas of agriculture are leveraging cutting-edge technological innovations, farming remains far behind in adopting these advancements. Ayurvedic medicine and other plant-based medical systems employ the classification of plants by their leaflets for a variety of purposes, including ecosystem, farming, disease diagnosis, rare plant maintenance, and Herbal remedies. Due to the utilization of biological nomenclature, standard methods of plant recognition and classification are difficult, time-consuming, and aggravating for non-experts. For learners who are seeking to learn about species, this poses a real challenge. Based on this consequence, this paper recommends Conv2D Xception Adadelta Gradient Descent (CXAGD) Deep Learning to classify the plant species based on the leaflet feature structure. The proposed CXAGD model uses Leaf classification dataset from KAGGLE with 4500 plant leaflets of various species. One input, a dense average pooling layer, and an output layer were used in the development of the proposed Conv2D Adadelta Gradient Deep CNN model. The CXAGD model was constructed using 36 depth wise separable convolutional layers and a maximum pooling layer for each convolution and trained with learning rate of Adadelta optimizer. The CXAGD built with Xception network with the Adadelta optimizer with the dropout of 0.5, weight decay of L2 regularization. The leaf dataset subjected to the training and testing data division and the training data applied to CXAGD model and also applied to other CNN models for comparing the efficiency of the proposed CXAGD model. Python was used for programming, with a batch size of 64 and 30 training epochs, on a Geforce Tesla V100 Graphics card server. Implementation outcomes projects that proposed CXAGD model exhibits accuracy of 97.85%, Precision of 97.42%, Recall of 97.75% and FScore of 97.76% when compared with existing CNN models.