Effectiveness of Depthwise Separable CNNs for Disease Identification among Various Crops

Agriculture is one of the most important fields worldwide encouraging food security in the face of numerous difficulties including climate alteration, resource availability, and population increase. More recently, Deep Learning (DL) algorithms have appeared as the revolutionary processing technologies capable of handling many of these issues. Deep learning is a subfield of machine learning that has several uses in agriculture. This paper focuses on analyzing the effectiveness of depthwise convolutional models on disease prediction at various crops. Various CNN architectures such as Conventional CNN, Residual model, DenseNet model and Depthwise separable models are compared. Among all the models, depthwise models are giving better accuracies than all other models. In this work, two depthwise models are considered, (i) MobileNet CNN and (ii) Custom Depthwise CNN. Both the depthwise models are giving much better results than convolutional blocks, Residual block and DenseNet blocks. MobileNet gives 93.42%, 94.35%, 93.07% and 93.71% in the metrics Accuracy, Precision, Recall and F1 scores respectively. Custom Depthwise CNN gives next best level of performance than other CNN models.