Multivariate Regressive Gradient Spiral Optimized Deep Belief Learning For Crop Yield Prediction

Abstract

Agriculture plays an important role in financial system of many countries like India. Crop yield prediction in agriculture sector involves amount of crops estimation that are harvested from land. Different machine learning models face challenges like time-efficient crop yield prediction with improved accuracy level. In order to improve the crop yield prediction accuracy with minimal time consumption, a novel deep learning model called Multivariate Regressive Gradient Spiral Optimized Deep Belief Learning (MRGSODBL) method is developed. The proposed MRGSODBL method comprises of five methods namely data acquisition, preprocessing, feature selection, classification and hyperparmater tuning to enhance the accuracy with minimum prediction error. In the data acquisition phase, a
number of data samples are collected from the dataset. After that, the proposed deep belief learning classifier is used for accurate yield prediction with minimal time consumption and error rate. The proposed deep belief learning classifier comprises of different layers, namely one input layer, one output layer and numerous hidden layers for crop yield prediction. Initially, the numbers of data samples are given to input layer of deep learning architecture. After that, the collected data are transmitted to the hidden layer 1. In that layer, data preprocessing is carried out to obtain the suitable format of the dataset by handling the missing data. Followed by, Censored Feature Regression Analysis is carried out in hidden layer 2 for selecting the suitable features to minimize the dimensionality of the dataset. Then the selected significant features are transmitted to the next hidden layer where the Multivariate Adaptive Regression Splines is employed in hidden layer 3 for crop yield prediction. Then, the adaptive gradient spiral optimization algorithm is employed in hidden layer 3 for tuning the hyperparameters of the deep belief learning classifier to minimize the error rate in the crop yield prediction. Finally, the accurate prediction results with minimal error are displayed at the output layer. Experimental evaluation considered the factors like crop yield prediction accuracy, precision, recall, specificity, root mean square error, F1 measure, and prediction time with respect to the number of data samples. The quantitatively analyzed results indicate that the proposed MRGSODBL method attains higher crop yield prediction accuracy with minimal time consumption when compared to conventional techniques.