Adaptive generalized regressive deep convolutional reinforcement learning for crop yield prediction in smart precision farming

Abstract

Accurate crop yield assessment is essential for improving agricultural productivity and resource management. However, yield prediction is challenging due to multiple influencing factors such as soil properties, weather conditions, crop practices, pests, and diseases. While deep learning techniques have improved prediction capabilities, achieving high accuracy with low error and reduced computation time remains a key concern for large-scale datasets. To address this, an Adaptive Generalized Regressive Deep Convolutional Reinforcement Learning (AGR-DCRL) model is proposed to enhance prediction accuracy for smart farming. The model consists of input, convolution, pooling and dense layers. Data is first collected in data harvesting using IoT technologies that monitor weather, soil conditions, and pesticide usage. During data augmentation, Adaptive proximity sampling process is utilized in input layer to create new data samples. In the convolution layer, preprocessing is performed by using weighted local similarity-based imputation and Generalized Tietjen–Moore test. The missing values are handled and outlier’s are detected. Feature selection is applied in pooling layer using Camargo’s adaptive diversity index to select the most relevant features and remove irrelevant features for reducing the dimensionality. Crop yield prediction is performed at dense layer for analyzing both extracted features and data samples via polytomous logistic regression. The output layer employs a softmax activation function to create multi-class prediction results. Then, the error rate is measured with each prediction outcome and rewards. The Q-values are iteratively updated based on rewards until the model converges. Lastly, the accurate crop yield prediction is obtained with higher accuracy and lesser error. Experimental assessment of the proposed technique is implanted in Python using Smart Farming Sensor Data for Yield Prediction dataset with several metrics. Experimental results demonstrate that the AGR-DCRL model achieves higher accuracy by 4%, lesser error rates by 69%, and faster prediction time by 20% compared to conventional deep learning approaches.