Advanced Agricultural Decision System using Recurrent Polynomial Network for Multi-Crop Recommendation Tasks

Abstract

Agriculture has transitioned from traditional manual practices to data-driven approaches with the advancement of digital technologies and Machine Learning (ML). Earlier, farmers relied on experience and basic statistical methods, which are insufficient for handling the increasing volume of agricultural data from sensors, weather records, and satellite systems. Traditional systems fail to process large datasets and capture complex relationships among soil properties, climate conditions, and crop health, resulting in low prediction accuracy and inefficient decision-making. This study addresses the lack of an integrated framework capable of handling multiple agricultural prediction tasks simultaneously. A multi-task agricultural analysis framework based on the One Classification and Two Regression Tasks (1CA2RT) approach is proposed. It includes crop disease classification and two regression tasks: Normalized Difference Vegetation Index (NDVI) estimation and harvest time prediction. The framework employs models such as Support Vector Machine (SVM)-1CA2RT, AdaBoost (AB)- 1CA2RT, and Ridge (R)-1CA2RT, along with a hybrid model Hybrid Recurrent Polynomial Ensemble (HRPE)-1CA2RT. The hybrid model integrates a Recurrent Polynomial Network (RPN) using Bidirectional Long Short-Term Memory (BiLSTM) with Ensemble Tao Tree Classifier (ETTC) and Ensemble Tao Tree Regressor (ETTR). Experimental results show that HRPE-1CA2RT achieves superior performance, with 100% accuracy, precision, recall, and F1-score for classification, and an R² score of 1.0000 for NDVI and harvest prediction. This unified framework improves prediction accuracy, consistency, and efficiency, supporting reliable decision-making in modern agriculture.