A Three-Tiered Hierarchical Memory Network With Digital Twins for Adaptive Crop Growth Optimization in Smart Agriculture

As a core driver of global agricultural modernization, smart agriculture still faces prominent bottlenecks including high data collection costs, low efficiency of multisource information fusion, and real-time decision-making delays, which hinder its large-scale application. To address these challenges, this study proposes an intelligent agricultural monitoring and optimization framework integrating Internet of Thing (IoT) sensor networks, digital twin technology, and a three-layer hierarchical memory network (HMN). The key innovations of this work are threefold: to begin with, the hierarchical memory structure originally designed for natural language processing (NLP) is extended to a three-tier architecture (short-term, long-term, and permanent memory layers) customized for the spatiotemporal characteristics and physiological dynamics of agricultural data; furthermore, a closed-loop integrated system is constructed, where IoT sensor networks reduce data redundancy via dynamic sampling, digital twin technology enables real-time data synchronization and control strategy prevalidation, and the HMN realizes adaptive regulation throughout the production process; finally, agriculture-specific algorithms are designed for each HMN layer, specifically including the crop growth prediction and control (CGPAC) algorithm for real-time irrigation and fertilization control, the intelligent collaborative regulation (ICR) algorithm for simulating plant hormone regulatory mechanisms to capture physiological trends, and the intelligent monitoring and planting optimization (IMAPO) algorithm for efficient network topology optimization and resource scheduling. Experimental results demonstrate that the proposed framework outperforms state-of-the-art methods [gated recurrent unit (GRU), long short-term memory (LSTM), Transformer, convolutional neural network-long short-term memory (CNN-LSTM)] across multiple performance metrics, achieving lower prediction errors, higher coefficients of determination, reduced energy consumption, and competitive inference time. This work provides critical technical support for advancing the precision and sustainable development of smart agriculture.