Scholarlink

<p>Precision agriculture is evolving with the integration of IoT and deep learning, enhancing efficiency and productivity in crop management. Smart farming demanded accurate yield prediction systems to optimize resource allocation, reduce waste, and enhance food security. However, traditional methods often fail to capture complex relationships among environmental factors, soil conditions, and crop growth patterns, leading to inaccurate forecasts. This study presented an efficient IoT-driven framework that integrated multi-modal agricultural information with advanced deep learning architectures. The framework was trained and evaluated on multiple datasets, including SICKLE (388 farm plots, 2370 season-wise samples, covering rice, maize, and sugarcane in the Cauvery Delta, India), and benchmark datasets CropNet (2291 U.S. counties) and CropHarvest (~ 95,186 global samples) to ensure robustness and generalizability. Structure-Aware Adaptive Bilateral Texture Filtering (SABTF) to reduce noise while preserving critical structural patterns in sensor and imagery data. A Multi-Scale Parallel Hybrid Network (MSPHN) was designed to extract hierarchical features from sensor, image, and metadata sources through parallel branches and cross-scale fusion. Plot-level embeddings generated by MSPHN were mapped into a farm graph, where Random Graph Diffusion Attention Network (RGDAN) simulated stochastic diffusion processes to capture spatial correlations, irrigation connectivity, and inter-plot dependencies. Attention-driven aggregation of diffusion-aware features enabled robust modelling of complex agronomic interactions. A regression head predicted per-plot yields, and Cleaner-Fish Optimization was employed to fine-tune the network parameters. Experimental results demonstrated that the framework reduced RMSE by 17.6%, decreased MAE by 15.3%, and improved R<sup>2</sup> by 12.8% compared with state-of-the-art baselines. The integration of IoT sensing, MSPHN, and RGDAN established an effective decision-support tool for smart farming, enabling precise yield prediction and sustainable agricultural management.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Efficient IoT-Driven Smart Farming Framework Leveraging Multi-Scale Random Graph Diffusion Parallel Hybrid Networks for Accurate Crop Yield Prediction

  • 期刊论文

  • 发表日期：

    2026年1月27日

查看全文
  • Kunal Devidas Gaikwad,
  • S. Sankara Narayanan

摘要

Precision agriculture is evolving with the integration of IoT and deep learning, enhancing efficiency and productivity in crop management. Smart farming demanded accurate yield prediction systems to optimize resource allocation, reduce waste, and enhance food security. However, traditional methods often fail to capture complex relationships among environmental factors, soil conditions, and crop growth patterns, leading to inaccurate forecasts. This study presented an efficient IoT-driven framework that integrated multi-modal agricultural information with advanced deep learning architectures. The framework was trained and evaluated on multiple datasets, including SICKLE (388 farm plots, 2370 season-wise samples, covering rice, maize, and sugarcane in the Cauvery Delta, India), and benchmark datasets CropNet (2291 U.S. counties) and CropHarvest (~ 95,186 global samples) to ensure robustness and generalizability. Structure-Aware Adaptive Bilateral Texture Filtering (SABTF) to reduce noise while preserving critical structural patterns in sensor and imagery data. A Multi-Scale Parallel Hybrid Network (MSPHN) was designed to extract hierarchical features from sensor, image, and metadata sources through parallel branches and cross-scale fusion. Plot-level embeddings generated by MSPHN were mapped into a farm graph, where Random Graph Diffusion Attention Network (RGDAN) simulated stochastic diffusion processes to capture spatial correlations, irrigation connectivity, and inter-plot dependencies. Attention-driven aggregation of diffusion-aware features enabled robust modelling of complex agronomic interactions. A regression head predicted per-plot yields, and Cleaner-Fish Optimization was employed to fine-tune the network parameters. Experimental results demonstrated that the framework reduced RMSE by 17.6%, decreased MAE by 15.3%, and improved R2 by 12.8% compared with state-of-the-art baselines. The integration of IoT sensing, MSPHN, and RGDAN established an effective decision-support tool for smart farming, enabling precise yield prediction and sustainable agricultural management.