中国农业大学学报2025，Vol.30Issue(10)：73-90,18.DOI:10.11841/j.issn.1007-4333.2025.10.07
深度学习驱动的土壤信息感知技术:进展、挑战与展望
Deep learning-driven soil information sensing:Advances,challenges,and prospects
摘要
Abstract
The rapid sensing of farmland soil information is as a improtant support for smart agriculture.To clarify the existing issues in the development of soil information sensing technologies,a systematic review and analysis are conducted on the current research status of soil information sensing driven by deep learning.This study focuses is placed on the innovative practices of various deep learning models in addressing traditional challenges associated with soil information inversion under different sensing scales and methods,along with an in-depth examination of the strengths and limitations of these methods and approaches.The findings indicate:1)For proximal sensing systems,deep learning has effectively addressed the coupling complexity of environmental factors and poor adaptability in traditional electrical sensor calibration by constructing various deep neural network models,thereby enhancing the accuracy and robustness of in-situ detection for soil moisture,salinity,and nutrients.Meanwhile,deep learning methods demonstrate significant advantages in overcoming key challenges such as moisture interference and poor cross-temporal/spatial generalization in Vis-NIR spectroscopy-based prediction of soil organic matter/total nitrogen content.2)A wide range of deep learning models and methods have been applied to process spaceborne and airborne remote sensing data,enabling the development of more accurate large-scale inversion models for soil organic carbon,moisture,salinity,and other parameters.These approaches have substantially improved the reconstruction,calibration,and downscaling capabilities of remote sensing data.3)Deep learning-driven multimodal,multi-scale,and multi-temporal information fusion models have gained widespread adoption.The integration of lightweight networks,hybrid models,and attention mechanisms has facilitated effective interaction and fusion of heterogeneous data,significantly enhancing the accuracy and real-time performance of synchronous multi-parameter prediction.4)Emerging research trends highlight the importance of developing inversion models with strong generalization capabilities for complex agricultural environments and constructing multi-scale(proximal-airborne-spaceborne)soil information sensing networks based on swarm intelligence and collaborative sensing.Furthermore,integrating soil continuum model theory,future efforts should prioritize the exploration of in-situ monitoring methods and quantitative correlation models for soil microbial dynamics and organic matter turnover processes,aiming to transition from single physicochemical parameter monitoring to comprehensive soil ecosystem health assessment.关键词
土壤养分/深度学习/传感器/智慧农业Key words
soil nutrients/deep learning/sensors/smart agriculture分类
农业科技引用本文复制引用
李佳怿,刘楠,杨玮,李民赞..深度学习驱动的土壤信息感知技术:进展、挑战与展望[J].中国农业大学学报,2025,30(10):73-90,18.基金项目
国家重点研发计划(2024YFD1500800) （2024YFD1500800）
中国农业大学2115人才工程联合资助 （）
