Under the dual challenges of global climate change and food security, agricultural environmental monitoring is at a critical juncture of transitioning from “passive remediation after the event” to “proactive prevention and control in advance.” With the development of smart agriculture, although mobile robots and multimodal sensors have addressed the problems of low efficiency and high cost in traditional manual sampling, most existing monitoring solutions remain at the “post-event detection” stage—that is, pollution has already occurred before an alarm finally arrives, causing control measures to often lag behind. To truly achieve green and sustainable agricultural development, there is an urgent need to give mobile robots a “smart brain”, enabling them not only to perceive the present but also to accurately predict the dynamic trends of pollutants based on complex environmental data.
However, high-precision prediction is by no means easy. The time series of pollutant concentrations in agricultural environments (such as CO, non-methane hydrocarbons (NMHC), and NOx) are highly nonlinear, non-stationary, and complex, and are highly susceptible to environmental noise interference. When faced with such data, existing machine learning methods often struggle to effectively separate different frequency components in the signal, resulting in insufficient feature extraction; traditional attention mechanisms are inefficient in processing high-dimensional data and ignore the intrinsic information entropy characteristics of the time series itself. How to accurately lock onto key features from noisy data while reducing computational costs has become a bottleneck constraining model deployment on mobile robots.
Recently, the team led by Professor Hui LIU at Central South University published a research paper titled “A hybrid prediction model based on decomposition algorithm and sample entropy-weighted attention mechanism for mobile robot agricultural environment monitoring” in ENGINEERING Agriculture (DOI: 10.15302/J-FASE-2026674). The study innovatively proposed a hybrid deep learning prediction model named VMD-SE-DLAM-LSTM (where VMD is variational mode decomposition, SE is sample entropy, DLAM is double-layer attention mechanism, and LSTM is long short-term memory neural network), cleverly resolving the problem of data complexity.
The model designs a rigorous three-step strategy of “decomposition–weighting–prediction”: first, variational mode decomposition (VMD) is introduced as a data preprocessing step to adaptively decompose complex raw signals into multiple modal components with clear physical meaning in the frequency domain, effectively avoiding the mode mixing problem; second, sample entropy (SE) is innovatively used to measure the complexity and information content of each decomposed sub-sequence, and this is used as prior knowledge to weight features, enabling the model to automatically focus on key features with higher information value while suppressing interference from redundant information; finally, a long short-term memory (LSTM) neural network is used as the core predictor to process the weighted and optimized feature sequences, making it easier to capture long-term dependencies in the data.
Experimental results strongly demonstrated the superiority of this hybrid model. When predicting the three pollutants CO, NMHC, and NO₂, the VMD-SE-DLAM-LSTM model comprehensively outperformed mainstream models such as ARIMA, RNN, SVM, and GRU. In particular, in terms of the prediction correlation (R²) metric, compared with the best-performing GRU model, the proposed model improved by 6.9%, 10.1%, and 4.8%, respectively, showing significant and comprehensive performance advantages, while its computational complexity is fully within the acceptable range for mobile robots.
This study successfully overcame the technical challenges of strong nonlinearity and non-stationarity and high noise interference in agricultural environmental data, empowering mobile robots with the ability to accurately “predict” pollutant concentration trends and raising prediction accuracy to a new level.
The study not only achieved a breakthrough at the algorithm level but also has far-reaching industrial application value. It directly responds to the pain points of “inaccurate measurement and insufficiently early prediction” in agricultural environmental monitoring, significantly reducing key indicators such as mean absolute error (MAE). This provides mobile robots with a true “smart brain”, enabling them to shift from passive remediation to proactive prevention and control. From the perspective of national strategy, this achievement closely aligns with China’s development strategies of “Digital Village” and “smart agriculture”. By building a closed-loop system of “monitoring–prediction–decision-making”, it provides strong technical support for safeguarding national food security and ecological security.
DOI：10.15302/J-FASE-2026674