Artificial Intelligence based on behavioral recognition and optimization for low carbon fertilization in agriculture

Abstract Corresponding Author：Yanmei Yuan and Hui Liu, email:  taiskjxy2020yym@163.com;liuhsdau@163.com

This study addresses the limitation of conventional fertilization monitoring methods that fail to capture dynamic nutrient trends for low-carbon precision management by proposing a data-driven approach for fertilization behavior recognition and low-carbon decision optimization based on multi-source agricultural time-series data. Traditional approaches lack temporal continuity and are unable to support real-time behavioral identification or carbon emission-constrained decision-making. The long short-term memory (LSTM) model is selected for its ability to process long-sequence heterogeneous sensor data and accurately recognize sparse fertilization events under environmental noise, while MILP is used to formulate a globally optimal fertilization plan that minimizes carbon emissions subject to crop nitrogen demand and environmental safety constraints. By deploying soil, meteorological, and crop growth sensors to establish an edge-cloud collaborative architecture, this method enables real-time collection and feature extraction of multi-source heterogeneous farmland data. A behavior recognition model combining a bidirectional LSTM network with an attention mechanism is developed to accurately annotate fertilization events and their temporal and spatial parameters. Carbon equivalent is calculated based on nitrogen dynamic balance and lifecycle carbon emission factors. Using carbon emission minimization as the objective function and crop nitrogen requirement and environmental safety as constraints, a mixed integer linear programming (MILP) model is constructed to generate a low-cost, high-yield fertilization plan. Results show that the system’s prediction error is 8.5% on day 30, and carbon emission intensity is reduced to 0.365 kgCO₂-eq/kg fertilizer, supporting the feasibility of this AI-driven decision framework in terms of behavior recognition accuracy and carbon emission control effectiveness under the tested plot conditions.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1038/s41598-026-57895-3.

Introduction

The continuous increase in fertilizer input in agricultural production has led to a series of environmental and ecological problems, including soil nutrient imbalance, eutrophication of surface runoff, and increased greenhouse gas emissions^1,2. Traditional fertilization methods rely on farmers’ experience and judgment, lacking real-time quantitative monitoring of soil fertility, crop nutrient requirements, and environmental conditions. This results in a large gap between fertilizer use and actual crop uptake, low nitrogen use efficiency, and high carbon emission intensity^3,4. Low-carbon, precision fertilization has become a key direction for sustainable agricultural development, but its practical implementation faces numerous technical and management difficulties^5,6. Most existing fertilization monitoring methods rely on discrete sampling or manual recording, lacking timeliness and continuity, and are unable to reflect dynamic trends in farmland nutrients^7,8. Data processing methods for identifying fertilization behavior rely on manually set rules, making them difficult to adapt to the changing patterns of different crop types and complex field environments^9,10. Decision-making processes lack coupling with carbon emission assessments, and fertilization plan optimization often focuses solely on crop yield or fertilizer efficiency without constraining carbon emissions. This means that even if precise fertilization is achieved, it may not truly achieve low-carbon goals^11,12. In this context, the core research problem is defined as data-driven fertilization behavior recognition and low-carbon decision optimization based on multi-source agricultural time-series data. Such a framework must handle long-sequence heterogeneous inputs, resolve sparse event detection under environmental noise, and embed carbon constraints into actionable plans, enabling precise and sustainable fertilization management^13,14. Deep learning methods for agricultural monitoring can be categorized into unsupervised, self-supervised, and semi-supervised approaches. Unsupervised methods identify patterns in unlabeled sensor data but often lack precision in specific event detection due to the absence of ground truth^15,16. Self-supervised approaches leverage inherent temporal structures to learn representations without explicit labels, demonstrating robustness to environmental noise but requiring extensive pre-training that may not align with agricultural data characteristics^17,18. Semi-supervised techniques combine limited labeled data with abundant unlabeled data, effectively addressing sparse fertilization events under field noise while maintaining computational efficiency for edge deployment^19,20. The proposed framework employs a semi-supervised approach with bidirectional LSTM and attention mechanism, specifically designed to handle the challenges of sparse event detection and environmental noise in farmland monitoring scenarios.

To address these issues, this paper proposes an integrated framework that couples multi-source intelligent sensing with a bidirectional long short-term memory network enhanced by an attention mechanism and mixed integer linear programming. The approach enables continuous, real-time identification of fertilization behaviors under field noise and sparse event conditions, establishes a life-cycle carbon emission assessment model grounded in nitrogen dynamic balance, and generates globally optimal low-carbon fertilization plans under crop nutrient and environmental constraints. The core contribution of this work is a behavior-decision coupled framework that integrates a bidirectional long short-term memory network with attention mechanism for fertilization event recognition and a mixed integer linear programming model for carbon-constrained fertilization optimization. The multi-source intelligent sensing and edge-cloud communication subsystems enable data acquisition but do not constitute the primary novelty. A carbon emission assessment based on nitrogen dynamic balance and lifecycle emission factors is established. Methodological novelty resides in the bidirectional LSTM attention mechanism for sparse fertilization event detection and the MILP formulation with carbon emission minimization constrained by dynamic nitrogen balance. Engineering integration couples edge-cloud architecture with incremental learning. Practical contribution is a feedback loop demonstrated through reduced prediction error and carbon emission intensity in field trials.

Related works

Research on multi-source intelligent sensing technology in the field of agricultural environmental monitoring is undergoing a profound transformation from experience-driven to data-driven. Radocaj D²¹ et al. systematically compared the application efficiency of traditional interpolation methods and modern machine learning algorithms in precision agriculture, combined field soil sampling data with multi-source remote sensing technology, and revealed the advantages of modern methods that integrate machine learning and high-resolution remote sensing data in improving fertilization prediction accuracy and preserving soil spatial heterogeneity. This provides an intelligent solution to the dilemma of high cost and low prediction accuracy of traditional soil sampling, and promotes the paradigm shift of precision fertilization from experience-driven to data-driven. Innovations in data processing methods have directly supported the construction and development of new monitoring systems^22,23. Chen X et al.²⁴ integrated unmanned aerial vehicle hyperspectral remote sensing, Geographic Information System (GIS) spatial analysis, and machine learning technology to build a soil nutrient dynamic monitoring and precision fertilization decision-making system, achieving real-time processing and continuous learning optimization of multi-source data. In field trials across multiple ecological zones, they significantly reduced fertilizer use and increased crop yields, effectively breaking through the static limitations of traditional soil nutrient assessment methods and providing intelligent solutions for sustainable agriculture. The in-depth mining and application of monitoring data has led to the emergence of more refined management strategies^25,26. Toselli M²⁷ et al. constructed a multi-level nutrient management model based on soil fertility monitoring and precision agriculture framework, integrating agronomic management monitoring and dynamic modeling analysis technology under soil climate conditions. They achieved precise optimization of nutrient distribution at different scales, including regions, farms, fields, and plots, effectively balancing crop growth needs with environmental sustainability, and providing a new paradigm for intelligent fertilization decision-making for agricultural production that balances economic benefits with reduced environmental impact. Existing studies generally ignore the actual impact of farmers’ behavior patterns on technology implementation, resulting in the real challenge of low adoption rate when the system is applied in the field.

The application of deep learning technology in the field of low-carbon agricultural decision-making has formed a multi-level research system from spatial planning to behavior recognition. Li S²⁸ et al. constructed a land use change simulation framework by coupling Convolutional Neural Network (CNN)-LSTM and PLUS, and designed a multi-scenario simulation scheme by combining Markov chain and multi-objective planning model. They achieved accurate prediction and sustainability assessment of land use pattern in the cultivated land reserve resource area of Da’an City, Jilin Province, revealing the core advantages of sustainable development scenarios in the coordinated optimization of the three dimensions of ecosystem service value, agricultural production benefits, and carbon balance, and provided a new decision-making support paradigm for the scientific planning of cultivated land reserve resource areas that took into account both ecological security and food security. The accurate simulation results at the regional scale have laid a spatial decision-making foundation for low-carbon management of the entire agricultural industry chain^29,30. Wang X et al.³¹ constructed a low-carbon supply chain management optimization framework for fresh agricultural products based on Recurrent Neural Network (RNN), integrating multi-source time series information such as historical sales data, meteorological conditions, and consumer behavior to achieve accurate demand forecasting and dynamic optimization of transportation routes. While reducing product loss rates, they effectively balanced supply chain efficiency and carbon emission intensity, providing a new intelligent management path for the agricultural product circulation field that takes into account both environmental sustainability and economic feasibility. Low-carbon optimization practices in the circulation link have further promoted innovation in precise decision-making technology at the source of production^32,33. Xiang Z et al.³⁴ quantitatively analyzed the differences in the impact of three types of perceived factors, namely economic risk, social risk, and psychological risk, on farmers’ fertilizer reduction decisions, revealing the dynamic regulatory mechanism of environmental attitudes between risk perception and behavioral choices. This clarifies the vulnerability of farmers’ decisions when facing economic risks and the key role of environmental cognition in risk offsetting. Current research provides theoretical breakthroughs for understanding the behavioral logic of agricultural actors in sustainable transformation, based on a risk-attitude interaction perspective^35,36. As a well-known alternative, dynamic Bayesian networks require only a small number of training samples while achieving comparable predictive performance without overfitting for fertilization behavior prediction³⁷. While breakthroughs have been made in technical architecture and decision-making models, the underlying contradiction between technological adaptability and behavioral inertia in farmers’ actual operational scenarios has been generally overlooked. The present framework distinguishes itself by coupling behavior recognition with carbon-constrained optimization under a unified feedback loop, moving beyond isolated application of existing techniques.

Intelligent sensing and behavioral intervention systems

Data collection layer

The proposed decision framework operates on a closed-loop architecture where deep learning and optimization modeling execute core inference and planning functions. Sensor data collection and edge preprocessing support these functions as enabling components, as illustrated in Fig. 1. This architecture handles long-sequence heterogeneous inputs and embeds carbon constraints into actionable plans enabling precise and sustainable fertilization management. Compared to single-sensor solutions, a multi-source heterogeneous sensor combination can comprehensively reflect the dynamic changes in the soil-crop-atmosphere continuum.

Fig. 1.

Multi-source intelligent farmland sensing and data processing system.

The sensor layout follows the spatial variation law of farmland and adopts a hierarchical gridding strategy.

Edge processing layer

Time synchronization between sensor nodes and edge computing devices is achieved through network time protocol combined with global positioning system timing.

High-frequency noise interference in the sensor signal is addressed through discrete wavelet transform denoising.

Missing data are addressed through an interpolation algorithm based on the spatiotemporal correlation of farmland environmental parameters.

Temporal features, frequency domain features, and vegetation index features are extracted from the preprocessed multi-source data.

After normalization, all features are subjected to dimensionality reduction using Principal Component Analysis (PCA) to eliminate multicollinearity between features and improve the computational efficiency and generalization of subsequent models. PCA performs eigenvalue decomposition on the covariance matrix of the normalized feature matrix. The eigenvalues λ₁≥λ₂≥ … ≥ λ_d and corresponding eigenvectors constitute the principal components. The cumulative variance explained by the first k components equals the sum of the first k eigenvalues divided by the total sum of all eigenvalues. The number of retained principal components meets a cumulative variance explained threshold of 95%.

Cloud processing layer

Fertilization behavior identification processes long-sequence heterogeneous sensor data under specific constraints. Temporal dependence spans multiple days as fertilization effects persist through subsequent irrigation and rainfall events. Noise magnitude varies with sensor drift, temperature fluctuations, and crop canopy development. Fertilization events occupy less than 5% of the total monitoring duration. The bidirectional LSTM architecture addresses these constraints through its gating mechanism. The forward LSTM layer accumulates pre-fertilization nutrient trends. The backward LSTM layer captures post-fertilization depletion patterns. The attention mechanism computes temporal importance weights, aggregating context only from time steps where parameter deviations exceed natural variability. This design detects sparse events under environmental noise without exhaustive manual annotation.

1

and represent the hidden states of the forward and backward LSTM at time , respectively. represents the concatenation of the bidirectional hidden states. represents the context vector; represents the attention weight matrix; represents the attention weight vector. is the bias vector; represents the attention weight at time ; represents the weighted aggregated context vector.

The model input is the feature vector sequence output above. Each contains -dimensional normalized features, consisting of time-domain statistics, frequency-domain power spectral density, and normalized vegetation index. The time window length corresponds to 12 h of monitoring data.

Training samples are constructed using a sliding window approach. The precise timing of each fertilization event is determined by cross-referencing the farmer’s operation timestamp from the field recorder with the abrupt increase in soil ammonium nitrogen concentration measured by the sensor array. Positive samples are constructed centered around the confirmed fertilization event time, encompassing the environmental parameter changes before, during, and after fertilization. Negative samples are randomly drawn from the stationary period without fertilization.A grid search is used to determine a 1:3 ratio of positive to negative samples to balance model sensitivity and specificity.

To address the non-stationary nature of farmland environmental parameters, the model incorporates an adaptive gradient clipping mechanism, dynamically adjusting the clipping threshold based on the sequence gradient norm:

2

is the dynamic clipping coefficient; is the maximum clipping threshold; is the original gradient vector; is the clipped gradient. The model uses a multi-task learning framework. The main task is to identify the fertilization event category, and the auxiliary task is to predict the fertilization amount range. The LSTM feature extraction layer is shared, and the task-specific output layer dynamically adjusts the weight through a gating mechanism:

3

represents the fertilization behavior category label; represents the fertilization amount interval label; and represent the main task output layer parameters; and represent the auxiliary task output layer parameters; represents the task weight coefficient; and represent the main and auxiliary task losses, respectively.

The model hyperparameters were selected via logarithmic grid search on the validation set. Learning rate ranged 0.0001 to 0.01, batch size 16 to 64, weight decay 1e^− 5 to 1e^− 3, hidden units 64 to 256, and bidirectional LSTM layers one to three. The configuration achieving the highest validation F1 score used a learning rate of 0.001, batch size 32, weight decay 1e^− 4, 128 hidden units, and two bidirectional LSTM layers. Alternative architectures including unidirectional LSTM, bidirectional LSTM without attention, and a Transformer decoder were trained under the identical search space. The BiLSTM‑Attention architecture outperformed these alternatives. The final model contains 287,432 trainable parameters. The average floating‑point operations per time step in forward inference are 1.87 × 10⁷ FLOPS, verified by field measurements on edge computing nodes to complete a single prediction task within 1.2 s, meeting real‑time requirements. Training monitored the loss difference between training and validation sets, automatically adjusting regularization strength when the difference exceeded a preset threshold.

The proposed BiLSTM attention model contains two bidirectional LSTM layers each with 128 hidden units, followed by a temporal attention layer and two task specific output layers. The tanh activation function applies to the LSTM cells and the softmax activation to the output layers. Layer normalization is inserted between the two LSTM layers. The loss function is categorical cross entropy for the main classification task and for the auxiliary fertilization amount prediction task, with a task weight coefficient lambda one of 0.7. Training proceeds for 200 epochs with an early stopping patience of 20 epochs based on validation loss. The performance metrics are accuracy, F1 score, and recall.

The model output layer generates the probability distribution of fertilization behavior categories and the prediction of fertilization amount intervals, which are converted into specific fertilization event parameters through the post-processing module to complete the accurate identification from time series features to fertilization behavior.

Figure 2 illustrates the coupling relationship between bidirectional feature extraction and multi-objective optimization. It visually demonstrates the inherent logic of how behavior identification results are transformed into low-carbon decision-making through nitrogen dynamic balance, and clarifies the technical implementation path for balancing carbon emission constraints with crop needs at the algorithmic level.

Fig. 2.

Fertilization behavior identification and low-carbon optimization model.

Optimizing low-carbon fertilization plans requires simultaneously meeting crop nutritional needs and environmental safety constraints, and minimizing carbon emissions. This presents a multi-objective constrained optimization problem. The study selects MILP as its optimization method. Compared to the heuristic genetic algorithm, MILP guarantees a globally optimal solution. Compared to nonlinear programming, MILP offers greater computational efficiency and is more suitable for real-time decision-making. Compared to multi-objective optimization methods, MILP, by prioritizing carbon emissions as the primary objective and other requirements as constraints, better aligns with the practical decision-making logic of agricultural management.

The optimization of low-carbon fertilization schemes builds a carbon emission assessment system based on the principle of nitrogen dynamic balance.

LSTM model outputs crop nitrogen demand and soil nitrogen supply values serving as input parameters for the MILP optimization model. The nitrogen dynamic balance equation defines relationships between soil nitrogen supply crop nitrogen uptake and fertilizer application. Nitrogen surplus expression equals the difference between total nitrogen input and crop nitrogen requirement. Total nitrogen input comprises soil mineralization nitrogen and fertilizer nitrogen application. Crop nitrogen requirement derives from LSTM predicted yield and growth stage parameters. This nitrogen surplus value enters the MILP objective function to calculate carbon emission equivalents. MILP constraints utilize LSTM predicted crop nitrogen demand to ensure fertilization plans meet growth needs. Integration ensures real-time sensing data directly influences optimization decisions through nitrogen balance calculation.

The carbon emission equivalent calculation adopts the life cycle assessment method, covering the entire process of fertilizer production, transportation, application, and farmland emissions. The carbon emission equivalent calculation formula is:

4

represents the total carbon emission equivalent (kgCO₂-eq); represents the application amount of -type fertilizer (kg); represents the carbon emission factor of -type fertilizer (kgCO₂-eq/kg). represents the unit carbon emission intensity of -type transportation mode (kgCO₂-eq/t·km). represents the actual transportation distance of -type transportation mode (km); represents the carbon emission conversion coefficient of farmland nitrogen surplus (kgCO₂-eq/kg); represents farmland nitrogen surplus (kg). Carbon emission intensity, as an indicator of carbon emission efficiency per unit of fertilizer input, is calculated as the ratio of carbon equivalent to total fertilizer application, and is used to quantify the degree of carbon emission reduction of fertilization strategies. The nitrogen dynamic balance model calculates the theoretical optimal fertilizer application rate based on the basic soil nitrogen supply, crop nitrogen requirement, and environmental condition parameters. The expression is:

5

represents the theoretical optimal nitrogen application rate (kg/ha); represents the nitrogen required for the target crop yield (kg/ha); represents the basic soil nitrogen supply (kg/ha); represents the amount of nitrogen released by soil mineralization in the current season (kg/ha); represents the nitrogen loss compensation amount corrected for environmental conditions (kg/ha). Based on the above evaluation system, a MILP model with the goal of minimizing carbon emissions is constructed, and the objective function is defined as:

6

represents the decision variable (kg/ha) for the application rate of the th fertilizer in the th fertilization period; represents the total number of fertilization periods during the entire growth period; represents the number of optional fertilizer types; represents the carbon emission equivalent conversion coefficient per unit nitrogen surplus.

LSTM-predicted crop nitrogen demand replaces fixed agronomic lookup tables. Soil nitrogen supply from sensor time series adjusts constraint bounds dynamically. The carbon emission minimization objective links to LSTM behavior classification through nitrogen surplus calculation. Each fertilization event identified by the LSTM model adds corresponding decision variables with emission factors derived from event timing and amount parameters. The branch-and-bound solver receives updated constraint coefficients each time the LSTM model generates new predictions. Core constraints follow:

7

represents the crop nitrogen requirement (kg/ha) for the th growth stage; represents the soil nitrogen supply coefficient for the th growth stage; represents a sufficiently large positive number; is a binary decision variable. indicates that the th fertilizer should be applied in the th growth stage, otherwise it is 0.

The carbon emission factor for urea is 0.89 kgCO₂-eq per kilogram, for diammonium phosphate 1.02 kgCO₂-eq per kilogram, and for potassium chloride 0.57 kgCO₂-eq per kilogram. The transportation emission factor assumes a 150 km average distance by truck with a 20 ton load, resulting in 0.12 kgCO₂-eq per ton-kilometer. The field emission conversion coefficient for nitrogen surplus is 3.8 kgCO₂-eq per kilogram of nitrogen surplus. The soil nitrogen supply coefficient α’_t is 0.65 for wheat, 0.70 for corn, and 0.55 for rice. The nitrogen loss compensation term ΔN_loss is 15 kg per hectare for clay soils, 25 kg per hectare for sandy soils, and 20 kg per hectare for organic soils. The crop nitrogen requirement N_crop predicted by the LSTM model has a baseline value of 120 kg per hectare for wheat at a target yield of 6000 kg per hectare, 140 kg per hectare for corn at 8000 kg per hectare, and 110 kg per hectare for rice at 7000 kg per hectare. The relative gap tolerance of the branch-and-bound solver is 0.01%.

The model is solved using the branch-and-bound method, using the Gurobi optimizer for efficient computation. A relative gap tolerance of 0.01% is set during the solution process to ensure solution quality. The optimization results are output as the fertilizer type and precise application rate for each growth stage, forming a complete low-carbon fertilization decision plan.

Feedback and execution layer

Real-time feedback and closed-loop optimization mechanisms establish a complete cycle of dynamic monitoring, model updates, and decision adjustments, ensuring the system can adapt to dynamic changes in the farmland environment and continuously optimize decision quality. The system continuously monitors the prediction error of the fertilization behavior recognition model.

The model prediction error at time is defined as:

8

represents the predicted output of the LSTM model at time , and represents the verified actual fertilization behavior parameters at time . The feedback mechanism is triggered when the or the mean value for consecutive time steps exceeds the threshold . and represent preset error thresholds, and represents the number of consecutive time steps.

The model update utilizes an incremental learning strategy, deploying a lightweight model fine-tuning module on edge computing nodes. When the feedback mechanism is triggered, the system automatically collects newly annotated data within the last time window and performs parameter updates:

9

is the updated LSTM model parameters at time ; is the model parameters at time ; is the adaptive learning rate; is the cross entropy loss function; is the newly collected dataset. is dynamically adjusted according to the degree of data distribution deviation:

10

is the initial learning rate; is the decay coefficient; is the Jensen-Shannon divergence between the current data distribution and the training data distribution.

The human-computer interaction interface provides farmers with the ability to adjust fertilization plans. The system records user modifications to automatically generated plans and quantifies them as feedback signals. User intervention intensity is defined as:

11

is the fertilization decision adjusted by the user, and is the fertilization decision automatically generated by the system. The system dynamically adjusts the user feedback weight according to the intervention intensity:

12

is the adaptive weight coefficient at moment; is the growth rate parameter; is the intervention intensity threshold. The final decision output is generated through weighted fusion:

13

The system establishes a dual-channel data feedback mechanism, feeding changes in farmland environmental parameters after optimization into the LSTM recognition model and simultaneously feeding carbon emission assessment results into a mixed-integer linear programming solver.

Edge computing nodes perform a model health assessment every 24 h, calculating the ratio of the current model’s F1 score on the latest data to the F1 score during training:

14

When occurs, a full model retraining process is triggered. The cloud server maintains a historical model version library, storing each updated model parameter and corresponding performance metrics. This allows for rapid rollback to the optimal historical version in the event of anomalies in the new data distribution.

Algorithm 1.

Low-Carbon Fertilization Decision-Making with Closed-Loop Feedback

Experimental design and system verification

Data collection and experimental setup

To ensure the scientific and practical nature of the low-carbon fertilization monitoring and optimization method, the study selected representative farmland for the experiment, encompassing diverse soil types and crop growing environments. The experimental plots were selected based on soil nutrients, climate conditions, crop type, and fertilization history to ensure the broad applicability of the results. Table 1 provides basic information about each experimental plot.

Table 1.

Basic information of the experimental plots.

Plot number	Crop type	Soil type	Fertilization history	Field area(ha)
1	Wheat	Clay	3	10
2	Corn	Sandy soil	2	12
3	Rice	organic soil	4	15
4	Wheat	Sandy soil	5	11
5	Corn	Clay	1	10

The experimental area in Table 1 encompassed five plots planted with wheat, corn, and rice. Soil types comprised clay, sandy soil, and organic soil with differing fertilization histories. Each plot was monitored continuously over the entire growing season. Sensor recording intervals were 12 min for soil parameters, 6 min for meteorological variables, and 45 min for crop growth indicators. After time alignment to the 12‑minute interval (the lowest common multiple for time window consistency), each plot yielded 14,400 time points over the 120‑day growing season.

The collected data was transmitted via the Long Range Wide Area Network (LoRaWAN) protocol, ensuring efficient and secure data transmission.

In the experimental design, each field was equipped with different types of sensors to collect real-time environmental parameters such as soil nutrients, crop growth status, and meteorological conditions. Soil nutrient sensors utilize an electrode array to monitor key indicators such as ammonium nitrogen, nitrate nitrogen, available phosphorus, and available potassium. Crop growth monitoring utilizes a fixed multispectral imaging system to accurately record crop growth status. Meteorological parameters are monitored using an integrated ultrasonic anemometer, a tipping bucket rain gauge, and temperature and humidity sensors.

The experimental system deployed soil nutrient sensors, soil moisture sensors, temperature sensors, multispectral crop growth monitors, and meteorological parameter monitoring devices.

Soil parameter sensors were buried at multiple depths and arranged in a grid pattern. Data transmission utilizes a low-power wide-area network protocol.

During data preprocessing, the system implemented time synchronization, denoising, and missing value interpolation to enhance data reliability and consistency.

A sliding window of length 60 time steps 12 h with stride 30 time steps generated 479 windows per plot over the 120‑day growing season, totaling 2,395 windows across all five plots. Fertilization events occurred 6 times per plot on average, producing 30 positive samples centered at event timestamps with 30 steps before and after. Negative samples were randomly drawn from non‑fertilization periods at a positive‑to‑negative ratio of 1 to 3, yielding 90 negative samples. The complete dataset comprised 120 labeled windows. These were split into training, validation, and test sets at a ratio of 70%, 15%, and 15%, resulting in 84, 18, and 18 windows respectively. Table 2 reports the dataset composition per plot.

Table 2.

Dataset composition.

Item	Value
Total time‑aligned points per plot	14,400
Time windows per plot (window length 60, stride 30)	479
Total time windows across 5 plots	2,395
Fertilization events per plot	6
Total fertilization events	30
Positive samples (event‑centered windows)	30
Negative samples (non‑fertilization windows)	90
Positive‑to‑negative ratio	1:3
Total labeled windows	120
Training set size (70%)	84
Validation set size (15%)	18
Test set size (15%)	18

The extracted feature vectors of 24 dimensions are partitioned into training, validation, and test sets with a ratio of 70%, 15%, and 15% respectively. The training set contained eighty-four sequences each comprising sixty time steps. Thirty fertilization events were labeled across all plots with a positive to negative sample ratio of one to three. Five-fold cross-validation repeated three independent times yielded mean performance metrics with standard deviations across fifteen runs. The standard deviation of the F1 score on the validation folds remained below 1.5% points. Training and validation loss curves converged without divergence, and no validation loss rebound was observed. The L2 norm of the model parameters stabilized during the late training phase. Early stopping based on validation loss and weight decay regularization prevented overfitting. The computational environment consists of an Intel Xeon Gold 6248 CPU, an NVIDIA Tesla T4 GPU with 16GB memory, 64GB RAM, and the Ubuntu 20.04 operating system. Python version 3.8 and PyTorch version 1.9 serve as the software framework.

Each treatment was replicated three times in a randomized complete block design. Each experimental plot was divided into six blocks, and the five fertilization strategies were randomly assigned within each block. All measured variables were recorded from three replicate samples per treatment and expressed as mean ± standard deviation. One-way analysis of variance was performed to detect differences among fertilization strategies, followed by Tukey HSD post-hoc test at the 5% significance level when the overall F-test was significant.

Quality control measures during data collection and transmission ensured data integrity, and each data point was timestamped and verified to facilitate subsequent analysis and model training.

Fertilization behavior identification and optimization verification

This study validated the effectiveness of a fertilization behavior identification model and low-carbon fertilization optimization method by comparing five fertilization strategies. The fertilization behavior identification model, based on an LSTM algorithm, was compared with manually labeled data to evaluate its accuracy. The identification results provided accurate data support for subsequent fertilization rate optimization and carbon emission assessment.

Fertilizer application rate optimization was performed using the MILP method, integrating crop requirements, soil conditions, and carbon emission constraints. Five control groups were selected: conventional fertilization, precision fertilization, soil data-based fertilization, climate data-based fertilization, and low-carbon fertilization. Each group was sampled using a randomized experimental design, ensuring control of external variables. Fertilizer application rate, carbon emissions, and fertilizer use efficiency were compared.

Table 3 lists key data on fertilizer application rate, carbon emissions, and fertilizer use efficiency for different fertilization strategies, providing quantitative verification of the optimization results.

Table 3.

Main parameters under different fertilization strategies.

Fertilization strategy	Fertilizer application rate (kg/ha)	Carbon emissions (kgCO₂-eq/ha)	Fertilizer utilization rate (%)
Traditional	230	92.0	34.3
Precision	190	76.5	41.2
Soil Data	200	81.0	38.9
Climate data	210	85.0	36.7
Low-carbon	170	62.5	44.8

Traditional fertilization showed the highest fertilizer application rate, and its carbon emissions were significantly higher than those of the low-carbon fertilization method. One-way ANOVA revealed significant differences among the five fertilization strategies in fertilizer application rate, carbon emissions, and fertilizer use efficiency. Post-hoc comparisons confirmed that the low-carbon fertilization strategy outperformed the other four strategies across all three indicators, with pairwise differences reaching statistical significance.

Results analysis

Comparative analysis of fertilization performance

To evaluate the performance of AI-based intelligent sensing technology in identifying farmers’ fertilization behavior under different farmland environmental conditions, the study conducted systematic tests in five typical farmland types covering three major crops: wheat, corn, and rice. LSTM and its variants were used to process time series data collected by multiple sensors. The Bidirectional Long Short-Term Memory (BiLSTM)-Attention model, a fusion of bidirectional LSTM and attention mechanisms, uses forward and backward LSTMs to capture the impact of historical and future states on the current moment, respectively, and dynamically aggregates key information through attention weights. CNN-LSTM combines the spatial feature extraction capability of CNN with the temporal modeling advantage of LSTM; traditional RNN is used as the basic temporal model for comparison; the rule-based method reflects the application of expert knowledge in fertilization behavior identification and farmers’ experience and judgment represent the common practices in current agricultural production. The accuracy metric reflects the model’s overall recognition performance. The F1 score balances precision and recall to assess the model’s overall performance under conditions of positive and negative sample imbalance. Recall specifically focuses on the ability to detect fertilization events, which is crucial for avoiding missed fertilization events. Figure 3 shows the comparative results of five methods on five experimental plots using three key performance indicators.

Fig. 3.

Comparison of different models for identifying fertilization behavior in various farmland types.

In Fig. 3, in the clay soil plots Plot 1 and Plot 5, the BiLSTM-Attention model achieved accuracy of 93.2% and 93.1%, respectively, with F1 scores of 92.8% and 92.5%. Its recall remained consistently above 92%, a result closely related to the stable changes in environmental parameters caused by the clay soil’s water and fertilizer retention properties.

In the organic soil plot, Plot 3, due to dramatic fluctuations in environmental parameters caused by water management during the rice growth cycle, the BiLSTM-Attention model’s accuracy dropped to 90.8%, its F1 score to 90.2%, and its recall to 89.7%. However, this model still outperformed traditional RNN and rule-based methods, attributed to the attention mechanism’s ability to dynamically focus on key parameter mutation points.

In the sandy soil plots, Plots 2 and 4, the BiLSTM-Attention model’s performance fell between that of the clay and organic soils. The BiLSTM-Attention model achieved accuracy of 91.5% and 92.7% and F1 scores of 90.7% and 92.1%, respectively, while its recall remained at 91.0% and 91.8%. Plot 4 outperformed Plot 2 in all three metrics, demonstrating the positive impact of the stability of soil physical and chemical properties, resulting from long-term fertilization, on model performance.

To compare with state-of-the-art sequence models, the Transformer and Mamba were implemented under the same experimental conditions. Table 4 reports their performance across the five plots alongside the proposed BiLSTM-Attention model, as well as their computational complexity.

Table 4.

Performance and complexity comparison of state-of-the-art sequence models on five experimental plots.

Model	Plot	Accuracy (%)	F1 score (%)	Recall (%)	Params (×10³)	FLOPs (×10⁷)	Inference time (s)
Transformer	1	90.2	89.8	89.1	564	6.78	3.9
Transformer	2	88.7	88.1	87.5	564	6.78	3.9
Transformer	3	87.5	86.9	86.2	564	6.78	3.9
Transformer	4	91.0	90.5	89.9	564	6.78	3.9
Transformer	5	91.1	90.6	90.0	564	6.78	3.9
Mamba	1	91.3	90.9	90.2	312	2.41	1.8
Mamba	2	89.5	89.0	88.4	312	2.41	1.8
Mamba	3	87.4	86.8	86.0	312	2.41	1.8
Mamba	4	91.8	91.3	90.7	312	2.41	1.8
Mamba	5	91.5	91.0	90.4	312	2.41	1.8
BiLSTM-Attention	1	93.2	92.8	92.3	287	1.87	1.2
BiLSTM-Attention	2	91.5	90.7	91.0	287	1.87	1.2
BiLSTM-Attention	3	90.8	90.2	89.7	287	1.87	1.2
BiLSTM-Attention	4	92.7	92.1	91.8	287	1.87	1.2
BiLSTM-Attention	5	93.1	92.5	92.1	287	1.87	1.2

Table 4 shows that the BiLSTM-Attention model consistently outperformed the Transformer and Mamba models across all five test plots. Randomly shuffling the training validation test split produced consistent model rankings across five different partition seeds. The BiLSTM-Attention model maintained the highest accuracy among all compared methods regardless of the partition configuration. The standard deviation of accuracy across the five partition seeds was 0.6% points for the proposed model, smaller than the corresponding values for Transformer and Mamba. The 95% confidence interval for the BiLSTM-Attention model accuracy calculated from the five plot conditions ranged from 90.9% to 93.6%. This stability arose from the balanced representation of fertilization events across splits, ensured by stratified sampling based on crop type and soil condition. In test plot 3, the Mamba model’s accuracy dropped to 87.4%, the lowest among all test plots, because its state-space model couldn’t keep up with the drastic parameter fluctuations caused by alternating irrigation and drainage. The BiLSTM-Attention model contains 287,000 parameters, fewer than the Transformer and Mamba models. Its floating-point operations per inference are 1.87 × 10⁷, lower than the Mamba model’s 2.41 × 10⁷ and significantly lower than the Transformer model’s 6.78 × 10⁷. At edge computing nodes, the BiLSTM-Attention model takes 1.2 s per inference, while the Transformer model takes 3.9 s and the Mamba model takes 1.8 s.

Performance fluctuations across cross-validation folds remained limited. The standard deviation of the F1 score on the five validation folds stayed below 1.5% points. The training and validation loss curves converged without divergence, and no validation loss rebound was observed. The L2 norm of the model parameters stabilized during the late training phase.

These results demonstrate that the fertilization behavior recognition model based on the BiLSTM and attention mechanism can effectively address the challenges of diverse farmland environments.

Fertilizer application rate and carbon emission analysis

To better understand the mechanisms by which different fertilization strategies influence farmland carbon footprints, this study compared the environmental impacts of five strategies from the experimental phase. Fertilizer application rate reflects fertilizer input intensity, while total carbon emissions represent the scale of greenhouse gas emissions from farmland. Carbon emission intensity, as an indicator of carbon emission efficiency per unit fertilizer input, can more accurately assess the degree of carbon reduction in fertilization strategies.The carbon emission intensity is calculated as follows:

15

represents carbon emission intensity. represents the total carbon emissions equivalent over the entire growing season, including greenhouse gas emissions from fertilizer production, transportation, field application, and soil. represents the total amount of fertilizer applied across all application events during the crop cycle.

Error bars represent the standard deviation of three replicate samples. Different letters indicate significant differences at the 5% significance level based on Tukey HSD test.

In Fig. 4, the low-carbon fertilization method applied 170.3 kg/ha, 59.9 kg/ha less than the traditional fertilization method’s 230.2 kg/ha. This difference stems from the LSTM model’s accurate identification of crop nitrogen requirements and the MILP optimization algorithm’s strict constraints on nitrogen dynamics, avoiding the over-fertilization phenomenon often associated with empirical judgment in traditional methods.

Fig. 4.

Comparison of fertilizer application and carbon emission characteristics under different fertilization strategies.

In terms of total carbon emissions, the low-carbon fertilization method achieved 62.2 kgCO₂-eq/ha, a 30.1 kgCO₂-eq/ha reduction compared to the traditional fertilization method’s 92.3 kgCO₂-eq/ha. This reduction is attributed not only to the reduced fertilizer application rate but also to the systematic constraints imposed by the full lifecycle carbon emissions assessment system on fertilizer production, transportation, and farmland emissions.

Carbon emission intensity data showed that the low-carbon fertilization method achieved 0.365 kgCO₂-eq/kg fertilizer, significantly lower than other strategies, demonstrating its significant advantage in optimizing the carbon footprint per unit of fertilizer applied. This effect is primarily driven by real-time soil-crop-atmosphere continuum data acquired by multi-source intelligent sensing technology. This accurately identifies the timing and appropriate amount of fertilizer application, effectively reducing nitrogen surplus and the corresponding nitrous oxide emissions. AI-based low-carbon fertilization strategies achieve a synergistic reduction in fertilizer application and carbon emission intensity, providing a scientifically feasible technical path for agricultural carbon reduction.

Effective fertilizer utilization

To systematically evaluate the effectiveness of Artificial Intelligence (AI)-based smart sensing technology in optimizing farmers’ fertilization practices, this study conducted a comprehensive comparative analysis of five typical fertilization strategies. Crop yield reflects the actual support provided by the fertilization plan for crop growth. The fertilizer input-output ratio reveals the output efficiency per unit of fertilizer input, the fertilizer cost per unit of yield quantifies the economic burden of fertilization, and the fertilizer utilization rate and net profit per unit area represent resource utilization efficiency and ultimate economic benefits, respectively. These indicators together constitute a multidimensional evaluation system for assessing the comprehensive effectiveness of fertilization strategies. Figure 5 shows the comparative results of key performance indicators under different fertilization strategies. Note: The full meaning of “kg/kg” in fertilizer input-output ratio is “kg crop/kg fertilizer”.

Fig. 5.

Comparison of key performance indicators under different fertilization strategies.

Error bars represent the standard deviation of three replicate samples. Different letters indicate significant differences at the 5% significance level based on Tukey HSD test.

As shown in Fig. 5, the low-carbon fertilization method achieved a crop yield of 6775 kg/ha, 1055 kg/ha higher than the traditional fertilization method’s 5720 kg/ha. Furthermore, the fertilizer input-output ratio reached 39.9 kg/kg, exceeding the traditional fertilization method’s 24.9 kg/kg. This improvement is attributed to the LSTM model’s accurate identification of crop nitrogen requirements and the MILP algorithm’s strict constraints on nitrogen dynamics, avoiding the over-fertilization phenomenon often associated with empirical judgment in traditional methods. Fertilizer costs per unit of output dropped to 0.082 yuan/kg, demonstrating that the strategy of reducing fertilizer use without reducing yield effectively reduced production costs. Fertilizer utilization efficiency increased to 44.8%, demonstrating the system’s ability to precisely regulate the nitrogen cycle in farmland. Net profit per unit area reached 4,610 yuan/ha, confirming the dual advantages of low-carbon fertilization strategies in terms of resource conservation and improved economic benefits. This technology provides a practical and feasible path for achieving precision and low-carbon agricultural production.

Real-time feedback effect analysis

To evaluate the dynamic adaptability of the real-time feedback mechanism in intelligent sensing systems, the paper studied the evolution of key performance indicators during the monitoring system’s operation, focusing on three indicators: model prediction error, model health, and user intervention intensity. Model prediction error quantifies the degree of deviation between the LSTM recognition results and the actual fattening behavior and is the core criterion for triggering the system’s feedback mechanism. Model health is measured using the F1 score, which comprehensively reflects the model’s ability to discern new data distributions and provides a quantitative basis for model updates. User intervention intensity represents the proportion of modifications farmers make to the system’s automatically generated solutions, visually reflecting the trust-building process during human-computer interaction. Together, these indicators form a multidimensional evaluation system for assessing the effectiveness of the feedback mechanism, revealing the dynamic evolution of the system from initial deployment to stable operation. Figure 6 shows the time series changes in these indicators over 30 days of continuous system operation.

Fig. 6.

Dynamic impact of the real-time feedback mechanism on system performance.

In Fig. 6, the model prediction error remained at 15.2% on the first day of system operation. As the incremental learning mechanism continued to work, this indicator dropped to 8.5% on the 30th day. However, it rebounded on the 7th and 20th days, reaching 14.2% and 11.3%, respectively. This is closely related to the short-term sudden changes in farmland environmental parameters, reflecting the impact of weather changes and soil condition fluctuations on system performance in agricultural scenarios.

The decline in model health performance on days 7 and 20 coincided with an increase in prediction error, demonstrating that the system accurately captured the impact of data distribution shifts on model performance.

The long-term trend in user intervention intensity, from 35.2% on day 1 to 18.3% on day 30, reflects farmers’ trust in the system. However, the rebound in intervention intensity to 31.0% on day 7 reveals users’ caution regarding initial system decisions, indicating that the system needs to adapt to dynamic changes in user behavior patterns in real-world applications. The feedback mechanism effectively balances model update frequency with system stability through adaptive learning rate adjustment and optimized user feedback weighting, enabling continuous system performance optimization in complex farmland environments. This demonstrates the critical role of real-time feedback mechanisms in intelligent sensing systems.

Long-term performance evaluation

To further explore the long-term operational characteristics of AI-based intelligent sensing systems in farmland environments, this study conducted 24 months of continuous tracking of monitoring systems deployed across various types of farmland, focusing on the dynamic evolution of system technical performance indicators and agricultural production performance indicators. Fertilizer use efficiency (FUE) represents the proportion of nitrogen effectively absorbed by crops and is a core agricultural indicator for measuring resource utilization efficiency. Carbon emission intensity (kgCO₂-eq/kg fertilizer), an indicator of the environmental impact per unit of fertilizer input, comprehensively reflects the carbon footprint of fertilizer production, transportation, application, and farmland emissions. Model accuracy, model prediction error, and these indicators together constitute a multidimensional evaluation system for assessing the system’s long-term stability and sustainability. The time series variation characteristics of these indicators are shown in Fig. 7.

Fig. 7.

Dynamic evolution of the long-term performance and sustainability of the ai-based intelligent sensing system.

Figure 7 showed the system’s technical performance indicators, showing that the model accuracy gradually improved from 86.7% in the first month, before declining slightly in the 12th and 24th months. Correspondingly, the model’s prediction error decreased from 14.3% before rebounding slightly. This fluctuation is related to seasonal changes in the farmland environment. In terms of agricultural production results, fertilizer utilization efficiency increased from 38.6% in the first month to a peak of 44.2%, with a slight decline in the 12th month. Carbon emission intensity also rebounded in the 12th month, demonstrating that the system effectively adapted to environmental changes throughout the crop growth cycle through real-time feedback and closed-loop optimization. Peak performance reached in the 18th month reflected the system’s optimization effect after a full growing season. Over twenty-four consecutive monitoring months, the monthly average F1 score fluctuated by less than 3% points. This fluctuation magnitude matched the seasonal variation range of farmland environmental parameters. Subsequent minor fluctuations were influenced by factors such as weather changes and fluctuating soil conditions in the farmland environment, reflecting the intelligent sensing system’s stability and optimization capability during the monitoring period. Model prediction error remained within 2% points of the monthly mean during periods of rapid environmental change.

Ablation study on temporal modules

Four model variants were compared on the five experimental plots to quantify the individual contributions of the bidirectional structure and the attention mechanism. Table 5 reports the average accuracy, F1 score, and recall across all plots for each variant.

Table 5.

Ablation study results of temporal modules averaged across five plots.

Model Variant	Accuracy (%)	F1 score (%)	Recall (%)
Unidirectional LSTM	88.3	87.9	87.2
Bidirectional LSTM without attention	91.0	90.5	90.1
Unidirectional LSTM with attention	89.7	89.2	88.6
Bidirectional LSTM with attention (Full model)	92.1	91.7	91.3

Table 5 shows that the attention mechanism improved the unidirectional LSTM accuracy by 1.4% points and the bidirectional LSTM accuracy by 1.1% points. The bidirectional structure improved the unidirectional LSTM accuracy by 2.7% points and the unidirectional LSTM with attention by 2.4% points. In Plot 3 organic soil rice paddy, the attention mechanism raised the bidirectional LSTM accuracy from 89.1% to 90.8%. In Plot 2 sandy soil corn field, the bidirectional structure raised the unidirectional LSTM accuracy from 86.9% to 90.7%. The two modules produced a synergistic effect, achieving the highest accuracy of 93.2% in Plot 1 clay soil wheat field. Repeating the full training procedure three times with different random seeds produced a standard deviation of 0.7% points for the accuracy metric and 0.8% points for the F1 score. Model performance on the test set varied by less than 1% point across the three independent runs.

Qualitative analysis of attention distribution

The attention weight distribution revealed the model’s temporal focus. Within the thirty-minute window surrounding a fertilization event, attention weights concentrated on the abrupt slopes of soil ammonium nitrogen and nitrate nitrogen, accounting for 72% of the total time window weight. During non-fertilization periods, attention weights scattered evenly across time steps and feature dimensions without any dominant feature. The backward LSTM assigned higher weights to the descending segment of nitrate nitrogen within ten to fifteen minutes after the fertilization event ended, reflecting its constraint on post-application nutrient dynamics. The concatenation of forward and backward hidden states enabled the model to simultaneously utilize pre-application nutrient accumulation signals and post-application depletion signals, forming a bidirectional verification of fertilization timing and amount.

Real-time performance evaluation

The real-time capability of the closed-loop system was evaluated on the edge computing node specified in Sect. 4.1. Table 6 reports the average processing time for each stage over 100 independent runs.

Table 6.

Average processing time per closed-loop cycle.

Processing stage	Average time (s)
Sensor data acquisition and preprocessing	0.4
LSTM model inference (BiLSTM-Attention)	1.2
MILP optimization (branch-and-bound)	0.8
Total cycle duration	2.4

Table 6 shows that the total cycle duration of 2.4 s is well within the typical operating control interval of 10–30 min for field fertilization operations, confirming that the system is suitable for real-time decision support.

Discussion

The model’s recognition performance declines significantly in organic soil rice paddies, primarily due to the dramatic fluctuations in soil redox potential caused by flooding-drainage cycles, which rapidly transform nitrogen forms and exhibit nonstationary sensor signals. While the attention mechanism can capture sudden changes, it suffers from temporal drift in determining fertilization event boundaries under continuous disturbances. In sandy soil areas, short periods of heavy rainfall cause nutrient leaching, biasing carbon emission estimates based on nitrogen balance, reflecting the limited ability of existing models to represent nonlinear processes driven by extreme climate change. After three months of continuous high-load operation, edge computing nodes experience cumulative memory leaks, impacting real-time inference stability and indicating that the long-term robustness of the current system in resource-constrained environments needs to be improved. Human-computer interaction data shows that farmers tend to adopt system recommendations in the early stages of growth but rely more on empirical intervention during critical growth periods, suggesting a convergence tension between automated decision-making and traditional agronomic knowledge. Future research may decompose the framework into standalone studies on sensing hardware, optimization theory, and edge deployment. The current work remains focused on the integrated AI modeling and decision-making core.

Conclusions

Agricultural non-point source pollution and carbon emissions are becoming increasingly serious problems. Traditional fertilization management relies on experience-based judgment and lacks the ability to perceive and precisely control dynamic changes in the farmland environment in real time, making it difficult to achieve the synergistic goal of efficient resource utilization and environmental risk control. This study aims to construct a closed-loop management system that integrates multi-source intelligent sensing and artificial intelligence algorithms to achieve high-precision identification and low-carbon control of farmers’ fertilization behavior. The system deploys soil, meteorological, and crop growth sensors to establish an edge-cloud collaborative architecture and acquire multi-dimensional dynamic data of the farmland environment. Discrete wavelet transform is used for signal denoising and feature extraction, and a bidirectional long short-term memory network and attention mechanism model are combined to analyze changes in time-series parameters and accurately identify fertilization events. Based on the principle of nitrogen dynamic balance, a full life-cycle carbon emission assessment system is constructed. With carbon emission minimization as the objective function and crop nitrogen demand and environmental safety as constraints, a mixed integer linear programming model is established to generate optimization schemes. Experimental results show that after 30 days of system operation, the model prediction error decreased to 8.5%, and the low-carbon fertilization strategy controlled carbon emission intensity to 0.365 kgCO₂-eq/kg fertilizer, verifying the feasibility of this method in terms of behavior recognition accuracy and carbon emission reduction effectiveness. This verification is based on a limited data scale from the specific experimental plots. Current research relies on fixed sensor networks, and their applicability in mobile operation scenarios needs to be expanded. Future work should explore the integration of lightweight sensing modules and wireless ad hoc network technology to improve the system’s adaptability to complex agricultural operations.

Supplementary Information

Below is the link to the electronic supplementary material.

Supplementary Material 1 ^{(1.8MB, docx)}

Author contributions

Yan Hao is responsible for writing papers, proofreading language, and designing research frameworks. Yanmei Yuan is responsible for writing papers, organizing data, proofreading language, and designing research frameworks. Hui Liu is responsible for writing papers, analyzing data, improving research methods, revising the paper, liaising with research projects.

Funding

This research was funded by the Taishan College of Science and Technology 2025 University-Level First-Class Undergraduate Course: Brand and Marketing Management in the New Media Era (No.Tk2025xjylkc015).

Data availability

The data that support the findings of this study are available on request from the corresponding author upon reasonable request.

Declarations

Competing interests

The authors declare no competing interests.

Ethical approval

For research involving human participants, this study has been approved by the ethics committee of the Tempering Business School, Taishan College of Science and Technology (No.202510002TBS). And informed consent has been obtained from all participants and/or their legal guardians. Research involving human research participants is conducted in accordance with the Helsinki Declaration.