The remainder of this study is structured as follows: Sect. 1 introduces the study background, outlines key questions, identifies gaps in the existing literature, and presents the study’s structure. Section 2 reviews related work on agricultural transformation, deep learning models, and optimization algorithms. Section 3 details the methodology, including the development of the hybrid deep learning model and the implementation of the SMA. Section 4 presents the experimental design, dataset construction, and model performance evaluation. Section 5 discusses the implications of the findings, summarizes the study’s main contributions, and outlines potential applications and directions for future research.
Since the Chinese government announced the “dual carbon” goals in 2020 — targeting peak carbon emissions by 2030 and carbon neutrality by 2060 — various industries have encountered both pressures and opportunities related to transformation and upgrading. The agricultural sector, as a traditionally high-energy-consuming industry, faces distinct challenges. Its transformation is essential not only for environmental protection but also for ensuring food security and promoting sustainable development. Consequently, evaluating the outcomes of agricultural transformation has become a key area of interest for both researchers and policymakers.
In recent years, rapid advances in artificial intelligence (AI), particularly in deep learning, have significantly impacted the agricultural sector. These technologies have been widely applied in areas such as pest and disease identification and crop growth monitoring. Such innovations have improved production efficiency while accelerating the transition toward low-carbon and intelligent agricultural practices. However, research focused on evaluating agricultural industry transformation using deep learning — especially within the framework of the dual carbon strategy — remains limited. For instance, Srivastava et al. (2022) developed a CNN-based model employing one-dimensional convolution operations to capture time-dependent relationships between environmental variables for winter wheat yield prediction. Their results showed the CNN outperformed other models in predictive accuracy. Similarly, Saleem et al. (2022) applied various machine learning and deep learning algorithms in agricultural robotics, achieving notable performance gains. Their findings revealed that deep learning-based models outperformed traditional methods, with RCNN reaching an 82.51% success rate in plant disease and pest detection. ResNet-18 and FCN also demonstrated superior performance in tasks such as crop-weed differentiation and land cover classification. Mohamed et al. (2023) provided a comprehensive review of trends in agricultural sustainability and technological innovation, highlighting emerging strategies such as precision farming, climate-smart techniques, and the integration of deep learning. They proposed innovative approaches to optimize crop yields, reduce ecological impact, and enhance global food security. These include quantum intelligence, meta-learning, deep reinforcement learning, curriculum learning, intelligent nanotechnology, blockchain technology, and CRISPR gene editing. Cob-Parro et al. (2024) introduced an open-source infrastructure specifically optimized for developing and deploying AI-based models in agricultural environments. In another study, Mujeeb and Javaid (2023) combined Spearman Correlation Analysis (SCA) and an improved shallow denoising autoencoder (ISDAE) with deep neural networks (DNNs), further enhanced by particle swarm optimization (PSO), to forecast carbon emissions in power systems. Despite these technological advancements, several challenges persist in practical applications — particularly issues related to data acquisition and the limited generalizability of many deep learning models. Addressing these challenges remains critical to fully realizing the potential of AI in supporting agricultural transformation under the dual carbon framework.
Swarm intelligence metaheuristic algorithms — such as PSO, Genetic Algorithms (GA), and Ant Colony Optimization (ACO) — have attracted considerable attention for their effectiveness in optimizing deep learning and machine learning models. These algorithms simulate collective behaviors observed in nature, enabling them to efficiently search high-dimensional solution spaces while avoiding the local optima that often hinder traditional optimization methods. Their success has been particularly notable in tasks such as feature selection, model tuning, and parameter optimization for complex problems. Swarm intelligence algorithms have demonstrated unique advantages across various application domains. For example, they have been widely applied in optimizing machine learning models for tasks including cloud computing instance price prediction, early diagnosis of type 2 diabetes, and software defect detection. Salb et al. (2024) used an improved PSO algorithm to fine-tune multi-head deep learning models, significantly enhancing both prediction accuracy and model stability. Similarly, Navazi et al. (2023) applied swarm intelligence algorithms to improve the accuracy and early-warning capabilities of diabetes diagnosis models. In software defect detection, Petrovic et al. (2024) demonstrated that ACO effectively extracted key features from large-scale defect datasets, leading to improved predictive accuracy. Zivkovic et al. (2023) employed swarm intelligence to optimize the Extreme Gradient Boosting (XGBoost) model and incorporated Shapley values to enhance feature interpretability, further boosting model performance. In the realm of social media data analysis, Dobrojevic et al. (2024) used Genetic Algorithms (GAs) to optimize the XGBoost model, achieving substantial improvements in the detection of cyberbullying, gender discrimination, and harassment. These studies collectively underscore the remarkable effectiveness of swarm intelligence metaheuristic algorithms in fine-tuning complex machine learning models, particularly in scenarios involving high-dimensional, nonlinear, and intricate datasets.
In the field of climate and environmental prediction, the application of deep learning models has become an emerging research focus. Guo et al. (2023) proposed a hybrid approach combining artificial neural networks (ANNs) with deep learning models to forecast monthly average and extreme temperatures in Zhengzhou. By integrating the linear modeling capabilities of ANNs with the nonlinear feature extraction strength of models such as CNN, their study significantly improved prediction accuracy — particularly for extreme high-temperature events. However, the analysis was limited to a single city and did not involve multi-regional comparisons or the integration of dynamic spatiotemporal interactions. Expanding on this work, Guo et al. (2024a) explored the joint use of Deep Convolutional Neural Networks (DCNNs) and LSTM networks for monthly climate prediction. DCNNs were used to extract spatial features from climate data, while LSTM networks captured temporal dependencies. This hybrid model achieved high accuracy in predicting various climate indicators, including precipitation and wind speed. The experiments demonstrated lower prediction errors compared to traditional models under complex climate conditions. Nonetheless, the study did not incorporate optimization algorithms for hyperparameter tuning, which may limit the model’s adaptability in more dynamic scenarios.
For comparative performance analysis, Guo et al. (2024b) systematically evaluated DCNNs, LSTMs, GRUs, and Transformer models. The findings showed that LSTM was most effective for modeling long-term dependencies, while Transformer excelled in capturing global correlations across climate variables. However, the absence of optimization strategies — such as metaheuristic algorithms — and inadequate handling of missing values during data preprocessing may have impacted the robustness of the results. In the context of air pollution prediction, He and Guo (2024) assessed the performance of LSTM, Recurrent Neural Networks (RNNs), and Transformer models in forecasting monthly PM2.5 concentrations in Dezhou. LSTM achieved the lowest prediction errors due to its ability to model long-term dependencies in time series data, while the Transformer model was hindered by high computational complexity, limiting its practical application. The study also noted that nonlinear relationships between meteorological variables (e.g., wind speed and humidity) and pollutant concentrations remain a key modeling challenge. Further research by Guo, He, and Wang (2025) focused on ozone concentration prediction in Liaocheng, comparing the performance of LSTM and ANNs. LSTM models showed better performance for daily ozone forecasts. However, their heavy reliance on real-time meteorological inputs reduced overall effectiveness. Additionally, the lack of multi-source data — such as satellite remote sensing — to enrich model inputs limited the ability to accurately predict sudden pollution events.
Recent advances in agricultural multisource data fusion and intelligent optimization have significantly expanded the application scope of deep learning and system modeling. For instance, Rokhva et al. (2024) developed an image recognition framework for the food industry using a pretrained MobileNetV2 model. Their lightweight network design enabled high-accuracy, real-time classification, and the proposed model compression strategies offer valuable insights for deploying agricultural remote sensing data in edge computing environments. Building on this work, Rokhva and Teimourpour (2025) introduced a food classification model that integrated EfficientNetB7, the Convolutional Block Attention Module (CBAM), and transfer learning. Their use of attention mechanisms and data augmentation techniques presents effective solutions for aligning features in agricultural multimodal data, such as crop disease imagery and meteorological heat maps. Additionally, Abbasi et al. (2025) applied circular economy principles to design a sustainable, closed-loop supply chain network. Their robust optimization framework addresses uncertainty and offers interdisciplinary support for dynamic resource allocation and policy resilience assessment in the context of agricultural low-carbon transitions. While these studies span different domains, their contributions to lightweight model design, attention-based feature enhancement, and sustainable system optimization provide a strong interdisciplinary foundation for developing a multisource, data-driven evaluation model for agricultural transformation in this study.
Artificial intelligence and metaheuristic algorithms are increasingly applied across agriculture and environmental sciences, reflecting a growing multidisciplinary trend. For example, Gaber and Singla (2025) used Random Forest Regression (RFR) to predict groundwater distribution, with multivariate feature selection providing a data-driven framework for agricultural irrigation planning. Similarly, El-Kenawy et al. (2024) introduced the Greylag Goose Optimization (GGO) algorithm, inspired by migratory bird behavior, offering novel swarm intelligence strategies for agricultural resource allocation. In climate adaptation research, Alzakari et al. (2024) developed an enhanced CNN-LSTM model for early detection of potato diseases. Their multi-scale feature fusion effectively integrates local pathological features from leaf images with temporal disease progression patterns. Elshabrawy (2025) reviewed waste management technologies for sustainable energy production, emphasizing the potential of hybrid metaheuristic algorithms — such as combinations of ACO and GAs — for agricultural waste recycling. Regarding crop yield prediction, El-Kenawy et al. (2024) compared machine learning and deep learning approaches for forecasting potato yields. They found that integrating gradient boosting with Grey Wolf Optimization (GWO) substantially reduced errors caused by climate variability. Despite these advances, most studies have yet to address spatiotemporal asynchrony in multisource agricultural data — including remote sensing, meteorological, and policy datasets — and often lack validation of cross-regional generalization. Furthermore, existing research on agricultural low-carbon transformation evaluation tends to focus on isolated technological pathways, rarely incorporating the dynamic effects of policy interventions. In contrast, this study proposes a hybrid CNN-LSTM model optimized by the SMA. SMA dynamically tunes model parameters, encoding regional climate adaptability as hyperparameter constraints. This approach enables the joint optimization of carbon footprint assessment and policy resilience. By overcoming the limitations of traditional methods in fusing cross-scale data and modeling nonlinear relationships, it provides an extensible evaluation framework to support the sustainable transformation of agricultural systems.
Table 1 summarizes the core methods, contributions, and limitations of representative studies on agricultural transformation evaluation. Existing research generally faces three key challenges. First, most models analyzing spatiotemporal features rely on fixed architectures and lack dynamic optimization. Second, metaheuristic algorithms are usually applied in isolation and are not fully integrated with deep learning models. Third, multisource data integration remains limited, hindering the ability to capture nonlinear interactions among climate, policy, and crop variables. In contrast, this study harnesses the global search capability of the SMA to coordinate CNN-LSTM spatiotemporal modeling. It also introduces region-specific feature weighting coefficients to enable a multidimensional, dynamic evaluation of agricultural transformation effects.
Most existing studies have concentrated on case-specific applications involving individual technologies or crop types, often lacking systematic and comprehensive methodologies. Additionally, many have overlooked the impact of regional variability on agricultural transformation. To address these limitations, this study proposes an integrated evaluation framework that combines deep learning models — specifically CNN and LSTM networks — with the SMA, a swarm intelligence optimization technique. Unlike traditional methods, SMA excels at global search and is particularly effective for handling the spatiotemporal complexity of agricultural data, thereby enhancing the accuracy and stability of production forecasts. This study introduces three key innovations over conventional CNN-LSTM optimization frameworks: First, traditional models often rely on fixed hyperparameters and lack the flexibility to adapt to the dynamic nature of heterogeneous agricultural data. The proposed model overcomes this limitation by using SMA to automatically adjust critical hyperparameters — such as network architecture and learning rate — during training. This dynamic tuning enables the model to better adapt to complex agricultural features, significantly improving predictive accuracy and convergence speed. Second, unlike previous frameworks that depend on a single type of spatiotemporal data, the proposed model adopts a multi-channel input design. It simultaneously incorporates satellite remote sensing imagery, meteorological time series, and field management data. During the feature fusion stage, region-specific adaptive weighting factors are introduced to enhance the model’s sensitivity to geographic variability. This design improves generalizability across different agricultural regions — a common weakness in existing models. Third, validation experiments conducted across multiple agricultural subregions show that the proposed model maintains high predictive accuracy while exhibiting substantially lower error variance compared to traditional CNN-LSTM approaches. These results highlight its superior robustness and resilience to environmental and data-related disturbances. In summary, the CNN-LSTM-SMA model presents significant advancements in structural optimization, multisource data integration, and regional generalization. It offers a practical and adaptable framework for evaluating agricultural transformation and holds strong potential for supporting data-driven decision-making in low-carbon agricultural policy.
Key innovations include: (1) A deep learning framework that integrates diverse data sources — including satellite remote sensing, meteorological data, and field management information — to enable dynamic monitoring and evaluation of agricultural transformation. (2) Regionally adaptive model adjustment strategies that enhance predictive accuracy and generalization across varying geographic and climatic conditions.
