Massive emissions of greenhouse gases (GHGs) have a negative impact on the development of sustainable agriculture. While techniques of imputation and forecasting facilitate the observation of GHG emissions with improved accuracy, there is a lack of an integrated model for both GHG emission data imputation and forecasting, particularly in few-shot learning scenarios. To address this issue, this paper proposes a pre-trained large language model dubbed Agri-LLM for GHG emission data imputation and forecasting in smart agriculture. Notably, this model develops an information fusion embedding layer that fuses missing patterns, temporal irregularities and incomplete time series into multi-level patched tokens. A global temporal similarity informed prompting module is further elaborated on to generate suitable prompts for target time series, based on similar temporal characteristics captured from other nodes. Finally, the model aligns the pre-trained knowledge language with multi-level integrated tokens directly without altering the large language model’s backbone. The experimental studies demonstrate that our model outperforms state-of-the-art baselines in both tasks of imputation and forecasting using full-sample training. Extensive experiments also confirm that the Agri-LLM exhibits superior performance in few-shot learning scenarios and the effectiveness of each proposed model component.

To address the conflict between the limited availability of spectrum resources and the swiftly growing number of frequency equipment in the Internet of Vehicles, this article starts from the solution of dynamic access of radio equipment and studies the problem of power spectrum density prediction of different channels, which is called spectrum prediction. We first propose the graph structure learning problem of electromagnetic spectrum data under the graph contrastive learning (GCL) framework from the causal perspective, and establishes a graph structure representation model between channel signal activity rules. Then, we establish the statistical distribution models of different channels at fine resolution based on the Gaussian mixture model. Then, the statistical model is embedded as prior knowledge using a graph convolutional network (GCN), and the channel association features are mapped to the association knowledge embedding using a graph structure encoder. This article proposes a novel spectrum prediction architecture based on GCL and multiscale feature fusion GCN (CL-MFGCN) to mine the time-frequency implicit knowledge of spectrum data, and the above knowledge embedding is integrated. This article visualizes the channel association relationship in the form of a graph structure. The experimental results indicate that the CL-MFGCN, reduces the average MAE by 14.6% (from 1.436 to 1.226) and the average MAPE by 12.9% (from 0.017 to 0.0148) compared to the second-best Pyraformer model, while maintaining a lower model complexity.

Radio map estimation (RME) is a crucial method for analyzing spectrum space utilization and network coverage, serving as an essential tool for the mobile communication. However, physical constraints, security, privacy, and other issues often render some areas inaccessible, resulting in extremely sparse and unevenly distributed measurement data. To address these challenges, we propose a multimodal feature fusion graph convolution network (MFFGCN). The model incorporates a dual-encoder architecture with an adaptive multi-feature fusion module to exploit environmental information and learn the shadowing effects of radio-signal propagation. We then convert the coarse estimation into regional feature patches and construct a graph over these patches. A graph neural network aggregates contextual information among them, thereby alleviating the impact of uneven spatial sampling. Extensive experiments on open datasets demonstrate that our method achieves state-of-the-art performance, effectively reducing the effects of uneven sampling.

With the increasingly serious shortage of spectrum resources, spectrum dynamic access based on spectrum prediction technology is widely recognized. Due to the high burstiness and complex intrinsic correlation of spectrum monitoring data, high-precision multi-channel spectrum prediction is challenging. This paper constructs spectrum monitoring data as a kind of graph structure data based on the correlation of spectrum itself, and designs a graph network model combining Graph convolution network(GCN) and Long-short term memory network(LSTM) for multi-channel spectrum prediction. This paper creatively introduces the method of graph network. And GCN is used instead of CNN to extract the correlation of channels, so as to improve the accuracy of multi-channel prediction. Experiments are conducted based on a real-world spectrum measurement dataset. The results show that the model proposed in this paper has better predictive performance compared with other methods.