Traffic flow prediction plays a crucial role in the development of smart cities. However, existing studies face challenges in effectively capturing spatio-temporal contexts, handling hierarchical temporal features, and understanding spatial heterogeneity. To better manage the spatio-temporal correlations inherent in traffic flow, we present a novel model called Dynamic Spatio-Temporal Memory-Augmented Network (DSTMAN). Firstly, we design three spatial–temporal embeddings to capture dynamic spatial–temporal contexts and encode the unique characteristics of time units and spatial states. Secondly, these three spatial–temporal components are integrated to form a multi-scale spatial–temporal block, which effectively extracts hierarchical spatial–temporal dependencies. Finally, we introduce a meta-memory node bank to construct an adaptive neighborhood graph, implicitly representing spatial relationships and enhancing the learning of spatial heterogeneity through a secondary memory mechanism. Evaluation on four public datasets, including METR-LA and PEMS-BAY, demonstrates that the proposed model outperforms benchmark models such as MTGNN, DCRNN, and AGCRN. On the METR-LA dataset, our model reduces the MAE by 4% compared to MTGNN, 6.9% compared to DCRNN, and 5.8% compared to AGCRN, confirming its efficacy in traffic flow prediction.

The accurate detection and identification of pests and diseases on cucumber leaves is a prerequisite for scientifically controlling such issues. To address the limited detection accuracy of existing models in complex and diverse natural backgrounds, this study proposes an improved deep learning network model based on YOLOv8, named SEDCN-YOLOv8. First, the deformable convolution network DCNv2 (Deformable Convolution Network version 2) is introduced, replacing the original C2f module with an improved C2f_DCNv2 module in the backbone feature extraction network’s final C2f block. This enhances the model’s ability to recognize multi-scale, deformable leaf shapes and disease characteristics. Second, a Separated and Enhancement Attention Module (SEAM) is integrated to construct an improved detection head, Detect_SEAM, which strengthens the learning of critical features in pest and disease channels. This module also captures the relationship between occluded and non-occluded leaves, thereby improving the recognition of diseased leaves that are partially obscured. Finally, the original CIOU loss function of YOLOv8 is replaced with the Focaler-SIOU loss function. The experimental results demonstrate that the SEDCN-YOLOv8 network achieves a mean average precision (mAP) of 75.1% for mAP50 and 53.1% for mAP50-95 on a cucumber pest and disease dataset, representing improvements of 1.8 and 1.5 percentage points, respectively, over the original YOLOv8 model. The new model exhibits superior detection accuracy and generalization capabilities, with a model size of 6 MB and a detection speed of 400 frames per second, fully meeting the requirements for industrial deployment and real-time detection. Therefore, the SEDCN-YOLOv8 network model demonstrates broad applicability and can be effectively used in large-scale real-world scenarios for cucumber leaf pest and disease detection.

Traffic forecasting plays a crucial role in smart city research, significantly influencing urban traffic management and facilitating precise route planning. In recent years, numerous deep learning techniques have gained prominence for their effectiveness in spatiotemporal modeling. However, achieving accurate predictions remains a formidable challenge due to the inherent dynamics and heterogeneity present in spatiotemporal data. To address this challenge, we introduce a novel Heterogeneity-aware Graph Convolutional Recurrent Network (HGCRN) specifically designed for traffic prediction. The HGCRN framework begins by constructing a set of spatial-temporal meta-embeddings that effectively encode temporal context and node heterogeneity. It then utilizes time-varying traffic features to generate dynamic graph learners, integrating dynamic graph convolutional networks with gated recurrent units to proficiently model spatial and temporal dependencies. In the decoding phase, we incorporate temporal embeddings of the target predictions to enhance the forecasting process, thereby improving the model’s capacity to capture trend variations. Extensive experiments conducted on four widely recognized benchmark datasets demonstrate the superiority of the proposed model while highlighting its robust interpretability.

Short-term traffic flow forecasting becomes especially difficult when incomplete observations, within-window frequency variation, and state-dependent sensor relations occur together. Missing readings can affect both node features and the spatial dependencies inferred from them, yet these issues are commonly modeled separately. We therefore propose RIFT-STGNN, a Robust Interleaved Frequency–Trend Spatio-Temporal Graph Neural Network for multi-step traffic flow forecasting. RIFT-STGNN follows a coordinated information flow: observation status is retained during temporal–frequency encoding, the frequency representation supports both node features and graph construction, and four graph sources are fused before each Graph-GRU update. Trend-aware temporal attention then produces direct multi-step forecasts. Experiments on PEMS03, PEMS04, PEMS07, and PEMS08 show competitive numerical performance relative to selected literature-reported baselines under the 12-step setting. On PEMS04 and PEMS08, the three-run mean MAE values are 17.73 and 13.14, respectively. These values are numerically 4.63% and 9.00% lower than the corresponding literature-reported sAMDGCN values. Component ablations, state-dependent graph analysis, and controlled missing-rate experiments support the roles of dynamic graph learning, within-window frequency encoding, and mask-aware input handling under the evaluated settings.

Accurate traffic-flow forecasting remains challenged by abrupt and irregular states even after dominant periodic patterns are captured. Existing predictors model the resulting difficult errors implicitly through their parameters and cannot explicitly reuse specific historical errors at inference. We find that multi-horizon base-prediction errors recur conditionally across similar de-cycled node-local states. Based on this observation, we propose De-Cycled Residual Memory with Utility-Gated Retrieval (DeCyReM). It constructs queries from de-cycled residuals, spatial context, and temporal information, retrieves historical base errors from a node-local difficult-and-diverse memory, and uses a utility gate to control their contribution as forecast corrections. Across PeMS03, PeMS04, PeMS07, and PeMS08, DeCyReM records the numerically lowest reported value in seven of the twelve dataset–metric combinations and ranks within the top three in eleven. Ablation and mechanism analyses show that retrieval improves error alignment and mainly benefits difficult samples; similarity alone cannot determine correction utility, whereas gate values track helpful retrievals. Case and robustness analyses further characterize successful correction, safe attenuation, and mismatch-induced failure, identifying reliable key–value correspondence as essential. Overall, DeCyReM converts reusable historical error experience into selective forecast corrections, providing an error-centric approach to forecasting abrupt and irregular traffic states.

Accurate multi-step traffic forecasting requires models to respond to genuine traffic-state transitions while resisting deviations caused by missing or noisy observations. Existing deviation-aware methods characterize departures from recurrent patterns, but deviation magnitude alone does not indicate whether the resulting correction will improve a node-wise forecast. Guided by the insight that correction strength should depend on forecast gain rather than deviation magnitude, we propose Spatiotemporal Traffic Forecasting with Forecast-Gain-Guided Deviation–Invariance Routing (ST-FGDIR). ST-FGDIR constructs an anchor-conditioned stable endpoint and a deviation-corrected candidate through weight-shared current–anchor graph encoding and multi-view deviation evidence. Signed candidate-error differences provide soft forecast-gain targets for a node-wise continuous router, while selective risk extrapolation and a false-activation loss control stable-endpoint risk across perturbation environments and suppress perturbation-induced increases in residual injection. Experiments on METR-LA, PEMS-BAY, PEMS04, and PEMS08 show competitive forecasting performance relative to published baseline results. The inference gates exhibited Spearman correlations of 0.49–0.53 with realized correction gain, and the complete model yielded an average relative MAE increase of 9.38% across 12 controlled dataset–perturbation conditions. These results demonstrate that forecast-gain-guided routing coordinates stable periodic prediction with adaptive deviation correction under both clean and perturbed inputs.