Temporal Transfer Research Articles

Optimizing parameter estimation in hydrological models with convolutional neural network guided dynamically dimensioned search approach

Hydrological model calibration plays a crucial role in estimating optimal parameters for accurate simulation. Estimation of parameters is inevitable in hydrological modeling due to the challenge of directly measuring them, as most parameters are conceptual descriptions of physical processes. Modellers commonly employ optimization algorithms for calibrating hydrological models. However, these algorithms often pose computational challenges, especially when dealing with complex physics-based and distributed models. To address these challenges, our study introduces a novel approach called hydroCNN+DDS. By leveraging the strengths of Convolutional Neural Networks (CNN) and the Dynamically Dimensioned Search (DDS) algorithm, hydroCNN+DDS simplifies the model calibration process in complex physics-based models. This approach enables to capture the general patterns and relationships between discharge time series and parameters without compromising the underlying physics. We use hydroCNN+DDS to estimate parameters in the highly parameterized hydrological model, Structure for Unifying Multiple Modeling Alternatives (SUMMA) using hourly observed discharge. Notably, hydroCNN quickly generates sub-optimal parameters, serving as a good initial solution for DDS. This initialization aids DDS in converging faster towards an optimal solution. One of the notable advantages of the hydroCNN+DDS approach is its potential for spatial and temporal transferability. This feature proves valuable in dynamic systems and regions with limited historical data, expanding the applicability of the methodology. Furthermore, our proposed methodology is versatile and can be applied to any simple or complex models, accommodating any variables of interest. The best practices of good model calibration are followed in our approach.

Robust evaluation of big data-driven winter weather traffic models using six weigh-in-motion sites as testbeds in Alberta's highway network

ABSTRACT This study assesses the temporal transferability of winter traffic models using dummy variable regression at six Weigh-in-Motion (WIM) sites on Alberta's highway network. Models for two vehicle classes were developed using five years of traffic and weather data. To evaluate transferability, an additional year of data was collected, and two model structures—dummy variable and naive—were tested. The models’ estimation accuracy was measured using R² and five error metrics. Results indicate successful transferability of the models to a different year, with performance varying by road function and vehicle class. The study highlights that different model types might be needed for each vehicle class to ensure high temporal transferability. Additionally, the quality of the test data is crucial for obtaining reliable transfer results. This research addressed gaps in previous studies by collectively testing the temporal transferability of winter traffic models across an entire highway network.

A Multi-Scale Feature Fusion Deep Learning Network for the Extraction of Cropland Based on Landsat Data

In the face of global population growth and climate change, the protection and rational utilization of cropland are crucial for food security and ecological balance. However, the complex topography and unique ecological environment of the Qinghai-Tibet Plateau results in a lack of high-precision cropland monitoring data. Therefore, this paper constructs a high-quality cropland dataset for the YarlungZangbo-Lhasa-Nyangqv River region of the Qinghai-Tibet Plateau and proposes an MSC-ResUNet model for cropland extraction based on Landsat data. The dataset is annotated at the pixel level, comprising 61 Landsat 8 images in 2023. The MSC-ResUNet model innovatively combines multiscale features through residual connections and multiscale skip connections, effectively capturing features ranging from low-level spatial details to high-level semantic information and further enhances performance by incorporating depthwise separable convolutions as part of the feature fusion process. Experimental results indicate that MSC-ResUNet achieves superior accuracy compared to other models, with F1 scores of 0.826 and 0.856, and MCC values of 0.816 and 0.847, in regional robustness and temporal transferability tests, respectively. Performance analysis across different months and band combinations demonstrates that the model maintains high recognition accuracy during both growing and non-growing seasons, despite the study area’s complex landforms and diverse crops.

Space-Based Mapping of Pre- and Post-Hurricane Mangrove Canopy Heights Using Machine Learning with Multi-Sensor Observations

Coastal mangrove forests provide numerous ecosystem services, which can be disrupted by natural disturbances, mainly hurricanes. Canopy height (CH) is a key parameter for estimating carbon storage. Airborne Light Detection and Ranging (LiDAR) is widely viewed as the most accurate method for estimating CH but data are often limited in spatial coverage and are not readily available for rapid impact assessment after hurricane events. Hence, we evaluated the use of systematically acquired space-based Synthetic Aperture Radar (SAR) and optical observations with airborne LiDAR to predict CH across expansive mangrove areas in South Florida that were severely impacted by Category 3 Hurricane Irma in 2017. We used pre- and post-Irma LiDAR-derived canopy height models (CHMs) to train Random Forest regression models that used features of Sentinel-1 SAR time series, Landsat-8 optical, and classified mangrove maps. We evaluated (1) spatial transfer learning to predict regional CH for both time periods and (2) temporal transfer learning coupled with species-specific error correction models to predict post-Irma CH using models trained by pre-Irma data. Model performance of SAR and optical data differed with time period and across height classes. For spatial transfer, SAR data models achieved higher accuracy than optical models for post-Irma, while the opposite was the case for the pre-Irma period. For temporal transfer, SAR models were more accurate for tall trees (>10 m) but optical models were more accurate for short trees. By fusing data of both sensors, spatial and temporal transfer learning achieved the root mean square errors (RMSEs) of 1.9 m and 1.7 m, respectively, for absolute CH. Predicted CH losses were comparable with LiDAR-derived reference values across height and species classes. Spatial and temporal transfer learning techniques applied to readily available spaceborne satellite data can enable conservation managers to assess the impacts of disturbances on regional coastal ecosystems efficiently and within a practical timeframe after a disturbance event.

A systematic quality-integrated diagnostic method for complex product assembly using multi-task spatial–temporal transfer learning

The assembly process is generally considered one of the primary factors influencing the quality of complex products. Currently, most existing quality-integrated diagnostic methods for products tend to deteriorate over different processes and degrade over time. To address this issue, this paper introduces a systematic quality-integrated diagnostic method for complex product assembly processes. First, the influence factors and error sources in the complex assembly process are analyzed using the 5M1E and FAHP methodologies. Next, similarity-based multi-task clustering and dismantling using RGMM is applied to divide the assembly tasks. Finally, MMD-MSE is employed to develop a quality prediction model for the complex assembly process using a spatial–temporal transfer learning approach. Experiments were conducted on an array antenna assembly task, comparing the proposed method with conventional methods. The results show that the accuracy and PrUP of the proposed model are 97.6% and 95.2%, respectively, for the quality-integrated diagnostic of complex assembly processes, with a fluctuation in accuracy of less than 6%. The diagnostic results effectively meet expert evaluations and provide a stable, reliable, and practical solution for addressing quality fluctuations in complex production assembly processes.

A new methodology for establishing an SOC content prediction model that is spatiotemporally transferable at multidecadal and intercontinental scales

Quantifying and tracking the soil organic carbon (SOC) content is a key step toward long-term terrestrial ecosystem monitoring. Over the past decade, numerous models have been proposed and have achieved promising results for predicting SOC content. However, many of these studies are confined to specific temporal or spatial contexts, neglecting model transferability. Temporal transferability refers to a model’s ability to be applied across different periods, while spatial transferability relates to its applicability across diverse geographic locations for prediction. Therefore, developing a new methodology to establish a prediction model with high spatiotemporal transferability for SOC content is critically important. In this study, two large intercontinental study areas were selected, and measured topsoil (0–20 cm) sample data, 27,059 cloudless Landsat 5/8 images, digital elevation models, and climate data were acquired for 3 periods. Based on these data, monthly average climate data, monthly average data reflecting soil properties, and topography data were calculated as original input (OI) variables. We established an innovative multivariate deep learning model with high spatiotemporal transferability, combining the advantages of attention mechanism, graph neural network, and long short-term memory network model (A-GNN-LSTM). Additionally, the spatiotemporal transferability of A-GNN-LSTM and commonly used prediction models were compared. Finally, the abilities of the OI variables and the OI variables processed by feature engineering (FEI) for different SOC prediction models were explored. The results show that 1) the A-GNN-LSTM that used OI as the input variable was the optimal prediction model (RMSE = 4.86 g kg−1, R2 = 0.81, RPIQ = 2.46, and MAE = 3.78 g kg−1) with the highest spatiotemporal transferability. 2) Compared to the temporal transferability of the GNN, the A-GNN-LSTM demonstrates superior temporal transferability (ΔR2T = −0.10 vs. −0.07). Furthermore, compared to the spatial transferability of LSTM, the A-GNN-LSTM shows enhanced spatial transferability (ΔR2S = −0.16 vs. −0.09). These findings strongly suggest that the fusion of geospatial context and temporally dependent information, extracted through the integration of GNN and LSTM models, effectively enhances the spatiotemporal transferability of the models. 3) By introducing the attention mechanism, the weights of different input variables could be calculated, increasing the physical interpretability of the deep learning model. The largest weight was assigned to climate data (39.55 %), and the smallest weight was assigned to vegetation (19.96 %). 4) Among the commonly used prediction models, the deep learning model had higher prediction accuracy (RMSE = 6.64 g kg−1, R2 = 0.64, RPIQ = 1.78, and MAE = 4.78 g kg−1) and spatial transferability (ΔRMSES = 1.43 g kg−1, ΔR2S = −0.13, ΔRPIQS = −0.50, and ΔMAES = 1.09 g kg−1), and the linear model had the higher temporal transferability (ΔRMSET = 1.46 g kg−1, ΔR2T = −0.14, ΔRPIQT = −0.45, and ΔMAET = 1.29 g kg−1). 5) The deep learning models necessitated the OI, whereas the linear and traditional machine learning models necessitated FEI to achieve higher prediction accuracy. This study presents an important step forward in integrating multiple deep learning models to build a highly spatiotemporal transferability SOC prediction model.

Using Duolingo in Teaching and Learning Vocabulary: A Systematic Review

This systematic review examines the extant literature on the efficacy of the Duolingo application for vocabulary development in instructed second language (L2) contexts. Drawing on several empirical studies, the review synthesises findings related to (a) the comparative effectiveness of Duolingo versus traditional pedagogies, (b) the impact of Duolingo on learner motivation and engagement, and (c) practical implications for classroom integration. While the majority of evidence attests to measurable gains in receptive vocabulary knowledge, the findings must be interpreted with caution due to methodological limitations, including small sample sizes, short intervention periods, and construct underrepresentation. The motivational appeal of Duolingo's gamified interface emerges as a prominent theme, corroborated by self-report data. However, the limited research employing psychometrically robust motivational measures precludes definitive conclusions about the temporal stability and transferability of this motivational effect. Theoretically, the provision of individualised feedback aligns with interactionist theories, yet the restricted opportunities for the negotiation of meaning raise doubts about the depth of lexical engagement. Pedagogically, the reviewed studies advocate the careful integration of Duolingo as a supplementary tool within formal instructional contexts. However, optimal implementation strategies remain unspecified, necessitating further classroom-based research employing mixed-methods designs. Significant gaps are identified, including the lack of investigations into the differential impacts of Duolingo's features, vocabulary transfer to authentic communication, and the development of self-regulated learning strategies.

On the temporal transfer function in STEM imaging from finite detector response time

Faster scanning in scanning transmission electron microscopy has long been desired for the ability to better control dose, minimise effects of environmental distortions, and to capture the dynamics of in-situ experiments. Advances in scan controllers and scan deflection systems have enabled scanning with pixel dwell times on the order of tens of nanoseconds. At these speeds, the finite response time of the electron detector must be considered as the signal from one electron detection event can contribute to multiple pixels, blurring the features within the image. Here we introduce a temporal transfer function (TTF) to describe and model the effects of detector response time on imaging, as well as a framework for incorporating these effects into simulation.

Climate covariate selection influences MaxEnt model predictions and predictive accuracy under current and future climates

The performance and transferability of species distribution models (SDMs) depends on a number of ecological, biological, and methodological factors. There is a growing body of literature that explores how the choice of climate covariate combinations and model parameters can affect predictive performance, but relatively few that delve into covariate reduction methods and the optimisation of model parameters, and the resulting spatial and temporal transferability of those models. The present work used the citrus pest, Diaphorina citri Kuwayama (Hemiptera: Psyllidae), to illustrate how MaxEnt models trained on the insect’s native range in Asia varied in their predictions of climatic suitability across the introduced range when eight different covariate reduction methods were applied during model building. Additionally, it showed how model sensitivity varied across these different covariate combinations using three sets of independently validated occurrence points in the invaded range of the psyllid. Climatically suitable areas for D. citri differed by as much as two-fold between the best and worst-performing models in selected areas. Great care should be taken in the selection of the highest-performing predictor combinations and model parameter settings for SDMs, particularly in the case of invasive species where the assumption of environmental equilibrium is likely violated in the introduced range. Understanding how the predictive ability of SDMs can be influenced by the methodological choices made during the model building phase is vital to ensuring that ecological and invasion management programmes do not over- or underestimate climatic suitability and subsequent invasion risk.

Temporal Interference Wireless Power Transfer Strategy With Passive Flexible Electrode for Vagus Nerve Stimulation

Temporal Interference Wireless Power Transfer Strategy With Passive Flexible Electrode for Vagus Nerve Stimulation

A temporal information transfer network approach considering federal funds rate for an interpretable asset fluctuation prediction framework

This study explores the complex interdependencies among five major financial assets—S&P 500, Bitcoin, Crude Oil, Gold, and USD/EUR—from April 2015 to September 2022, emphasizing the importance of understanding global financial dynamics for robust financial management. We quantitatively analyze daily causal relationships using conditional transfer entropy, a method that surpasses traditional correlation analyses. By incorporating the effective federal funds rate into our models, we enhance predictive accuracy and account for monetary policy impacts, ensuring our findings are relevant to current economic conditions. Our results reveal significant causal networks, providing key insights into asset interdependencies that support advanced hedging strategies and effective diversification. This research improves prediction models through the innovative use of network-based features and offers practical strategies for managing multinational financial assets, with relevance across various economic scenarios.

A new transferable deep learning approach for crop mapping

ABSTRACT Accurate crop mapping provides important information for government decision-making and agricultural management. Recent advancements in deep learning have significantly enhanced the capabilities of remote sensing-based crop mapping. In this study, we developed a new deep learning approach, DSH, which comprises three modules, DeepLabV3+ (D), channel self-attention (S), and histogram matching (H). The DeepLabV3+ module learns the spatial distribution of crops in the spatial and spectral dimensions. The channel self-attention module was used to enhance the weights of the important features. The histogram-matching module addresses the domain gap between the training and testing images and enhances the transferability of the DSH. In addition, the input data of the DSH are monthly synthesized Sentinel-2 data, thus relaxing the data collection requirements. The performance of the DSH was evaluated and benchmarked against HRNetV2, DeepLabV3+, and random forest (RF) at eight sites in the U.S. China, and France with different farming structures, climate conditions, and landscape complexities. Temporal transfer revealed that the classification performance of DSH improved as the growth season progressed, peaked in the peak growth season (August), and then declined at each study site. The DSH model trained during the peak growth season demonstrated superior temporal transferability. Spatial transfer experiments demonstrated that transferring DSH to sites with similar planting structures achieved higher precision than full spatial transfers. Spatiotemporal transfer experiments conducted at eight sites over three years (2020–2022) demonstrated an average overall accuracy (OA) of 88.1% for DSH, highlighting the importance of similar planting structures for effective spatiotemporal transfer. DSH outperformed HRNetV2, DeepLabV3+, and RF in OA, producer’s accuracy, and user’s accuracy at all eight study sites. Notably, DSH exhibited comparable performance during both the peak and full growth seasons, with an average OA of 93.9%. The classification accuracy of test images after histogram matching remained stable from 2020 to 2022, whereas the accuracy of raw test images fluctuated annually. This indicates that histogram matching is effective in improving the transferability of DSH. Embedding the channel self-attention module into the position of the high-level features had a more significant impact on improving DSH’s performance than other configurations. Visualizing features from the channel self-attention and high-level layers revealed that integrating the channel self-attention module with DeepLabV3+‘s high-level features significantly improved crop mapping by enhancing relevant semantic information. Overall, DSH exhibits robust spatiotemporal generalizability and requires minimal input data, making it suitable for large-scale crop mapping.

Deep learning-based quality prediction for multi-stage sequential hot rolling processes in heavy rail manufacturing

Traditional quality control practices in heavy rail manufacturing primarily rely on experiential knowledge and numerical simulations. However, these methods come with significant drawbacks, including high costs, safety concerns, and limited accuracy due to the inadequate consideration of various influencing factors. Despite the accumulation of extensive data by some heavy rail manufacturers in recent years, achieving data-driven quality modeling remains a formidable challenge, mainly due to the complexity of the multi-stage sequential production process. This study presents the Temporal-Spatial Mapping approach to transform isochronous sampling data into spatially aligned data. Additionally, it introduces a multi-objective prediction model, Temporal-Fusion-LSTM (TFL), designed for evaluating the quality of heavy rails. The proposed model presents the Double-Layer LSTM with Time-Steps Fusion (DLL) architecture. The first layer of the DLL architecture targets the temporal transfer regularity within a single stage, while the second layer addresses the transition dynamics across multiple rolling stages in the heavy rail rolling process. Following this, the Local Process Correlation Matrix (LPCM) based on the Maximal Information Coefficient (MIC) is employed to empower the model to comprehend the intricate interactions among parameters in each stage of the rolling process. Finally, the Multi-Gate based Shared-Bottom architecture is employed to accentuate manufacturing processes highly correlated with the target quality indicators during multi-objective prediction. Real production data from a heavy rail factory is utilized to validate the effectiveness of the method. The experimental results unequivocally demonstrate that the proposed method accurately predicts heavy rail quality and facilitates cross-stage prediction, thus establishing the groundwork for data-driven quality diagnosis and pre-control in heavy rail production.

Cropping intensity mapping in Sentinel-2 and Landsat-8/9 remote sensing data using temporal transfer of a stacked ensemble machine learning model within google earth engine

This article aimed to map Cropping Intensity Patterns (CIPs) in the southwest region of Iran using Google Earth Engine and monthly composites of Sentinel-2 and Landsat-8/9 data. To detect CIPs with high inter- and intra-class variability of crops, a heterogeneous Stack ensemble of machine learning model was developed. The model incorporated the Minimum Distance (MD) approach as a meta-classifier, combining multiple base models, including Support Vector Machines (SVM), Random Forest (RF), Classification and Regression Trees (CART), and Gradient Boosted Trees (GBT). In 2021, the Stack model was trained and evaluated using Ground Truth (GT) samples from the same year, achieving an Overall Accuracy (OA) of 94.24%. This performance surpassed the base models by about 4% in OA and was also reflected in the detection accuracies, including User’s Accuracy (UA), Producer’s Accuracy (PA), and F1-score, of the target classes. Subsequently, the trained stack model was temporally transferred to generate CIP maps for other years. The model achieved high OAs of 91.82% and 90.97% based on GT samples from 2020 and 2022, respectively. Finally, the time series of CIP maps (2019-2023) were utilized by the Cellular Automata-Markov model to forecast the map for 2024.

Temporal Transferability of Tree Species Classification in Temperate Forests with Sentinel-2 Time Series

Detailed information on forest tree species is crucial to inform management and policy and support environmental and ecological research. Sentinel-2 imagery is useful for obtaining spatially explicit and frequent information on forest tree species due to its suitable spatial, spectral, and temporal resolutions. However, classification workflows often do not generalise well to time periods that are not seen by the model during the calibration phase. This study investigates the temporal transferability of dominant tree species classification. To this end, the Random Forest, Support Vector Machine, and Multilayer Perceptron algorithms were used to classify five tree species in Flanders (Belgium) with regularly spaced Sentinel-2 time series from 2018 to 2022. Cross-year single-year input scenarios were compared with same-year single-year input scenarios to quantify the temporal transferability of the five evaluated years. This resulted in a decrease in overall accuracy between 2.30 and 14.92 percentage points depending on the algorithm and evaluated year. Moreover, our results indicate that the cross-year classification performance could be improved by using multi-year training data, reducing the drop in overall accuracy. In some cases, gains in overall accuracy were even observed. This study highlights the importance of including interannual spectral variability during the training stage of tree species classification models to improve their ability to generalise in time.

Transformer-based novel framework for remaining useful life prediction of lubricant in operational rolling bearings

Accurate prediction of the remaining useful life (RUL) of lubricants in rolling bearings is crucial for maintaining efficient operation of rotating machinery and ensuring timely lubricant replacement. We propose a comprehensive framework that integrates the temporal variation transfer function (TVTF), harmonic-sideband Matrix, and the harmonic frequency transformer (HarFT), a transformer-based model. This approach effectively utilizes vibration characteristics to enhance the accuracy of lubricant degradation prediction in rolling bearings. Validation with rolling bearings experiencing lubrication failure confirms that our framework significantly outperforms alternative methods in RUL prediction. The proposed framework excels in extracting and analyzing harmonic components from vibration responses, enabling detection of minute status variations due to lubricant degradation. By applying explainable artificial intelligence (XAI), it is possible to ascertain the rationale behind the RUL predicted by the HarFT model, facilitating evidence-based decisions. Our research provides a novel strategy for lubricant RUL assessment in rolling bearings, thereby improving reliability and maintenance efficiency in industrial applications.

Optimizing airflow in spiral blast freezers

Spiral blast freezing is a common unit operation used in food processing facilities for rapidly freezing a variety of foodstuffs. The purpose of a blast freezer is to generate high velocity, low temperature air flow over food products being conveyed within refrigerated enclosures to accomplish the freezing process. However, air flow patterns observed within field operating blast freezers are often suboptimal, resulting in diminished system performance. This paper applies a Monte Carlo simulation technique to a food product freezing simulation in order to identify velocity profiles that optimize the freezing process. A one-dimensional food product model is used to evaluate the interplay between the time variation in the magnitude of the air velocity over food products conveyed through the freezing system and the resulting dwell time needed to achieve a target product core temperature at the blast freezer exit. Temporal heat transfer coefficients derived from field measurements made in a newly installed spiral blast freezer serve as a basis to calibrate the one-dimensional product model.The results of the Monte Carlo analysis show freezing system performance is improved when high and stable air velocities over the product are achieved early in the freezing process dwell time. Air flow patterns within a freezing system that result in high air velocity later in the freezing process dwell time are suboptimal. Field-measured data on a newly installed spiral blast freezer showed this suboptimal air flow pattern and the use of baffling within the spiral enables improved airflow leading to an estimated 10 % increase in production throughput.

Decoding Urban Mobility: Application of Natural Language Processing and Machine Learning to Activity Pattern Recognition, Prediction, and Temporal Transferability Examination

Decoding Urban Mobility: Application of Natural Language Processing and Machine Learning to Activity Pattern Recognition, Prediction, and Temporal Transferability Examination

Temporal spatial transfer characteristics and cross-scale fusion behavior mechanism of interface temperature in induction welding of plain carbon fiber polyphenylene sulfide

The intricate multiphysics coupling conditions and nonlinear characteristics inherent in carbon fiber composite induction welding render the temperature transfer properties of the interface and the fusion connection mechanism across different scales ambiguous. In this study, a transient 3D finite element model was developed to simulate the induction welding process of carbon fiber reinforced polyphenylene sulfide (CF/PPS) composites, utilizing an electromagnetic-thermal multiphysics coupling mechanism. This investigation revealed that the electromagnetic effect leads to the accumulation of magnetic flux and spatial constraints, thereby triggering an eddy current concentration effect at the interface, which serves as an additional influencing factor in induction welding's edge effect. Simultaneously, the temperature distribution at the induction welding interface exhibits a gradient characteristic over both time and space: fluctuations in the current heat source result in a pronounced gradient distribution pattern of 225 %-680 % in welding interface temperature differences, accompanied by a 43.29 % reduction in shear strength. Notably, this study delves into the intricate origin of the induction welding heat source, elucidating that the generation of induced current in the magnetic field by carbon fiber overlap contributes to volumetric heat induction. This facilitates the welding process to some extent, despite the significantly lower heat contribution compared to thermally conductive wire meshes. In conclusion, to gain deeper insights into the impact of heat input on the fusion behavior of the bonding interface, SEM analysis was conducted on the experimental samples to examine the fusion morphology of the induction welding interface. Additionally, the method of EDS elemental analysis has also been used to investigate the generation and fusion mechanisms of welding defects. The amalgamation of numerical simulation technology and multi-scale material characterization serves as a methodological guide for researching the welding mechanisms of composite materials and even extends to investigations within the domain of metal connections. Furthermore, it paves the way for exploring a multidimensional research horizon in the realm of material joining and molding.

A temporal statistical assessment of the effectiveness of bicyclist safety helmets in mitigating injury severities in vehicle/bicyclist crashes

This study estimates mixed logit models taking into account heterogeneity in means and partially constrained parameters in order to explore possible shifts within parameters over time to study factors influencing bicyclist injury severity outcomes. Separate statistical models are estimated for two bicyclist helmet-wearing scenarios (helmet and non-helmet) using a comprehensive dataset from Florida covering a three-year period to assess COVID-19 effects from the 1st of January 2019 to the 31st of December 2021. This research evaluates several factors influencing helmeted and non-helmeted bicyclist injury severity, encompassing the attributes of drivers and cyclists, the environment and weather, the features of the roads and their temporal aspects, and the different types of vehicles. The performed analysis further enhances model robustness by assessing the temporal stability and transferability across different contexts through likelihood ratio tests, alongside an in-depth examination of the temporal consistency of explanatory variables via marginal effects analysis, confirming significant variations between non-helmeted and helmeted bicyclist models and revealing temporal shifts in factors affecting injury severity during the study period. Findings from the model estimations identify several significant variables with consistent parameter estimates across years. Stop signs, cycling with traffic, and dark, unlit conditions increase severe injury risk in non-helmet models, while the stop sign indicator consistently reduces severe injury risk in helmet models. Statistically significant random parameters are identified across different years and helmet-wearing scenarios, including the male driver indicator, which exhibits varying effects on injury severity. Out-of-sample prediction analysis suggests helmets reduce severe injury probability but may increase minor injuries and decrease no-injury accidents, indicating potential risk compensation behavior among helmeted bicyclists. Although helmets offer protection against severe injuries for bicyclists, it is crucial to adopt a comprehensive safety approach, particularly given the evolving demographics of bicyclists amid the COVID-19 outbreak. This entails considering factors like bicyclist and driver behavior, environmental conditions, and infrastructure enhancements. Policymakers, road safety professionals, and advocacy groups should collaborate to develop holistic strategies to address the determinants of bicycle crash severity outcomes and enhance safety measures for bicyclists across diverse road environments.