Accurate mapping of field-scale soil moisture (SM) is a critical component of precision agriculture (PA). This study proposes a machine learning (ML) based framework that leverages multi-modal observations collected by an unmanned aircraft system (UAS) to generate high-resolution SM estimates over 210 × 110 m (2.31 ha) corn and cotton fields at Mississippi State University's research farm. The proposed ML approach uses in-situ SM measurements acquired from SM probes as ground-truth labels. The dataset spans over three growing seasons (2021-2023) and integrates global navigation satellite system reflectometry (GNSS-R) observables, multi-spectral vegetation indices, light detection and ranging (LiDAR)-derived canopy structure, and temporal features such as Day of Year to characterize dynamic soil-vegetation interactions. Three ML models, namely random forest (RF), extreme gradient boosting (XGBoost), and light gradient boosting machine (LGBM) were evaluated under challenging cross-validation strategies to assess spatial, temporal, and cross-crop generalization. In addition, multiple feature-selection techniques were applied to identify compact and physically meaningful predictor subsets. Results show that XGBoost achieves the strongest overall performance (e.g., root mean squared error (RMSE) = 0.0324 m3m-3 under probe-wise validation), with realistic SM maps that align with normalized difference vegetation index (NDVI) and canopy height maps. A modality-wise ablation study and comparison with a classical linear regression (LR) baseline were also conducted to evaluate the contribution of individual sensing modalities and the effectiveness of nonlinear multi-modal learning. A voting-based aggregation of selected features across all models and validation schemes highlights a small set of consistently informative variables related to temporal variability, UAS orientation, satellite geometry, and surface reflectivity behavior. The findings provide a detailed understanding of the opportunities and limitations of UAS-based GNSS-R SM prediction and point to promising directions for improving model generalization in future systems.

Accurate and unobtrusive measurement of upper-limb kinematics is critical for advancing wearable sensing technologies used in industrial ergonomics, human-machine interaction, and real-time biomechanics monitoring. This study evaluates the performance of two soft, flexible wearable sensors-BendLabs biaxial angular displacement sensors and StretchSense capacitive stretch sensors-for quantifying wrist and elbow motions during simulated dynamic industrial tasks. Wrist flexion-extension and radial-ulnar deviation were measured using BendLabs sensors mounted on the dorsal hand, while elbow flexion-extension was captured using StretchSense sensors positioned along the elbow joint. A multi-camera optical motion capture system served as the reference standard. Sensor data were preprocessed using baseline correction, smoothing, denoising, and normalized cross-correlation techniques to support temporal alignment with motion-capture recordings. Across all activities, the BendLabs sensors demonstrated moderate agreement with motion capture for wrist kinematics, with generally better performance for radial-ulnar deviation than for flexion-extension. StretchSense sensors demonstrated stronger agreement with motion capture for elbow flexion-extension, with performance that was generally consistent across task types. These findings support the feasibility of soft wearable sensors for capturing upper-limb kinematics during simulated occupational tasks and highlight their potential for integration into ergonomic assessment, occupational monitoring systems, and future industrial wearable platforms.

Accurate crop yield estimation is crucial for decision-making and planning in modern agriculture with increasing challenges with food security. Yield predictions provide farmers with insights into expected production, facilitating optimized resource allocation, improved agricultural management strategies, and enhanced profitability. This study investigates the application of machine learning (ML) techniques, including Feedforward Neural Networks (FNN), Long Short-Term Memory (LSTM), and Random Forest (RF) models, for predicting crop yields using multi-sensory time-series data that has been collected on two fields over a four-year timeframe. The focus is on corn (Zea mays) and cotton (Gossypium hirsutum) yield, two of the top critical crops in Mississippi region. A multi-sensory dataset was collected using multispectral cameras and LiDAR sensors mounted on unmanned aircraft systems (UAS), along with soil moisture and temperature data from volumetric probes and environmental data from a nearby weather station. Over four years, more than 30 features were extracted weekly from five major categories, with 235 ground truth yield records from plots in the field. The study outlines the methodology for feature selection and examines its impact on yield prediction accuracy. Using percentile root mean square error (RRMSE) and mean absolute percentage error (MAPE) as performance metrics, the study found that the proposed LSTM model produced lower field-wise errors (9 - 21 % MAPE) compared to other models and validation, indicating superior performance in predicting yields across selected weeks. The proposed ML-based approach, validated through year-based and field-wise cross-validation methods, demonstrates the effectiveness of using UAS-collected multi-sensor data for accurate yield estimation in corn and cotton.

Reliable perception of roadway signals is critical for autonomous vehicles operating in complex urban environments, particularly when traffic lights convey safety-critical instructions through flashing and arrow indications that extend beyond conventional red, yellow, and green states. However, most existing vision-based approaches focus primarily on static traffic-light recognition and lack robust mechanisms for interpreting temporal behaviors such as flashing signals. To address this limitation, this paper proposes a unified real-time perception framework, termed HybridSignalNet, for multi-class recognition of traffic lights, road signs, and lane-related roadway elements. The framework combines spatial detection with temporal state reasoning to interpret both steady and flashing signal patterns in video streams. Experimental evaluation demonstrates strong performance across multiple object classes, achieving an average detection F1-score of 91.3%, while traffic-light state classification reaches 96.7%, including reliable identification of flashing and arrow-based signals. The proposed system operates in real-time and provides an interpretable and deployable solution for intelligent transportation systems and autonomous driving applications, particularly at signalized intersections where temporal signal understanding is essential for safe decision-making.

Accurate crop yield estimation is essential for decision-making and planning in modern global agriculture, including in the USA, and is becoming increasingly critical for sustainable productivity to tackle climate change and food security challenges. Yield estimation provides farmers with insights into potential yields, enabling optimized resource allocation (water, fertilizer, pesticides, etc.), refined agricultural management strategies, and enhanced overall profitability. This research focused on feature selection for yield estimation of two of the top six crops in Mississippi and the US: corn and cotton. The study was conducted over a 2.31-hectare field encompassing both crop fields at the R. R. Foil Plant Science Research Center in Mississippi, US. The experiment took place over four consecutive years (2020 similar to 2023) during the crop growth period, from late April to mid-October each year. Weekly data were collected utilizing multispectral cameras and LiDAR sensors mounted on unmanned aircraft systems (UAS). Soil moisture (SM) map and soil temperature data were obtained from volumetric soil moisture probes, while all weather and environmental parameters were sourced from a nearby field weather station. The dataset for each plot consisted of over 30 features across five major categories, with ground truth yield data collected for 235 plots (30 cotton and 30 corn plots) over the 4-year period. The investigation of feature selection techniques included Pearson's correlation coefficient filtering, recursive feature elimination wrapping, and recursive groupwise wrapping, with the aim of identifying the most relevant features for yield estimation. For performance evaluation, crop yield prediction was conducted using a Long Short-Term Memory (LSTM) network, with root mean square error (RMSE) used as the performance metric. Among the methods, recursive groupwise wrapping produced comparatively lower RMSEs, indicating superior performance over selected weeks. Ultimately, the feature selection process enhances crop yield estimation accuracy compared to the usage of all available features.

Detecting, classifying, and tracking fish in underwater environments is a challenging task, even with the use of advanced machine learning models. Inaccurate annotations can significantly affect model performance, making it crucial to meticulously review and correct these annotations to enhance overall accuracy. The Fish Annotation Review Tool (FishART) is a customized labeling software designed to support active learning (AL) by allowing users to efficiently review and refine fish image annotations developed using AL algorithms. This tool displays detailed information such as image identifiers, bounding box coordinates, class labels, confidence scores, and track information, enabling users to carefully assess and improve the quality of their annotations. By leveraging the Southeast Area Monitoring and Assessment Program Dataset 2021 (SEAMAPD21) and the Gulf Fisheries Independent Survey of Habitat and Ecosystem Monitoring Dataset 2024 (GFISHERD24), FishART enables users to meticulously review and correct fish image annotations. Its key capabilities include the ability to merge multiple track information into a single object to ensure consistent tracking of fish across different frames. It also offers powerful filtering and sorting options, allowing users to filter annotations by class name, confidence score, or track information, helping them focus on specific aspects of the dataset. Furthermore, this tool facilitates annotation editing by enabling the addition, deletion, or modification of annotations, ensuring the dataset's accuracy. With its support for batch processing, the tool allows for efficient handling of large datasets, enabling users to apply bulk edits and filters with ease. Through our collaboration with the National Marine Fisheries Service (NMFS), a division in the National Oceanic and Atmospheric Administration (NOAA) and Northern Gulf Institute (NGI), FishART is playing a vital role in enhancing the accuracy and reliability of underwater fish detection and classification models, contributing to advancements in marine research and conservation efforts.

In this paper, we introduce an automatic method for classifying radar and communication waveforms using complex-valued convolutional neural networks (CV-CNNs) with parameterized, learnable sinc filters. Unlike traditional approaches that depend on computationally expensive preprocessing and transformations, our system processes complex-valued raw IQ RF data directly. By integrating learnable bandpass-like sinc filters, we efficiently extract time-frequency features while reducing the number of trainable parameters, thereby improving both interpretability and robustness in dynamic RF environments. We evaluated the proposed model on a synthetic dataset of 18 types of waveform modulation with signal-to-noise ratios (SNRs) ranging from -20 dB to 20 dB. The results show that our method outperforms conventional models in all key metrics, achieving an average accuracy of 76.47%. Moreover, it demonstrates improved noise robustness and effectively discriminates among a wide variety of modulation classes, even under challenging conditions.

Counter-drone systems can confuse drones, birds, and bird-like drones. Their radar micro-Doppler (mu-D) effects can be used to differentiate these targets from each other. Due to the lack of availability of mu-D signature datasets, mu-D return models from quadcopters with spinning rotors, birds with flapping wings, and bird-like drones with flapping wings were implemented for classification purposes. To showcase their use, these models were used to generate a synthetic dataset of spectrograms for binary classification (Drone vs. Bird) and ternary classification (Quadcopter vs. Bird-Like Drone vs. Bird). The classifiers examined were support vector machine (SVM), knearest neighbors (KNN), Naive Bayes, and a convolutional neural network (CNN). All binary classifiers produced at least 92.0% accuracy and at least 93.9% F1 score. All ternary classifiers produced at least 89.5% accuracy and 89.4% F1 score. The classification results indicate the mu-D models possess suitable fidelity for further research and development in drone vs. bird classification.

Accurate distance estimation in dynamic environments is essential for Advanced Driver Assistance Systems (ADAS) and autonomous driving. The limitations of individual sensors-such as cameras or radar-under challenging conditions necessitate a sensor fusion approach to enhance system robustness and perception accuracy. This research proposes a multi-sensor fusion method based on the Extended Kalman Filter (EKF), which integrates camera and radar measurements with a dynamic vehicle model to improve distance estimation. The approach is benchmarked against individual sensor performance to highlight its advantages. Synthetic datasets are generated using the Mississippi State University Autonomous Vehicular Simulator (MAVS), a high-fidelity simulation platform that enables controlled experimentation with various levels of sensor noise and missed detections. Multiple noise levels for both camera and radar sensors, along with missed detection rates, are introduced to emulate real-world sensor imperfections. The fusion algorithm, along with single-sensor baselines, is implemented in MATLAB. Performance is evaluated using both qualitative visualizations and quantitative metrics such as Root Mean Square Error (RMSE). Results demonstrate that the EKF-based fusion approach consistently yields lower RMSE values compared to single-sensor methods, indicating improved accuracy. Additionally, the study identifies optimal sensor and process noise parameters for EKF performance. Overall, this research presents a simulation-based framework that effectively mimics real-world conditions and shows that multi-sensor fusion with EKF significantly enhances distance estimation in dynamic driving scenarios, contributing to more reliable and safer autonomous vehicle operations.

Accurate traffic light detection and classification are fundamental for autonomous vehicle (AV) navigation and real-time traffic management in complex urban environments. Existing systems often fall short of reliably identifying and classifying traffic light states in real-time, including their flashing modes. This study introduces FlashLightNet, a novel end-to-end deep learning framework that integrates the nano version of You Only Look Once, version 10m (YOLOv10n) for traffic light detection, Residual Neural Networks 18 (ResNet-18) for feature extraction, and a Long Short-Term Memory (LSTM) network for temporal state classification. The proposed framework is designed to robustly detect and classify traffic light states, including conventional signals (red, green, and yellow) and flashing signals (flash red and flash yellow), under diverse and challenging conditions such as varying lighting, occlusions, and environmental noise. The framework has been trained and evaluated on a comprehensive custom dataset of traffic light scenarios organized into temporal sequences to capture spatiotemporal dynamics. The dataset has been prepared by taking videos of traffic lights at different intersections of Starkville, Mississippi, and Mississippi State University, consisting of red, green, yellow, flash red, and flash yellow. In addition, simulation-based video datasets with different flashing rates—2, 3, and 4 s—for traffic light states at several intersections were created using RoadRunner, further enhancing the diversity and robustness of the dataset. The YOLOv10n model achieved a mean average precision (mAP) of 99.2% in traffic light detection, while the ResNet-18 and LSTM combination classified traffic light states (red, green, yellow, flash red, and flash yellow) with an F1-score of 96%.

Background: This study was performed to validate the addition of capacitive-based pressure sensors to an existing smart sock developed by the research team. This study focused on evaluating the accuracy of soft robotic sensor (SRS) pressure data and its relationship with laboratory-grade Kistler force plates in collecting ground force reaction data. Methods: Nineteen participants performed walking trials while wearing the smart sock with and without shoes. Data was collected simultaneously with the sock and the force plates for each gait phase including foot-flat, heel-off, and midstance. The correlation between the smart sock and force plates was analyzed using Pearson’s correlation coefficient and R-squared values. Results: Overall, the strength of the relationship between the smart sock’s SRS data and the vertical ground reaction force (GRF) data from the force plates showed a strong correlation, with a Pearson’s correlation coefficient of 0.85 ± 0.1; 86% of the trials had a value higher than 0.75. The linear regression models also showed a strong correlation, with an R-squared value of 0.88 ± 0.12, which improved to 0.90 ± 0.07 when including a stretch-SRS for measuring ankle flexion. Conclusions: With these strong correlation results, there is potential for capacitive pressure sensors to be integrated into the proposed device and utilized in telehealth and sports performance applications.

Our SEAMAPD21 fish dataset consists of underwater images, where the class distribution is highly imbalanced, making fish identification particularly challenging. Additionally, tracking individual fish in this dataset presents further difficulties due to varying environmental conditions and fish behavior. YOLOv10 delivers enhanced detection accuracy, particularly for imbalanced datasets like ours, which contain underwater fish images. By integrating YOLOv10 for detection with ByteTrack for tracking, we significantly improve both the identification and tracking of fish species, leading to better overall performance in challenging underwater environments. The primary goal of this paper is to enhance the performance of the algorithm for fish tracking in this complex dataset, with a focus on accurately tracking and counting fish species. This improvement is crucial to effectively monitor marine biodiversity and contribute to conservation efforts. Experiments on the Southeast Area Monitoring and Assessment Program Dataset 2021 (SEAMAPD21) and the Gulf Fisheries Independent Survey of Habitat and Ecosystem Monitoring Dataset 2024 (GFISHERD24) demonstrate enhanced detection accuracy and robust multiobject tracking, even in challenging underwater conditions, highlighting the approach's suitability for ecological monitoring and conservation research.