Graduate Student Research
Graduate Student Research
Graduate students from Mississippi State University's academic departments are using data science to address important human challenges in many domains.
Comprehensive Wind Speed Forecasting-Based Analysis of Stacked Stateful and Stateless Models
Wind speed is a powerful source of renewable energy, which can be used as an alternative to the non-renewable resources for production of electricity. Based on stats from the US Energy Information Administration, Non-renewables accounts for up to 81% of the total electricity generation in the US in 2021, with natural gas being the most common source (38%). Renewable sources are clean, infinite and do not impact the environment negatively during production of electrical energy. However, while eliciting electrical energy from renewable resources viz. solar irradiance, wind speed, hydro should require special planning failing which may result in huge loss of labor and money for setting up the system. We use about four deep recurrent neural networks viz. Stacked Stateless LSTM, Stacked Stateless GRU, Stacked Stateful LSTM and Statcked Stateful GRU which will be used to predict wind speed on a short-term basis for the airport sites beside two campuses of Mississippi State University. The paper does a comprehensive analysis of the performance of the models used describing their architectures and how efficiently they elicit the results with the help of RMSE values. We estimate which location out of the two considered locations is favourable for the installation of a wind turbine. A detailed description of the time and space complexities of the above models has also been discussed.
Electroencephalography and Biomechanics of the Basketball Free Throw
According to various studies, compared with novice athletes, experts exhibit superior integration of perceptual (e.g., quiet eye – eye fixation on target prior to motion), cognitive (e.g., sense of distance), and motor skills (e.g., movement control). This superior ability has been associated with the focused and efficient organization of task-related neural networks. Specifically, skilled individuals demonstrate a spatially localized or relatively lower response in brain activity, characterized as ‘neural efficiency’, when performing within their domain of expertise. Besides, previous works also suggested that elite basketball players can predict successful free throws more rapidly and accurately based on cues from body kinematics, which might be due to a long training of specific motor skills (e.g., basketball free throw) and associated with focused excitability of the motor cortex during the reaction, movement planning and execution phases. Thus, utilization of electroencephalography (EEG) and motion capture system (MoCap) can provide a deeper understanding of the relationship between neurophysiological activity and human biomechanics as well as their effects on the success rate of the motor skill. Additionally, there were no previous studies that combined both EEG and MoCap systems to analyze the performance of the subjects during specific motor skill execution. The reason might be to avoid motion artifacts occurring during movement initiation and execution. However, our protocol will ensure the subjects remain stable and avoid making sudden moves, which helps to fully capture the EEG signals during the movement planning step prior to movement initiation. As a result, by utilizing the recorded EEG signals and MoCap data of 16 participants, each performing 50 basketball free throws, this study aims to analyze the athlete’s biomechanical and neurological parameters during movement execution to evaluate its effect on shot accuracy. The study can be a practical approach in analyzing the sources that lead to better elite athletes’ performance in various sport-related tasks. Moreover, the acquired data can contribute to a deeper understanding of the connection between the mental and physical states of elite athletes during successful outcomes, thus, providing vital information for the overall improvement of athletic performance and guidance for sport-specific training needs.
Student Scientist: Phong Phan
Advisor: Dr. David Vandenheever
Department: Department of Agricultural and Biological Engineering
An Optimization Model for Assessing the Impact of Carbon Offset Programs in Timberland Assets
In the last decades, the carbon market has become a promising alternative to the current poor timber prices practices in the US South. However, it is not very clear what is the impact of a carbon offset program on long-term forest management. We gathered information about the yield curve, cost, and prices to build a harvest schedule model based on a profit maximization in a multi-stand timberland asset. After estimating the optimal rotation ages, we imposed a harvest constraint in the first rotation to simulate one-year carbon offset contracts. Under three different site productivity, our results indicate that imposing a one-year carbon constraint will change the silvicultural interventions in future rotations, thereby reducing the financial returns. Landowners must be compensated from $169 to $235 per acre for a one-year carbon offset contract. Our findings can support public and private decision makers to estimate the impact of carbon-offset projects on timber supply and forest conservation. In addition, landowners can better evaluate their economic and ecological trade-offs when applying for such programs.
Student Scientists: Nasir Qadir, Kalani Perera, Okikiola Michael Alegbeleye
Advisor: Dr. Bruno K. da Silva
Department: Department of Forestry, College of Forest Resources
Modeling and Forecasting Pine Sawtimber Stumpage Prices in the South-Central US
Timber return is composed of three main drivers, forest growth and yield, timber price, and land price. Forest growth-yield is stressed by literature in a wide range of models for different ecosystems, while land prices are related to inflation. Not having accurate predictions of timber prices over time for different regional timber markets affect optimal deterministic rotation. Consequently, affecting the periodic dividends and the decision-making process regarding strategies of timber production for each region to improve timber return. To overcome this problem, we evaluated time series methods applied to real pine sawtimber stumpage quarterly prices in timber regions across the southern US. Pine sawtimber stumpage prices from four different states in South-Central US (AR, LA, MS, and TX) were used from the 1977-2018 period. To provide a framework of how the time series methods behave with price prediction, we used the moving average as the benchmark and compared it with the univariate autoregressive integrated moving average (ARIMA). A data stationarity test was performed, and first-order differencing was used for predictions. Hence, it was possible to compare the performance of each modeling approach for in-sample and out-sample price forecasts (2016-2018). Moving average presented the lowest root mean squared errors for AR, LA, and TX (5.2%, 11.0%, and 7.3%), while ARIMA presented the performance when predicting prices for MS (16.0%). The results of this study proved to be very useful for timberland shareholders and investors.
Student Scientists: Mateus Sanquetta
Advisor: Dr. Bruno K. da Silva
Department: Department of Forestry, College of Forest Resources
WabileNet: Wavelet-Based Lightweight Feature Extraction Network
Designing a lightweight convolutional neural network with high accuracy is challenging due to the information loss, which mainly affects small objects when a deep network is designed due to the pooling operation at each layer of the network. This work presents a lightweight convolutional neural network using discrete wavelet transform (DWT) and inverse wavelet transform (IWT) as downsampling and upsampling operators without pooling and stride operations. The downsampling layer is built using DWT and downsampling block, whereas the upsampling layer is designed using IWT and upsampling block. The wavelet property assists in keeping the high-frequency information of images, which is discarded in the standard convolution. Additionally, in the standard convolution, a single filter is applied, but low- and high-frequency filters are applied in this work, which helps capture more input portions and increases the receptive field size. Depth-wise separable convolution is also adopted to reduce the number of parameters of the model. The proposed model has 5.69 times and 30.8 times fewer parameters than ResNet-50 and VGG16 backbone networks. The experimental results on the CIFAR-10, CIFAR-100, and ImageNet datasets show that the proposed feature extraction network’s accuracy is comparable to ResNet-50 and VGG16 networks with fewer parameters and are more suitable for small object classification. The code and pre-trained weights will be released after the paper gets published.
Class-aware Fish Species Recognition Using Deep Learning for an Imbalanced Dataset
Fish species recognition is crucial to identifying the abundance of fish species in a specific area, controlling production management, and monitoring the ecosystem, especially identifying the endangered species, which makes accurate fish species recognition essential. In this work, the fish species recognition problem is formulated as an object detection model to handle multiple fish in a single image, which is challenging to classify using a simple classification network. The proposed model consists of MobileNetv3-large and VGG16 backbone networks and an SSD detection head. Moreover, a class-aware loss function is proposed to solve the class imbalance problem of our dataset. The class-aware loss takes the number of instances in each species into account and gives more weight to those species with a smaller number of instances. This loss function can be applied to any classification or object detection task with an imbalanced dataset. The experimental result on the large-scale reef fish dataset, SEAMAPD21, shows that the class-aware loss improves the model over the original loss by up to 79.7%. The experimental result on the Pascal VOC dataset also shows the model outperforms the original SSD object detection model.
Student Scientist: Simegnew Alaba
Advisor: Dr. John Ball Responsible Unit: CAVS
Department: Electrical and Computer Engineering This research was funded by HPC2-NGI.
Project Information: Research Paper in Sensors 2022, 22(21)
Deep Learning-based Global Soil Moisture Estimation Using CYGNSS Delay Doppler Maps
Remote sensing-based soil moisture estimation has become much popular nowadays. Various satellite missions have been launched to retrieve soil moisture from the earth's surface. But the main challenge is to get soil moisture at higher spatial and temporal resolution. However, Global Navigation Satellite System (GNSS) signals at the L-band frequency can reflect off the land surface and deliver high-resolution land surface information, including surface soil moisture. Cyclone Global Navigation Satellite System (CYGNSS) constellation generates Delay-Doppler Maps (DDM)s, containing important earth surface information from GNSS reflection measurements. CYGNSS is intended to measure the ocean surface wind field with unprecedented temporal and spatial resolution, under all precipitating conditions and across the entire dynamic range of wind speeds experienced in a tropical cyclone. Mississippi State researchers from the Information Processing and Sensing (IMPRESS) lab utilized this satellite mission and proposed a new deep-learning algorithm to determine complex relationships between the reflected measurements and surface parameters which can provide improved soil moisture estimation. The model is trained and validated using the Soil Moisture Active Passive (SMAP) mission's enhanced SM products. The developed model is able to provide global surface soil moisture products at a daily 9 km resolution.
Developing a Pollen Nutrition Database for North America: Healthy Food for Healthy Bees
Poor nutrition is one of the major stressors of bee species and is a main contributor to loss in pollinator populations. Bees are currently faced with many nutritional challenges, including loss of forage habitat and monoculture. The aim of this project is to promote better nutrition for bees by learning which floral resources are nutritionally optimal for all bees based on the nutritional composition of their pollens. This study will first use various methods of pollen collection in order to collect sufficient pollen from each target species of plants. Next, the collected pollen samples will be analyzed in the lab using basic biochemical assays, as well as mass spectrometry-based methods, to determine the nutritional quality of the pollen. The pollen will be analyzed for its concentration of proteins, lipids, amino acids, sterols, metabolites, and phytochemicals. Finally, all of this data collected will be compiled into an online database showcasing the nutritional quality of each plant species’ pollen. This database can be used by beekeepers, conservation groups, researchers, growers, and policymakers to scientifically select forage plants for bee pollinators.
Student Scientist: Lauren Jennings
Advisor: Dr. Priyadarshini Chakrabarti Basu
Responsible Unit: PCB Lab, Funded by the United States Department of Agriculture AFRI
Department: Biochemistry, Molecular Biology, Entomology, and Plant Pathology
Autonomous Robot Coverage Path Planning and Mapping
In our daily life and varying society, some application scenarios using autonomous robots include vacuum robots, painter robots, demining robots, lawn mowers, unmanned harvesters, and window cleaners, which require complete coverage path planning (CPP) for robot navigation. Such scenarios typically necessitate completely covering of the entire workspace along the robot's running trajectory while optimizing travel distance, energy consumption, and time. CPP methods with its great research significance, due to its technological hurdles, have been developed in this project. CPP is extensively employed in precision agriculture worldwide. In the event of dynamic changes in the environment, however, practical implementations may confront computational complexity issues. Therefore, two different CPP frameworks are developed for indoor and outdoor settings of precision agriculture.
In outdoor settings, such as weeding and harvesting on farms. A three-layer framework is proposed: the first layer is based on the satellite map to CPP considering the shape of the workspace; the second layer is utilized to plan a collision-free path in light of the obtained onsite images by UAVs; the third layer is followed to avoid unknown and dynamic obstacles through onboard LiDAR sensors. In order to lessen computational cost, the proposed framework progresses more precisely layer by layer depending on the obtained environmental data.
In indoor settings with limited sensing capabilities, such as poultry barns, autonomous robots are required to eliminate broiler mortality. Two robots are developed using a multi-layered CPP system to detect and remove broiler mortality. One detection robot is utilized to search entire broiler barns and detect dead birds. The combination of informative path planning and CPP enables a more efficient search of the entire broiler barn based on the historical data on the distribution of dead birds. The dead bird detector is embedded to detect and indicate the locations of dead broilers. The other removal robot moves directly to the corresponding dead broiler position in an optimized order to clear the dead birds. The results of simulation, comparative studies and experiments demonstrate the effectiveness and the remarkable performance of the proposed framework.
Lei T, Luo C, Jan GE and Bi Z (2022) Deep Learning-Based Complete Coverage Path Planning With Re-Joint and Obstacle Fusion Paradigm. Front. Robot. AI 9:843816. doi: 10.3389/frobt.2022.843816
Lei T, Li G, Luo C, Zhang L, Liu L, Stephen Gates R. An informative planning-based multi-layer robot navigation system as applied in a poultry barn. Intell Robot 2022;2(4):313-32. http://dx.doi.org/10.20517/ir.2022.18
Evaluating the Efficiency and Effectiveness of Unoccupied Aircraft Systems (UAS) for Monitoring Animals
The use of remote sensing to monitor animal populations has greatly expanded during the last decade. Unoccupied Aircraft Systems (UAS) allow vast areas to be surveyed efficiently regarding cost and time. However, when monitoring mobile animals, counting errors may occur. This project aims to evaluate the consequences of counting animals using UAS and recommend an efficient and effective UAS survey approach. Agent-based model simulations will be used with varying animal distributions and movement patterns to simulate animals on a digital landscape. Multiple competing UAS flight patterns will be investigated to determine the best approach for reducing erroneous counts. These complex simulations will be conducted within a high-performance computing environment using multiple animal movement-UAS survey configurations to produce thousands of simulated surveys and thus provide robust inference for UAS survey applications in remote animal detection.
The accompanying animation depicts a UAS surveying an area and is simulating a photo being taken at each yellow square. The green dot represents the starting point of an animal with its movements tracked by the orange line moving across the screen.
Student Scientist: Emma Schultz
Advisors: Dr. Raymond B. Iglay and Dr. Kristine Evans
Additional Project Mentors: Dr. Natasha Ellison, Dr. Melanie Boudreau, and Dr. Garrett Street
Department: Wildlife, Fisheries and Aquaculture