This work in progress innovative practice paper represents the preliminary investigation of a two-year endeavor into the development of an introductory electromagnetics laboratory curriculum aimed at sparking the enthusiasm of undergraduate electrical engineering students. Departing from conventional lecture-based approaches, our endeavor immerses students in a realm of adventure labs and projects, placing active learning at the forefront. Through meticulously crafted prompts, students engage in a self-guided journey of critical thinking, fostering a deeper comprehension of theoretical concepts compared to passive methods. Our innovative approach breaks from tradition in several significant ways by contextualizing students' explorations within a historical framework in addition to recognizing and accommodating the diverse learning styles among our students. Over the past two years, the positive impact of this approach has been unequivocal. Students consistently express heightened enjoyment of the learning process, fueled by the hands-on, discovery-based nature of the labs. Emphasizing self-guided exploration and analysis has nurtured the development of critical thinking skills, empowering students to tackle complex problems with confidence. The incorporation of historical context, diverse communication channels, and a competitive element has resulted in a deeper understanding of the fundamental principles of electromagnetics.

Land use classification is a crucial component of the field of remote sensing due to its importance in the use and development of agricultural, commercial, and industrial applications, while also aiding in environmental planning and conservation. This paper introduces two innovative iterations of Multispec- tral Efficient Land Use Classification (MELC) based on vision transformers (ViTs), namely, MELC and MELC with PreCoding (MELC-PC). The MELC model utilizes an extremely compact vision transformer design characterized by tightly constrained dense layers and low-rank convolutional filters of small kernel sizes. Both models demonstrate state-of-the-art results, showcasing not only exceptional accuracy but also remarkable computational efficiency compared to existing models and techniques. The results demonstrate that both the MELC models achieve unprecedented efficiency and accuracy results in multispectral image classification.

Machine learning in geoscience and remote sensing has reached a pivotal stage, necessitating practical implementation in real-world settings. It is crucial that these models exhibit an efficiency suitable for application in computationally limited settings, coupled with maintaining competitive accuracy levels. This paper introduces a purpose-built model designed to meet both of these requirements, surpassing contemporary models in accuracy and efficiency. The proposed model demonstrates a remarkable ability to detect subtle changes in land use images with extreme accuracy while maintaining a minimal parameter count of 34,313 (134.04 kB). This multifaceted success not only raises the bar in terms of accuracy, but also illustrates the importance of custom models for the future of machine learning.

Autonomous Unmanned Aerial Vehicles (UAV) hold the potential to revolutionize logistics and transportation. To become truly viable, this technology must prove its capability to operate safely across a wide range of environments and conditions. Factors like wind, rain, hail, birds, and the presence of other drones in the airspace must all be considered in the decision-making process. While traditional control systems struggle with the complexity of this problem, machine learning has shown promise in tackling these challenges efficiently and effectively. This work proposes to advance independent drone operation through object avoidance, data collection, and smart navigation.

With the onset of a global pandemic, it became necessary to reevaluate the nature of Science, Technology, En-gineering, and Math (STEM) outreach. In many areas of the United States and beyond, in-person events were prohibited for multiple years. It was therefore necessary to find an alternative approach to continue to impact the interest of K-12 students in exploring hands-on STEM activities. This paper presents the Future Innovators Workshop (FIW) model for conducting successful virtual STEM outreach. While still a work in progress, lessons from the FIW model can be applied to both other STEM outreach efforts in addition to future iterations of the virtual Future Innovators Workshop.

Satellite-based land surface temperature (LST) data characterizes the Earth’s surface temperature central to a wide range of applications including ecological monitoring, urban heat distribution mapping, and many others. However, the data is often plagued by atmospheric interference causing corrupted retrievals. Additionally, the spatio-temporal resolution of many of the available products is limited, constraining the scope of applications characterizing features with both high spatial and temporal heterogeneity. While solutions have been proposed for both the retrieval corruption and resolution problems, they often suffer from data-quality issues and only solve one of the two aforementioned LST product limitations. In this paper, we propose a novel attention-enhanced CNN architecture capable of rectifying both the corruption and resolution problems by reconstructing gaps in low spatial-resolution LST products and performing 33x super-resolution using land classification and elevation models as a baseline. The model produces a daily, 30-meter resolution LST image with an overall MAE, RMSE and R2of 1.96°C, 2.59°C, and 0.961, respectively.

Land use classification has been a topic of great interest for many years. The ability to automatically recognize land cover as well as detect changes in remote sensing images is important for a variety of industries. Due in part to the size of satellite images, as well as the inefficiency of current methods, machine learning solutions are often very large. With a growing need for on device analysis, it is becoming more important to perform accurate processing without over utilizing resources. To meet this growing need, this paper presents an efficient algorithm for land use classification. The efficacy of the proposed solution is verified through studies to accurately classify remotely sensed images within the EuroSAT dataset.

Satellite imagery enables spatially-temporally continuous monitoring and understanding of global environmental factors for a range of applications. This data, however, often suffers from gaps in retrieval from sensor malfunction or atmospheric interference, particularly dense clouds that obscure parts or all of an area. For dynamic datasets such as the MCD19A2 Aerosol Optical Depth (AOD) dataset, gap filling is especially challenging. The difficulty lies in the often large, continuous blocks of cloudy pixels with missing data that limit the ability of spatial filling and the daily fluctuation in features such as AOD that incur high difficulty in gap filling from temporal trends. In this study, we propose a spatiotemporal long short-term memory (LSTM) convolutional autoencoder method that effectively reconstructs missing data resulting from thick cloud interference for MODIS AOD data. The proposed method outperforms previous methods of reconstructing data lost to thick cloud interference in AOD retrievals with a generalized network achieving a weighted average PSNR, SSIM, and R2of 47.2, 0.992, and 0.941, respectively, between original, cloud-free days and those same days masked with simulated thick cloud interference without the need for additional covariates.

Through the development of machine learning and computer vision, image scene classification has made immense progress over the last decade. Remote sensing land use analysis remains a topic of great interest. Using deep learning methods from computer vision, we develop a novel approach that combines a convolutional structure and gated recurrent unit layers with a fully connected neural network to solve land use classification tasks. Simulation studies confirm the proposed method can more accurately classify and recognize remote sensing images in the EuroSAT dataset than current state-of-the-art algorithms.

As methods for detecting hidden data evolve, there exits an ever increasing need to develop new steganographic solutions. This paper introduces novel spread spectrum (SS) and improved spread spectrum (ISS) multimedia data embedding techniques using L1principal component signatures. The design presented performs well in terms of bit error rate and the structural similarity index metric.