An Intelligent Bee Health Assessment System Using Cross-Attention-based Multimodal Neural Network for Visual and Audio Signals

Overview

Honeybees are crucial for pollinating approximately one-third of the world’s food supply. However, honeybee colonies have declined by nearly 40% over the past decade due to threats such as parasites. Traditional beehive monitoring, including manual inspect ions, is often subjective, disruptive, and time -consuming. Machine learning models have been used to improve beehive health assessments. However, previous studies have primarily relied on single-source data, such as honeybee images or sounds, and lacked co mprehensive solutions. To overcome these limitations, this study presents a Cross-Attention-based Multimodal Neural Network (CAMNN) that integrates visual and audio signals in a shared deep feature space. CAMNN achieves an 85.8% accuracy, significantly outperforming the image-only models by 25.2% and audio-only models by 19.8% across four health categories. Additionally, it demonstrates strong prediction robustness, maintaining an F1 score above 75% across all four assessed health conditions. To provide a practical solution, an advanced system has been developed to offer near real-time live streaming of hive entrance activities and CAMNN -powered hive health assessment through a mobile application. With access to real -time data and actionable insights, beekeepers can monitor their hives remotely, identify stressors, and quickly intervene to reduce colony losses. Opti-Scan: A Low Cost, Non-Mydriatic Retinal Imaging System with a Three Stage Deep Learning Pipeline for Automated Diabetic Retinopathy Diagnosis Paarth Nawani Cupertino High School, Cupertino, CA Diabetic Retinopathy affects 103.12 million people annually, with the number expected to climb to 191.0 million by 2030. Unfortunately, this issue disproportionately impacts those in low - resource regions, where 90% of all blindness occurs due to minimal ba sic eye care systems. Here, we propose Opti-Scan, an all-in-one non-mydriatic fundus camera integrated with a three- step machine learning pipeline designed for automatic diagnosis. For the fundus capture portion, we created a Raspberry Pi -powered system co nnected to a Raspberry Pi camera, 2 LEDs for fundus illumination, and an ophthalmic lens for magnification. We solely utilized off -the-shelf components to ensure that our device remained low -cost and accessible to all. Our machine learning pipeline consist s of three key components. First, we employed a Super -Resolution Generative Adversarial Network (SRGAN), specifically the Real -ESRGAN, which we fine -tuned for retinal images to enhance image quality. Next, we developed an ensemble model by combining DenseNet121, InceptionV3, and MobileNetV2, rigorously testing and fine-tuning these pre-trained models to achieve optimal performance. Finally, we implemented a YoloV5 Object Detection model to identify and draw bounding boxes around potential lesions in the fun dus, improving AI explainability. Ultimately, our ensemble model achieved an impressive accuracy of 95.2%, surpassing many current benchmarks. Opti -Scan, costing under $150, has the potential to bridge the gap in global eye care inequality, offering a cost -effective and accessible solution for early detection and diagnosis of diabetic retinopathy in underserved regions. 10 Minutes of Action Can Prevent 10 Years of Destruction: Detecting Wildfires from Space Using AI Ahvish Roy Saint Francis High School, Mountain View, CA Wildfires are among the most devastating natural disasters, threatening ecosystems, economies, and lives. The increasing frequency and intensity of wildfires, exacerbated by climate change and human activities, highlight the need for improved wildfire detection and forecasting. This s tudy explores the application of Artificial Intelligence (AI) and deep learning to enhance the accuracy and efficiency of wildfire detection using geostationary satellite imagery. By leveraging geostationary satellite data from GOES and Himawari -8 AHI, along with Low Earth Orbit (LEO) reference datasets such as MODIS and VIIRS, this research introduces an AI -driven framework to enable real-time wildfire monitoring and prediction. A Dual-Module Convolutional Neural Network (DM CNN) was developed to process both fire - related spectral signals and environmental variables, significantly improving wildfire detection by reducing false positives and increasing robustness across diverse geographic regions. By integrating both spectral and contextual data, the model adapts to varying atmospheric conditions, terrain types, and land cover classifications, ensuring more reliable fire identification in challenging environments. This dual-processing approach allows the model to distinguish real fire events from common false triggers, such as solar reflections, hot surfaces, and cloud edges. Experimental results on the LA Wildfires of 2025 show the AI -driven wildfire detection system outperforms traditional algorithms, providing higher accuracy and faster detection times. The combination of real -time meteorological updates and AI offers a comp rehensive wildfire management framework, capable of improving response times and mitigating fire -related damages. Future research will expand the training dataset for global wildfire scenarios as well as incorporate social media data into the system.