Progress 04/01/24 to 03/31/25
Outputs Target Audience:The project team participated in some capacity building, outreach and knowledge sharing events between April 1, 2024 and March 31, 2025 to share about the project as well as share progress made. These events included Lincoln University of Missouri's annual field day at the George Washington Carver Farm, which took place in June 2024 attracting variuos stakeholders from the farming, research and industry communities. During this event, a poster titled "A smart spotted wing drosophila monitoring system" was presented to field day attendees and a live demonstration was provided of the smart trap that is being tested in this project for monitoring the spotted wing drosophila (SWD). The next event we participated in was the 2024 Missouri State Fair which took place in Sedalia, MO, in August and attracted a broader audience. A demonstration was provided using the smart trap and flyers on the spotted wing drosophila obtained from North Central IPM Center were shared with the public. The third event was the Great Plains Growers conferencewhich took place in January 2025 anddrew hundreds of attendees representing farmers, researchers, extension personell, industry representatives and students from the Midwest. A poster presentation entitled 'Multilocation monitoring of spotted wing drosophila in Missouri small fruit farms' based on results obtained during the 2024 season were made by a Mastersgraduate studentsupported by this grant. In addition, an oralpresentation titled 'Smart monitoring system for spotted wing drosophila management in small fruit production' was made to introduce the audience to the on-going work as well as share progress made. In summary, we were able to reach a diverse group of stakeholders by targeting field days, grower conferences and even public events such as the State Fair. Changes/Problems:There were some challenges faced the previous year which are described below, alot of which we still need to make improvements too and hence why the upgrades described on the previous page. Challenges in Hardware Development and Deployment Power Supply Issues in Extreme Weather: Continuous rainy or low-light conditions prevent sufficient solar charging. Backup battery depletes, causing temporary shutdowns of smart traps. Unstable Data Transmission. LTE/4G signals are weak or unstable in remote deployment areas, affecting real-time data transmission. Device Maintenance and Cost. Sticky trap cards must be replaced every two weeks. Field inspections are conducted monthly, increasing labor and maintenance costs. Challenges in Model Training and Algorithm Optimization Small Object (SWD) Detection Accuracy. SWD adults are very small (2-3mm) and difficult to detect in complex backgrounds (e.g., varying light, vegetation occlusions). Data Annotation Efficiency and Quality Manually labeling large datasets is time-consuming and requires expert entomological knowledge. High labor costs for manual annotations. Model Generalization Issues. Model performance decreases when tested on new field environments not included in training. ?There were also a few identified hardware issues including. Solar power shortages occurred during extended rainy periods, leading to brief shutdowns. Rainfall affectedimage quality because rainwater stayedon the sticky cards What opportunities for training and professional development has the project provided?The participation in grower conferences and outreach events has resulted in expanded networks as there is interest in the project outcomes. In the coming year, we will have even more participation at both scientific and non-scientific conferences, as well as collaborations with growers as that will be crucial for the project. How have the results been disseminated to communities of interest?During the reporting period, we participated in three outreach events including a conference (Great Plains Growers conference), field day (Carver annualField day) and 2024 Missouri State Fair. During the field day, we had a live demonstration of AI- powered solar smart traps, and introduced attendees to the concept of real-time pest detection and its role in integrated pest management. This provided a hands-on experience for farmers andagricultural professionals. During the Missouri State Fari, we demonstrated real-time SWD detection using instance segmentation and spoke to attendees about how AI could potentatially reduce labor costs and improve pest control efficiency. Finally a presentation given during the Great Plains Growers conference was focused on introducing growers to the project and sharing research progress on SWD monitoring in Missouri small fruit farms. What do you plan to do during the next reporting period to accomplish the goals?In addition to increasing the number of monitoring locations andsmart traps deployed. We plan to monitor weekly and also address some of the challenges faced with the smart trap. Some of these plans include: Hardware System Optimization and Upgrades to enhance environmental adaptability Increase battery capacity from 12V 7Ah to a higher-capacity battery for improved power supply during prolonged rainy periods. Optimize solar panel efficiency and MPPT controller performance to enhance charging under low-light conditions. Develop a refined low-power mode that automatically reduces energy consumption during rainy weather to extend operational time. Improvements in Enclosure and Protectionfor better waterproofing, dust resistance, and corrosion protection; and to enhance durability and reliability for long-term outdoor deployment. Enhancing Communication Stability by: Testing and integrating high-gain LTE antennas or satellite communication modules to improve network stability in remote areas. Optimizinglocal data caching mechanisms to improve on-device storage and cloud synchronization, ensuring data security and completeness. Model Performance Improvement and Generalization by adopting advanced deep learning models and dataset expansion Evaluate Transformer-based architectures (e.g., Mask2Former, Segment Anything Model) for SWD instance segmentation. Assess the robustness and generalization of these models and deploy the best-performing model on edge devices. Collect more data under varied conditions, including different crops, backgrounds, and lighting scenarios, to build a more comprehensive and generalized dataset. Explore weakly supervised and semi-supervised learning techniques to reduce the need for manual annotation and enable automated dataset updates. Improving Algorithm Efficiency Implement model quantization (e.g., INT8 quantization) and pruning for deployment on low-power edge devices (e.g., Raspberry Pi). Evaluate domain adaptation techniques to enhance model performance in new environments. Smart Monitoring System and Agri-IoT Integration Develop and deploy a remote monitoring and real-time alert system for pest population dynamics. Design a user-friendly management and interaction platform for unified data management, aiding farmers and agricultural experts in data-driven decision-making. Integration with Sage Continuum Edge Computing Platform Incorporate the smart trap system with Sage Continuum's edge computing infrastructure for efficient real-time data processing and collaborative decision-making. Utilize Sage Continuum's AI toolkit to optimize model inference efficiency, reducing computational loads on edge devices. Building a Cyberinfrastructure for AI at the Edge Leverage the Sage Continuum network architecture to enable multi-device collaboration and data sharing for enhanced agricultural intelligence. Develop and deploy a distributed real-time alert system to support precision pest control strategies. Developing an SWD Population Prediction Model Collect long-term environmental data (temperature, humidity, light conditions) and historical image data for comprehensive analysis. Build predictive models using time series analysis techniques (e.g., ARIMA, Prophet, or deep learning-based LSTM) to forecast pest outbreaks and provide farmers with proactive control strategies. Develop a real-time dashboard with historical data tracking and trend analysis for farmers and researchers. Implement automated report generation to simplify the interpretation of complex data and aid in decision-making. We will also continue to share about the research and plan to give presentations and/or demonstrations at the Carver farm annual field day, International Elderberry symposium, and Entomological Society of America annual meeting. We will utilise all the newly formed networks to disseminate this information. This year we will also engage with high school students during a week-long high school STEAM camp.
Impacts What was accomplished under these goals?
Some progress was made towards objective 1 and objective 3. The major accomplishments include smart trap development and validation, data collection using smart traps and research dissemination.An AI-enabled "smart" trapwas developed using a high resolution camera, humidity and temperature sensors, edge computing unit (Raspberry Pi), cellular data network and solar power system. The system was tested and validated under laboratory and field conditions.Monitoring was performed using tensmart traps and 21 conventional bucket traps deployed in 6 locations including the Lincoln University Carver farm, Llloyd's farm, University of Missouri Southfarm, University of Missouri Southwest Research, Education and Extension Center, Columbia Elderberry farm and the Missouri State University State Fruit Experimental Station.The traps were servicedbiweekly between May and October 2024, and SWD were collected, identified and counted under a microscope in the laboratory and then stored.Simultaneously, image data was collected by the camera onthe smart trap every hour from 6 am to 9 pm daily. Two imagery data were uploaded to the cloud and then downloaded for further processing. Sticky cards attached to the smart trap were changed biweekly and all SWD were identified under the microscope, and individualimages were taken of all collectedsticky cards to mark identified SWD for model validation. A data processing and analysis pipeline is still under development and it involves multiple steps of image preprocessing and modeling. A total of 11 objective recognition and classification models were evaluated including You Only Look Once (YOLO) models, instance segmentation models, and transformer-based models. The best performing model was YoLoV11s achieving a recognition accuracy of mAP@0.5=0.84. Weather data was also collected from either the smart trap sensor or local weather stations. In addition we participated in outreach events including a conference (Great Plains Growers conference), field day (Carver annualField day) and 2024 Missouri State Fair. During the outreach events we had live demonstrations of the AI-powered smart traps and pest-monitoring system and also shared research progress on SWD monitoring in small fruit production.Two graduate students, one Masters and one Doctoral were recruited to assist with these research activities which will also go toward their theses.
Publications
|