An AI based Solution for Automated Forest Surveillance.
Country: Greece, Norway
Areas of Experimentation: Timeseries forecasting using both off-the-shelf and custom models, Geospatial models for location-based optimization.
EU Funding Source: https://imagineb5g.eu/
Our project in a few words
Agriculture and Forestry have been extensively researched and various technologies have been developed aiming to significantly optimize the processes under those sectors. Among the different research and innovation areas, there are plenty of important topics that need to be addressed. In this Project, we aim to develop and deploy solutions related to the accurate real-time monitoring of forests, considering a set of multiple metrics.
Furthermore, the set of diverse IoT ground sensors, on-site (forest) processors, and drones will be the main architectural components, hardware-wise. On the other hand, the software components will include the APIs, App UIs, dockerized software components, and ML/AI models for Computer Vision (CV) and data analytics. The combined ground and aerial system will provide a series of raw data that will be processed in the edge and far-edge servers. The scenario includes the drone with an onboard multispectral camera which will enable direct image and video streaming to the edge cloud server through the mobile application. There, the video can be fed to a CV model capable of identifying key elements related to the forest’s health while also identifying potential threats that are directly related to the health of the forest (trespassing, fire, illegal hunting). Additionally, the ground sensors provide real-time soil and air metrics in terms of temperature, humidity and PH. Thus, if any sudden or suspicious changes occur, the drone will be able to fly directly closer to the selected location for further inspection.
Who will help implement the AI solution?
Local AI is a recently established High Tech Startup in Kalamata, Greece, aiming to develop innovative solutions in the sustainability area powered by AI algorithms. Moreover, Local AI is aiming to get support from AI promotion national and EU funding to use the deep knowledge of its team in AI in Smart City, and Green Energy transition programs. Local AI is a member of the Smart Attica European Digital Innovation Hub (https://www.smartattica.eu/), which is providing us, among other things, the computational resources to train our AI models and establish a robust cloud presence for our software modules. One of the products under development is EasyRide, a green Urban Mobility navigation platform that, in cooperation with municipalities, is aimed to promote green mobility to citizens and visitors. The innovation of it is based on AI search algorithms that will optimize human experience, selecting the safest, coolest, and easiest routes when walking and cycling. We are also developing an AI -powered Charging Network Rollout optimization tool (CNROpt) and Hrvatski Telekom is one of the operators that we have agreed to pilot the tool with. This tool is already doing time-series forecasting for Charging Station utilization that is used for capacity expansion planning. We are also funded under the Interconnect H2020, for the AI4CS project with support from Hrvatski Telekom, for the pilot implementation of an AI-powered mobile app suggesting the best charging options for their EV Charging customers. Local AI team members possess extensive experience in both the theoretical and the practical aspects of AI applications and Reinforcement Learning, for resource allocation, object, and dynamic recommendation systems. Therefore, they can adapt and develop such notions to create an end-to-end solution that is specifically tailored to the domain-specific needs of the ImagineB5G project.
What is the AI solution the project plans to implement?
Concept and objectives
The proposed experiment within the framework of IMAGINE-B5G aims to pioneer a transformative approach to forest and environmental monitoring, leveraging cutting-edge technologies in agriculture, forestry, and 5G communication. Our objectives are strategically designed to address critical challenges in real-time forest management and environmental conservation. These objectives are clear, measurable, realistic, and fully achievable within the project’s specified duration, without the need for subsequent development.
Real-time Forest Metrics Monitoring: Our foremost objective is to establish a real-time monitoring system for critical forest metrics, including temperature, humidity, pH levels and other metric that serve as vital indicators of forest health and sustainability.
Aerial Imaging for Surveillance: We aim to utilize drones equipped with advanced cameras to capture high-resolution & multispectral aerial images of the forested area. The specific goals include real-time streaming of these images and video to an external edge server through RTMP (Real Time Messaging Protocol) communication protocol over 5G network and the integration of on-map analytics for forest metric visualization. This capability will enable an immediate response to alarming changes. For this reason, DJI Mavic 3 Multispectral will be used for the scenario deployment as it offers an RGB and a multispectral camera.
Deep Learning Model Implementation:
Our project aims to implement a comprehensive suite of deep learning models designed to operate on edge enablers, facilitating real-time detection and monitoring of the forest. The initial model will be an image segmentation deep neural network, designed to perform precise image segmentation on each frame captured by the multispectral camera mounted on the UAV. This process involves dissecting the image frame into enclosed pixel areas (referred to as blobs) with consistent color and texture attributes. The color data of each identified segment can be used to compute the Normalized Difference Vegetation Index (NDVI), a metric crucial for assessing forest health. Consequently, the segmentation module holds the potential to serve both short-term (utilizing 5G capabilities) and long-term forest health monitoring purposes. This segmentation model will be based on the widely adopted Mask RCNN paradigm used in RGB image segmentation. However, it will be fine-tuned with innovative unsupervised domain adaptation methods to effectively process aerial multispectral images. Moreover, the second model is an object detection neural network that excels at recognizing various tree species within aerial imagery frames. By employing this model, the system generates bounding boxes that precisely encapsulate each tree identified in images captured by the UAV’s multispectral camera. This information will be used to estimate the extent of thinning or deforestation required for specific forest locations.