SubterraAI has evolved into a highly modular platform that integrates edge computing and real-time image analysis from Minirhizotron tubes. These high-resolution cameras are strategically placed to capture root structures at multiple soil depths, enabling in-situ root phenotyping without disturbing the soil.

The development of SubterraAI now leverages split computing, a new feature designed to tackle the challenges of deploying large deep neural networks (DNNs) on memory-constrained devices. This approach splits the DNN into two parts:

• Edge Device: Executes a lightweight version of the model to handle preliminary data processing.
• Cloud Server: Offloads complex data processing tasks, optimizing energy efficiency and reducing communication latency.

By employing split computing, SubterraAI ensures rapid data processing in the field while maintaining model accuracy. Using the YOLOv8 architecture, we have demonstrated that splitting the network can compress the model while retaining the necessary accuracy for real-time root analysis. This innovation allows for faster processing and improved resource management, even in resource-constrained environments.

SubterraAI incorporates an advanced multimodal deep learning architecture (v3) that includes the following models:

• U-Net: Excels in consistent data environments, such as medical imaging, providing robust root segmentation.
• YOLOv8: Optimized for real-time predictions, suitable for scenarios requiring rapid root structure analysis.
• Detectron2: Particularly effective for heterogeneous datasets, adaptable to field conditions where environmental factors vary.

We achieved notable precision and recall scores:

• YOLOv8: Precision - 0.85, Recall - 0.85
• Detectron2: Precision - 0.98, Recall - 0.98

This combination of models allows us to flexibly select the best-suited model for each specific task, ensuring accurate root segmentation and trait identification.

1. Data Capture: Minirhizotron tubes capture high-resolution images of root structures at multiple soil depths.
2. Edge Processing: Initial image analysis and preprocessing are performed on edge devices using a split version of the DNN.
3. Cloud Processing: The preprocessed data is sent to the cloud server for advanced image analysis using the multimodal deep learning architecture.
4. Results: Real-time data analysis enables root phenotyping without soil disturbance, providing researchers with actionable insights.

• Energy Efficiency: Split computing reduces the computational load on edge devices, optimizing battery life and processing speed.
• Real-Time Analysis: YOLOv8's real-time capabilities enable rapid data processing and feedback in field conditions.
• Flexible Model Selection: The combination of U-Net, YOLOv8, and Detectron2 allows for tailored analysis depending on environmental and data conditions.
• High Accuracy: Achieved precision and recall scores ensure reliable root segmentation and trait identification.