Soil organic carbon estimation from airborne images using machine learning

Overview

This research develops a machine learning pipeline that uses drone-mounted RGB and infrared imagery to estimate soil organic carbon pools from bare soil surfaces. By correlating high-resolution aerial images with measured soil organic matter (SOM) values across experimental plots, the work aims to produce a practical Soil App capable of proximal sensing-based SOM detection. The long-term objective is to support scalable, cost-effective assessment of soil carbon sequestration potential in temperate agricultural landscapes, where finding and maintaining bare soils across a range of SOM levels is otherwise logistically difficult.

Technical specifications

• Imaging platform: Drone-mounted RGB (red, green, blue) camera paired with infrared (IR) sensors for proximal sensing of bare soil.
• Image resolution: Approximately 5 sq cm resolution aerial imagery collected over experimental plots.
• Ground sampling protocol: Five soil samples collected from 1 sq m grids of the top 10 cm layer of each plot, with four replications across SOM rates ranging from 1.5% to 6%, sourced from biochar and compost amendments.
• Modeling approach: Interpolated maps generated from aerial images and corresponding SOM measurements are used to train a machine learning model, including deep learning algorithms for image-based estimation of organic carbon pools.
• Deliverable concept: An image database and Soil App designed to detect SOM content of a soil through proximal sensing.
• Prior experience: The research team has prior experience processing aerial images for machine learning-based plant detection and variable rate spraying system development.

Technology readiness level

The project is at an early-to-mid research stage. Prior work has established competency in processing aerial imagery for machine learning-based detection tasks, including variable rate spraying. The next phase involves collecting drone imagery and paired soil samples across replicated SOM treatments, training the model, and building a prototype Soil App. The technology has not yet been validated for commercial deployment and requires further field validation, model refinement, and testing across diverse soil and environmental conditions before readiness for operational use.