Mapping Soil Organic Carbon for Regenerative Agriculture Using Remote Sensing and Ai

Overview

Conventional agricultural practices (like tillage and synthetic fertilizer use) have caused the world's soils to release billions of tons of carbon into the atmosphere and significantly contribute to global greenhouse gas emissions. Regenerative agriculture (including practices like cover cropping and no-till) has the potential to sequester large amounts of CO2 back into the soil as soil organic carbon (SOC) and help combat climate change. To facilitate regenerative agriculture, the ability to quantify SOC accurately and efficiently and predict agricultural practices' impact on SOC is critical. However, current methods involve manual soil sampling and are expensive and time-consuming. Farmers cannot monitor the effects of their agricultural practices on SOC and soil health, leading to the loss of SOC. The purpose of this research was to apply machine learning to develop an efficient and low-cost solution for (1) quantifying SOC by analyzing multispectral imagery from NASA's Landsat 8 satellite and (2) predicting SOC based on agricultural practices. Using the R programming language and Google Earth Engine, the Harmonized World Soil Database (containing soil properties including SOC) was linked to NASA Landsat 8 global satellite images (surface and top-of-atmosphere reflectance). Pansharpening (increasing spatial resolution) and topographic correction (correction of illuminations due to varying topographies) were applied to each image, and 151 spectral indices were extracted from each image to train multiple machine learning models, including eXtreme Gradient Boosting (XGBoost), Random Forest, and Light Gradient-boosting Machine (LightGBM). Additionally, AgEvidence datasets were used to train models for predicting future SOC based on agricultural practices. As part of this research, two novel spectral indices were developed using multiple linear regression to quantify topsoil and subsoil SOC. The models were deployed via REST API, and a mobile app was created to provide an interface to the trained models to help farmers follow regenerative agricultural practices—farmers can select their farms on a map and quantify and monitor SOC. The LightGBM model was the most accurate for SOC quantification, with a root-mean-square error (RMSE) of 0.97. For predicting the impact of agricultural practices on SOC, the RMSE was less than 5 for each category of practices. The models created are generalizable and can accurately quantify, monitor, and predict soil organic carbon to help reduce atmospheric carbon, fight climate change, and create sustainable food production.