Deploying Satellite-Imagery Based Machine-Learning Models for Large-Scale Mapping of Tillage Practices: Final Report

This BIOAg final report documents development and validation of a satellite-imagery and machine-learning platform for mapping tillage practices in eastern Washington. Ground-truth data were collected from 781 fields across Whitman and Columbia counties between 2021 and 2023, totaling more than 3,000 observations. A Random Forest classifier integrating multispectral optical imagery, Sentinel-1 radar data, terrain variables, crop type information, and customized importance weighting achieved 87% overall classification accuracy.

Inclusion of crop type improved model performance by approximately 5%, while radar data contributed an additional 4% improvement compared to baseline optical-only models. Spatially explicit maps were generated for multiple years and compared with USDA Census of Agriculture statistics, showing general agreement.

The project produced open-source code, field-scale tillage maps, stakeholder workshops, conference presentations, and a manuscript in preparation for submission to Remote Sensing of Environment. This work strengthens regional capacity for scalable verification of conservation tillage adoption and supports improved monitoring and reporting systems.