High-resolution winter cover crop mapping with PlanetScope imagery: Comparative analysis of Random Forest, Convolutional Neural Network, and unsupervised classification

Effective monitoring of agricultural conservation practices is essential for evaluating environmental outcomes and guiding land management strategies. This study assessed three satellite-based classification methods to estimate winter cover crop adoption across Benton County, Indiana, from 2021 to 2023, and compared their results with traditional field-based transect surveys. Using 3-m PlanetScope imagery and a consistent preprocessing and validation pipeline, we implemented (1) unsupervised Iso-cluster classification with five vegetation indices, (2) supervised Random Forest (RF) models, and (3) deep learning-based Convolutional Neural Networks (CNNs), tailored separately for December and April imagery. Supervised methods outperformed the unsupervised approach. RF models achieved F1 scores of 0.98 (December) and 0.96 (April), while CNNs reached 0.97 and 0.92, respectively. Unsupervised classification yielded lower accuracy (F1 ≤ 0.77), particularly under heterogeneous spring conditions. While transect surveys reported 24–128 % higher cover crop acreage than satellite-based estimates, the spatial and temporal patterns captured by both methods were similar, highlighting trends such as higher adoption after corn than soybean and substantial seasonal variation. Multi-year analysis revealed that less than 1 % of fields maintained continuous cover cropping across three consecutive winters, indicating predominantly intermittent adoption. These findings underscore the value of satellite imagery for full-coverage, repeatable assessments of conservation practice adoption. Scalable, remotely sensed classification enables timely evaluation of program effectiveness and supports adaptive land management to improve soil health and water quality at county scales.