Accurate cropland segmentation is essential for agricultural monitoring, with deep learning (DL) models relying heavily on training data quality. The USDA’s Cropland Data Layer (CDL) is widely used for this purpose, but it includes classification errors and noise. Although a confidence layer is provided, existing refinement methods often use arbitrary thresholds without systematic evaluation, limiting their effectiveness. To address these challenges, this study proposes a refinement method that integrates image filters with the CDL confidence layer to reduce noise and preserve reliable pixels, improving CDL data for major crop segmentation from Sentinel–2 time series data. The method’s effectiveness and generalizability were evaluated across two study areas with diverse crop types, including corn, cotton, rice, and soybeans, using two DL models (U–Net++ and DeepLabV3+) and two image filters (Majority and Expand–Shrink) applied with varying confidence intervals. To further assess generalizability, quantitative statistical significance tests were conducted. Results indicate that integrating image filters with confidence intervals from +5% to +55% improves CDL data quality for DL training, increasing segmentation accuracy by approximately 1.5% overall and up to 2% for corn and soybeans. The proposed method outperforms other refinement approaches, demonstrating its potential to enhance CDL–based DL segmentation.
Maleki et al. (Sat,) studied this question.