Dynamic monitoring of cropland via remote sensing imagery is vital for safeguarding cropland resources and ensuring food security through accurate change detection (CD). However, convolutional neural network (CNN)-based methods, reliant on extensive convolution and pooling operations to extract deep cropland features, often accumulate irrelevant features and lose critical ones, degrading performance, especially in dense, continuous cropland regions. To address this, we first modified the recurrent residual convolutional neural network (R2U-Net) to improve the ability to extract cropland boundary, and developed a novel feature-difference-enhanced network (FDE-Net) to capture pixel-based cropland change information. Furthermore, we introduced a majority voting mechanism method to generate parcel-based cropland change maps from the cropland boundary and pixel-based CD results produced by the CNN. Using Sentinel-2 imagery as our data source, experiments revealed that R2U-Net achieved precision and F1-scores of 89.93% and 85.78%, respectively, for cropland boundary extraction. The FDE-Net yielded pixel-based cropland change detection results with precision, F1-scores and IOU of 68.09%,75.08% and 84.26%, respectively. Ultimately, we produced parcel-based cropland change maps for the Huicheng District through the majority voting mechanism, resulting in precision and F1-scores of 78.91% and 81.64%. Additionally, we extracted various types of cropland changes in the Huicheng District from 2019 to 2020, including cropland to cropland, cropland to urban, and cropland to other uses. Compared to the pixel-based method, the overall accuracy for these three change types was 84.86%, 81.86%, and 82.36%, respectively. Our methodology can obtain a more comprehensive and precise depiction of land use changes. Critically, this approach has surmounted challenges posed by fragmented or ambiguous change detection regions, ultimately elevating the reliability and fidelity of our findings.