Automated remote sensing monitoring of cropland non-agricultural and non-grain conversion at parcel scale in complex environments through multi-source data fusion

Changes in cropland use, particularly the transition from agricultural to non-agricultural and non-food crop production, can diversify rural economies but may also pose challenges to regional food security, especially in densely populated and rapidly developing regions such as China. High-precision monitoring of cropland non-agricultural and non-grain conversion is essential for balance regional food security with rural income enhancement. This study focuses on the monitoring cropland non-agricultural and non-grain conversion in the rainy and cloudy regions of southern China. We aim to develop an automated process framework that accurately extracts parcel boundaries and identifies multiple types of changes. Quantitative experiments assessed the effectiveness of various solutions for key modules in the framework, including multisource data fusion, image segmentation, sample generation, and classification feature strategies. Validation using verification samples obtained through visual interpretation and field surveys revealed the following results: (1). The use of both optical and SAR images improved classification accuracy by 1.30% compared to using optical images alone. (2) Under the constraint of vector patch data, segmentation using high-resolution images outperformed both segmentation using medium-resolution images with the same constraint and segmentation using high-resolution images without the constraint, achieving Mean Intersection over Union (MIOU) improvements of 0.28 and 0.24. (3) Samples automatically generated from vector patch data achieved classification accuracy comparable to that of manually selected samples, with only a 0.64% decrease in overall classification accuracy. (4) Classification utilizing time-series feature extraction from reconstructed data outperformed classification based on temporal feature, with an overall accuracy increase of 1.94%. The optimized automated process framework achieved an overall accuracy of 89.00% in monitoring cropland conversion in the complex planting conditions of southern China. This framework represents an effective approach for the automated monitoring of cropland non-agricultural and non-grain conversion with precise parcel boundaries, providing valuable insights for similar monitoring objectives and application scenarios.