An Automatic Sample Augmentation Method for Paddy Rice Mapping Based on Segment Anything Model and Phenological Features—A Case Study in Southwest China

The remote sensing mapping of paddy rice in Southwest China faces challenges such as fragmented parcels and difficulties in field sample collection, hindering deep learning technology applications. To address sample scarcity for deep learning-based paddy rice mapping, this study proposed an automatic sample augmentation method for paddy rice mapping based on the Segment Anything Model (SAM) and phenological features. Multi-dimensional phenological feature dataset was constructed from Sentinel-2, Landsat 8/9, and MODIS imagery. Utilizing SAM's “point prompt” mode and temporal similarity measures based on Edit distance for Real Penalty (ERP), high-quality sample data were automatically augmented from limited field-collected samples. A novel one-dimensional deep learning model (MBiTGA) was developed to adapt point sample inputs. Results demonstrate that: (1) The augmentation method increased paddy rice samples by 25-fold and non-paddy rice samples by 95-fold, maintaining quality through strict multidimensional ERP distance controls; (2) Mapping accuracy showed a gradient improvement pattern of “composite sample set > SAM-augmented sample set > field-collected sample set,” with the composite sample set achieving a 17% improvement in Matthews Correlation Coefficient (MCC) and a 16% improvement in Mean Intersection over Union (mIoU) compared to field-collected sample set; (3) The MBiTGA model consistently outperformed five baseline models across all evaluation metrics, achieving superior results for mIoU (0.8806), Overall Accuracy (0.9365), MCC (0.8749), F1 score (0.9365), Precision (0.9370), and Recall (0.9379). This study provides an effective deep learning-driven approach for fine-scale paddy rice mapping in challenging agricultural environments, contributing valuable insights for agricultural monitoring, food security, and market stability.