Multi-modal recognition method for non-grain cropland using remote sensing time series

Large-scale non-grain cultivation has posed a serious risk to national grain security. Non-grain cultivated land can also deteriorate the ecological environment, such as soil quality, greenhouse gas emissions and agricultural pollution. Therefore, it is of great significance for the accurate monitoring of non-grain cultivated land using remote sensing data. However, the existing research has focused mainly on the distribution of specific crop types or cash crop expansion. It is still lacking in the non-grain croplands and their spatial distribution. In this study, the remote sensing time series were selected to identify the non-grain croplands in Gaocheng District, Shijiazhuang City, Hebei Province, China. The spatial distribution of non-grain cropland was also obtained from 2019 to 2022. The results demonstrated that: 1) Both pixel and object-oriented machine learning showed a high classification accuracy to identify non-grain cultivated land. Spectral bands were integrated to extract the phenological features from the NDVI time series. The growth status and temporal patterns of different crops were captured to improve the classification accuracy. The low accuracy was also found in the time series matching on the open-field vegetables, greenhouses, and uncultivated croplands, due to the high within-class variability. A better performance was achieved to distinguish between grain crops and other land cover classes. 2) The pixel approach was more sensitive to the land cover at the pixel level, thus capturing the differences among various types of land cover. The object approach effectively reduced the salt-and-pepper noise, and then significantly improved the confusion among land cover categories with the high within-class variability. 3) The object machine learning exhibited the highest accuracy for the classification and identification of non-grain cropland conversion, with overall accuracies of 87.00% and 81.00% for two growing seasons, respectively. 4) The annual non-grain cropland area was 2 753.09 hectares, with the orchards accounting for up to 62.55% of the total. The seasonal non-grain analysis showed that the autumn non-grain area (3 174.86 hectares) was significantly higher than the summer ones (1 060.27 hectares). In conclusion, the different remote sensing time series can be expected to accurately identify and monitor the non-grain cultivated land. The specific recommendations were selected for the practical applications, according to the application needs and data availability. Machine learning can be selected to fuse the harmonic and phenological features, when the specific types of non-grain cultivated land need to be distinguished. However, the time series matching can also be used for the limited sample sizes without considering the specific non-grain cultivated land.