Crop identification by synergistic Sentinel-2 and GF-3 multi-feature optimization

Global climate has posed the significant challenges on agricultural production and food security in recent years. Alternatively, satellite remote sensing can be expected to timely and accurately monitor the crop types and planting structures in precision agriculture. An agricultural survey is then required to estimate the crop yields and disaster impacts. Existing remote sensing data and ground sample can also be fully utilized to combine with feature engineering techniques and machine learning. The accuracy of crop classification can be enhanced after quantitative remote sensing using satellite-derived spatiotemporal big data. However, there are the substantial variations in the feature selection algorithms and machine learning under different scenarios, leading to the significant fluctuations in the recognition and classification accuracy. Therefore, it is very necessary for the individual remote sensing imagery and multi-feature selection during crop recognition and classification. This study aims to realize the crop identification using synergistic Sentinel-2 and GF-3 multiple feature optimization. The study area was taken from the Ying Shang County in Anhui Province, China. Sentinel-2 and GF-3 satellite imagery data was employed to extract 58 feature indicators, including spectral, index, texture, and polarization characteristics. Subsequently, three algorithms of feature selection and three machine learning were combined to design three experimental schemes, in order to explore the effects of feature selection and machine learning techniques on the crop classification. Feature dimensions and classification accuracy were then compared to evaluate the effectiveness of various classification schemes. Several key insights revealed that: (1) The feature selection algorithms were effectively reduced the dimensionality to avoid the data redundancy associated with excessively high dimensions. After that, Relief F algorithm performed the best to reduce the feature dimensions with the effective performance of crop classification. Specifically, Relief F selected 28 features, which was 4 fewer than the number selected by RF_RFE and 22 fewer than those selected by CST. (2) A comparative analysis was performed on the three algorithms of feature selection. It was found that the red-edge feature bands (B5, B6, and B7) were emerged in three algorithms during crop classification. Additionally, the bands B11 and B12 with texture features were further enhanced the accuracy of crop recognition. Different algorithms were varied in the preferences during feature selection; Specifically, Relief F algorithm was tended to favor the spectral and index features. (3) Three machine learning revealed that the combination of Relief F with Random Forest (RF) was achieved in the highest accuracy of crop classification in the study area. While the XGBoost and Support Vector Machine (SVM) shared the inferior performance. In Relief F feature selection, the RF was achieved in an overall accuracy of 93.39%, a Kappa coefficient of 0.8933, and an F1 score of 94.31%. These results indicated that the RF outperformed the XGBoost and SVM by 1.36 percentage point, 0.021 and 1.31 percentage point, while by 8.81 percentage point, 0.1312 and 8.78 percentage point, respectively. This experiment was strongly validated the effectiveness and advanced nature of combined Relief F feature selection with RF classification. Feature selection with suitable machine learning can be expected to significantly enhance the classification performance under various agricultural contexts. These findings can greatly contribute to the effective utilization of remote sensing technologies in agricultural monitoring and decision-making. The relationship between feature selection and classification accuracy can also provide the technical approach for the recognition and classification in precision agriculture.