Objective  Means to upgrade the resolution of the images obtained by the currently available Slentinel-2 optical imagery technology were explored for better identification of rapeseed crops in mountainous southwest China.

  Method  Sentinel-2 images of rapeseed crops acquired from the satellite in space were modified using image reconstruction and fusion technology to increase the spatial resolution by varying the spectral bands. Image quality as to how accurate it could recognize rapeseed crops was evaluated based on a random forest, complex terrain model.

  Result  ① The fusion treatment significantly enhanced the contrast on minute details and texture changes, greatly improved the sharpness, and increased the brightness of the images. Meanwhile, the gray curves of the main features remained basically unchanged before and after the treatment. ② The enhanced spatial resolution effectively facilitated vegetation classification. The overall accuracy and Kappa coefficient differed slightly at the resolution of 2m. However, the crop mapping accuracy was significantly improved from 91.30% to 95.65% by the red edge bands applied. ③ Different red edge bands exhibited varying effects on the recognition accuracy. The combination of C2 (visible light B2, B3, and B4-red edge B5-near infrared B8) and C1 (visible light B2, B3, and B4-near infrared B8) increased the accuracy by 4.75%. The combined C3 (visible light B2, B3, B4-red edge B5, and B6-near infrared B8) and C2 enhanced the accuracy by 1.21%. Although both red edge B5 and B6 bands could improve the overall accuracy, B5 was more effective than B6. The combination of C4 (visible light B2, B3, B4-red edge B5, B6, 7-near infrared B8) and C3 resulted in an increase on the mapping accuracy by 4.35% as well as a user accuracy by 0.57%. The red edge B7 was most effective of all. The random forest model showed, under the improved conditions, the normalized importance metrics of characteristic band for the blue band B2 to be 0.94; for the green band B3, 0.82; for the red band B7, 0.89; and, for the red edge B5, 0.75. The results, consistent with those obtained under the band combinations, indicated that B7 and B5 bands contributed more significantly to the accuracy improvement.

  Conclusion  The spatial resolution of Sentinel-2 images could be significantly enhanced through image reconstruction and fusion. The accuracy of rapeseed crop identification by various band combinations was analyzed by the quantitative measurements of the importance of characteristic bands under the random forest classification model to arrive at the conclusion. As the first comprehensive study of its kind, the information obtained would be of value for further applications of the Sentinel-2 imaging system.