Remote sensing identification for Artemisia argyi integrating multi-temporal GF-1 and GF-6 images

Artemisia argyi has been one of the typical Chinese herbs with a great an edible and medical value. The planting area has also increased significantly in southern China in recent years. It is of great practical significance to clarify the spatial distribution pattern of Artemisia argyi, particularly for the decision making on the regional crop planting structure and the optimization of industrial layout. In this study, the new identification of Artemisia argyi was made to integrate with the multi-temporal GF-1 and GF-6 satellite images. The study area was taken as the Qichun County, Hubei Province, the main production area of Artemisia argyi in China. A total of 20 spectral features were selected, including 8 single-band features, and 12 red-edge vegetation indices, according to the phenology of Artemisia argyi and the spectral bands of high-resolution images. A random forest classification was then performed to estimate the contributions of different red-edge vegetation indices to the Artemisia argyi identification. A systematic evaluation was also made to identify the potential of the integrated GF-1 and GF-6 images. The mapping accuracy was first assessed using the field samples, and then compared with four additional classification scenarios with different inputs of GF-1 and GF-6 images. In addition, the statistical data was used to verify the mapping areas extracted by remote sensing. The evaluation results showed that the integration of GF-1 and GF-6 images was generated the highest accuracy with the user's and producer's accuracy of 92.73% and 88.74%, respectively, indicating significantly higher than those of either a single GF-1 or GF-6 data only. Moreover, the fitting data of the mapping and statistical areas in each township showed that the determination coefficient R2 reached 0.70, indicating accurately matching the area and spatial distribution. The features importance was derived from random forest, where the number of red-edge bands and indices accounted for 54% of the top 50 features with the highest importance scores. The red-edge band I (B5) on June 23 (DOY204) of GF-6 data was contributed the most, which was considered as the best spectral feature to identify. The other newly added purple band (B7) of GF-6 was also valuable to distinguish rather than the traditional multiple bands. Furthermore, the optimal periods of identification were determined as early May and early September, when the first and second stubble of leaves were growing rapidly. Another optimal period was also found to extract the spatial distribution in late June and late September, when the plant was more distinguishable from others. Overall, the accuracy of crop identification was effectively improved under the newly added bands of GF-6 WFV images and the associated vegetation indices using the red-edge bands. The integration of GF-1 and GF-6 images can be widely expected to better capture the key phenological characteristics of crop types, where multiple temporal information was used to improve the classification accuracy. Consequently, the present crop identification was suitable for mapping Artemisia argyi. This finding can provide a typical application demonstration to fully realize the finer resolution of multi-source satellites.