1. College of Forestry, Southwest Forestry University 650233, China
2. College of Big Data and Intelligent Engineering, Southwest Forestry University 650233, China
3. Institute of Big Data and Artificial Intelligence, Southwest Forestry University 650233, China
4. Key Laboratory of National Forestry and Grassland Administration on Forestry and Ecological Big Data, Southwest Forestry University, Kunming 650233, China
Abstract:To explore the extraction effect of the rich spectral information of hyperspectral images on tea plantations in mountainous regions and to promote the application of domestic satellite hyperspectral images in tea plantations distribution mapping and resource monitoring. Taking the typical distribution area of tea plantations in the southern mountainous region of Pu'er City as the study area, an algorithm for tea plantations extraction in subtropical mountainous regions based on the random forest (RF) classifier was constructed, and the main data sources include the hyperspectral 5 AHSI (GF-5 AHSI) image, digital elevation model (DEM), and the field survey data. Firstly, the 250 bands of GF-5 AHSI image after removing the noise bandsused as spectral features (SF), Based on the spectral analysis of the main features (tea plantation, forest and cropland) in the study area and DEM data, 45 vegetation index features (VIF) and 3 topographic features (TRF) were constructed, respectively.. Moreover, the RF was used to rank the feature importance of each type of feature, and the features were input into the RF classifier for tea plantations extraction in order of feature importance from highest to lowest. The feature dimension of the optimal feature space is determined when the F1-Score of the tea plantations reaches saturation and no longer increases significantly with the continuous input of features. Finally, 12 classification schemes were constructed based on 3 feature factors (SF, VIF and TRF). And the accuracy of tea plantations extraction was compared among the 12 schemes and the optimal scheme was finally determined. The results showed that the producer’s accuracy (PA) and user’s accuracy (UA) of the 6 classification schemes after feature selection (FS) were better than those of the 6 schemes before FS, and the VIF+TRF scheme had the best extraction accuracy (PA: 89.72%, UA: 81.97%) among the 6 classification schemes before FS, while the best performance of tea plantations extraction accuracy after FS was the SF+VIF+TRE scheme (PA: 90.69%, UA: 83.09%). The F1-Score of tea plantations extraction with different feature combinations was ranked as SF+VIF+TRE+FS > TRF+VIF+FS > SF+TRF+FS > TRF+VIF > VIF+FS > SF+VIF+FS > SF+VIF+FS > SF+VIF+TRF > SF+FS > SF+TRF > VIF > SF+VIF > SF. Among the 6 classification schemes after FS, the bands that were selected twice in the 4 schemes in which SF was involved in classification were b4, b5, b6, b27, b133, b150 and b281; the indices that were selected four times in the 4 schemes in which VIF was involved in classification were REP, VOG2, SR2, SR3, WBI, TIP3 and TIP9 and all terrain factors were selected in the 4 classification schemes in which TRF participated. SF+TRF+VIF features combined with RF algorithm after FS can effectively identify the distribution of subtropical tea plantations with good recognition accuracy and credibility. The GF-5 AHSI satellite data has good potential and prospect for application in the fields of tea plantations distribution mapping and resource monitoring.