Abstract:

【Objective】Mining domestic satellite image data information to rapidly and accurately obtain the types of crop planting structures，to provide references for optimizing agricultural production layout.【Methods】 The random forest algorithm model was used in combination with sample data to analyze the importance of spectral features，vegetation index features，and texture features of the images. The optimal feature combinations based on GF-6 images were acquired by evaluating classification accuracy. The selected features were then used in object-oriented classification research（with smoothness of 0.5 and compactness of 0.3 as parameters and 10 as steps，setting a total of 11 different segmentation scales ranging from 40 to 140），to obtain the spatial distribution of the main crop planting structure in the research area.【Results】The GVI，EVI，Nir，GI，GNDVI，and Green features obtained by feature optimization method could effectively reduce data redundancy in crop classification and improve classification efficiency. Among the 11 segmentation scales set in the research area farmland，the optimal segmentation scale was 100. The segmentation results retained the integrity of the plot and reflected the heterogeneity of different crop types. The classification accuracy based on object-oriented classification method reached 96.2%，with a Kappa coefficient of 0.944，which was 5.3 percentage points higher than pixel based classification accuracy.【Conclusion】Using domestically produced GF-6 images as the data source and employing the feature optimization object-oriented classification method could effectively improve classification accuracy and serve as an effective means for crop planting structure monitoring.

Key words: GF-6 image, Eigenvectors, Object-oriented, Crop, Planting structure