Hyperspectral remote sensing extraction of pitaya number by combining spectral transformation and feature selection

【Objective】Rapid and non- destructive acquisition of plant spatial information and plant number of pitaya is an important prerequisite for accurate monitoring its growth and adjusting regional planting structure. Traditional field measurement is costly and inefficient, however, hyperspectral remote sensing is simple to operate and the data is more sensitive to vegetation, so it has become an effective means for non-contact acquisition of vegetation spatial information at a large scale.【Methods】The DJI M600 low-altitude UAV equipped with Pika XC2 sensor was used to collect hyperspectral remote sensing images of pitaya growing areas in Shangguan town, Guanling county, Guizhou province. Different regions were divided according to surface complexity and the spectral curves of major surface objects were calculated using Envi 5.3. After Savitzky- Golay second- order smoothing, first derivative spectrum (FDS) and continuum removal spectrum (CRS) were derived to explore the potential of hyperspectral image data, and a feature selection method was proposed to eliminate redundant variables by defining dimension reduction strategy from feature distance. Based on artificial neural network (ANN),support vector machine (SVM) and random forest (RF) machine learning models, different ground objects in the study area were divided, and the plant number was calculated by combining the projected area of pitaya measured on the surface.【Results】The results were as follows: (1) The reflectance of the primary hyperspectral curve of pitaya was lower in the visible wavelength region and higher in the near infrared wavelength region, and the reflectance between them was connected by red edge; The spectral reflectance of pitaya and other ground objects were different in different spectral types. The primary spectrum was located in the“red valley”and“high reflective platform”, the first derivative spectrum was located in the“red edge”and“green peak”, and the continuum removal spectrum was located in the“red valley”and“green peak”. (2) The feature selection algorithm defined by the feature distance had a better dimensionality reduction effect, and the number of feature bands was proportional to the surface complexity. The number of feature bands of pitaya ranged from 2 to 9 under different spectral types in each region, and the dimensionality reduction ratio was all above 97%. The spectral transformation could effectively reduce the number of feature bands and the distance between features. The characteristic bands of different spectral types in each region were mainly concentrated in the“red valley”, “red edge”and“near infrared”regions. (3) The classification accuracy of ground objects and the extraction accuracy of plant number were inversely proportional to the surface complexity. Among all classification models, the accuracy of CRS-RF models was the best, and the overall classification accuracy and Kappa coefficient were above 84% and 0.87, respectively, indicating that the training set and the result set were completely consistent. CRS-RF models had the best effect on the number of pitaya, and the accuracy in different regions was above 83.33%.【Conclusion】The combination of continuum removal transformation and random forest algorithm can accurately identify pitaya plant information, which can provide technical reference for obtaining the spatial information of pitaya plants in karst area at a large scale. In practical application, it is only necessary to input hyperspectral remote sensing image of the study area into the trained CRS-RF model, and then the space position and plant number of pitaya in the corresponding region can be output.