Remote sensing method for pitaya yield estimation based on hyperspectral fusion information

【Objective】This study aimed to investigate remote sensing method for pitaya yield estimation based on hyperspectral fusion information， thereby providing technical support for local governments to optimize agricultural industrial structure， guide agricultural production， and promote high-quality development of the fruit industry.【Method】Using single pitaya plants at full fruiting stage as the experimental subjects， the imaging and non-imaging hyperspectral information was fused， and the continuous projection algorithm （SPA） was employed to select yield-sensitive bands as independent variables and ground-measured yield per plant as the response variable. At a ratio of 2∶1， 126 sampling sites were divided into the modeling set （84 sites） and validation set （42 sites）， and yield inversion models based on multiple linear regression （MLR）， partial least squares regression （PLSR）， support vector regression （SVR）， and lion swarm optimization algorithm-support vector regression （LSOA-SVR） were established to evaluate estimation accuracy of single-sensor and multi-sensor fusion models.【Result】The imaging and non-imaging hyperspectral reflectance feature curves of pitaya plant canopy were basically consistent， with lower reflectance in the visible light region and higher reflectance in the near-infrared region. Within the visible light region， the green bands showed a reflection peak， while the red and blue bands showed absorption troughs. Reflectance of the canopy in non-imaging hyperspectral near-infrared region was negatively correlated with yield per plant， with low-yield plants showing higher spectral reflectance， and the imaging and non-imaging hyperspectral reflectance of canopy were 44.6% and 63.5% respectively. The extremely significant correlated （P<0.01） bands between imaging and non-imaging hyperspectral reflectance of fruit as well as pitaya yield per plant were mainly concentrated in 398-704 nm. According to pitaya yield inversion models based on single-sensor data， the MLR model had the lowest accuracy， followed by the PLSR model， and the SVR model was the best. Based on the four models based on multi-sensor fusion data， the SVR model had better accuracy than the MLR and PLSR models， and the LSOA-SVR model after lion swarm optimization algorithm further improved the performance of prediction.【Conclusion】By fusing imaging and non-imaging hyperspectral information， multi-source data collaboration can significantly improve the accuracy of remote sensing inversion for yield of pitaya at full fruiting stage. In yield inversion models based on single-sensor data， the nonlinear model SVR outperforms the linear models PLSR and MLR. As applying LSOA in parameter optimization of SVR model， the proposed LSOA-SVR method can effectively improve prediction accuracy of the remote sensing inversion model for pitaya yield per plant.