Detection Of Rape Canopy SPAD Based On Low-altitude Spectral Imaging Remote Sensing Technology

Oilseed rape is one of the most important oil crops in China. The use of nitrogen fertiliser has an effect on the yield and oil content of rapeseed in cultivating oilseed rape. Low-altitude spectral imagine remote sensing technology was used to detect the nitrogen content of oilseed rape canopy timely and efficiently, which was helpful to formulating detailed nitrogen nutrient management scheme, and then improved the yield and quality of rapeseed. In this research, Zheshuang 758, a kind of brassica napus, was used as the research object, and the SPAD value was regard as evaluation index of nitrogen content. A agriculture digital camera, carried by a set of unmanned aerial vehicle (UAV) simulation platform, was used to obtain multispectral images of oilseed rape canopy. In order to evaluate the feasibility of nitrogen detection by using low-altitude spectral imagine remote sensing technology, canopy SPAD value prediction models were set up, based on vegetation index and texture features of multispectral image respectively. The impact of three factors, included different image acquisition time, altitude and speed, on prediction models was evaluated. In addition, this research also discussed the efficiency of oilseed rape canopy SPAD value prediction models based on visible and near infrared hyperspectral imaging technology. The methods of spectral preprocessing and effective wavelength selection were optimized. Then several chemometrics methods were adopted to establish canopy SPAD value prediction models by using all bands, characteristic bands and texture features respectively. The main research conclusions obtained in this study are as follows:(1) When oilseed rape canopy SPAD value prediction models were set up based on vegetation index of multispectral images, both NIR/G and (NIR - G)/(NIR + G) had good correlation with SPAD value, and linear function models was better than quadratic function and exponential function models. PLS models were better than MLR models when the models were set up based on texture features. However, the models based on texture features were more susceptible to the influence of imaging quality than vegetation index, and the models' performance based on vegetation index was better in different condition.(2) The acquisition time, height and speed of multispectral images had different effects on predict models. When the predict models were set up based on vegetation index, the increase of growth time and acquisition height were significantly beneficial to the performance of prediction models. A slow acquisition speed improved the prediction models'performance slightly. When the models were set up based on texture features, the increase of growth time and acquisition height, the decrease of acquisition speed were significantly beneficial to the performance of prediction models. Models based on both vegetation index and texture features had the best performance in the third image acquisition time with a height of 1.9m and a speed of 0.1m/s. The correlation coefficient of prediction set (Rp) was 0.7354 and 0.7800.(3) When establishing SPAD value prediction modles based on hyperspectral images, Savitzky-Golay smoothing and successive projections algorithm were selected asspectral preproeessing and effective wavelength selection among several methods. When the modles were established by using all bands, the Rp was up to 0.8287 in the third image acquisition time. ELM modles had slightly better adaptability than PLS, MLR, BPNN and SVM in prediction modles based on feature bandsfor different rape growth time, and the biggest Rp was 0.8466 in the third image acquisition time. When the modles were based on texture feature of characteristic bands'image, the Rp of PLS model was up to 0.7341 in the second image acquisition time, which was better than MLR model.The models based on all bands and characteristic bands had better prediction performance than texture feature.This study realized the rapid detection of oilseed rape canopy SPAD value based on low-altitude spectral imagine remote sensing technology, and explored the influence law of multispectral images acquisition time, height, speed on the analytical model. This study laid theoretical foundation for application of unmanned aerial vehicle (UAV) low-altitude spectral imaging remote sensing technology to accessing a wide range of rape canopy nitrogen rapidly in the future.