Detecting Crop Water And Nutrient Stress Using Hyperspectral Remote Sensing

The yield and quality of crops will be significantly reduced when they are subjected to environmental stress.Water stress and nutrient stress will seriously limit the growth of crops.Monitoring crop moisture status is essential for implementing precision irrigation and balancing sustainable agriculture between crop production and water supply.In the process of Chinese agricultural production,the utilization efficiency of fertilizers is relatively low.Excessive fertilization will increase planting costs and seriously damage the natural ecological environment.Real-time and accurate monitoring of crop nutrient status is very important for increasing crop yield and improving fertilizer utilization efficiency.The study of detecting crop water and nutrient stress using hyperspectral remote sensing provides scientific technical support for modern precision agriculture.We chose winter wheat as the research object for water stress study.Two water stress experiments were conducted during the growth seasons of 2015-2016 and2016-2017.Using the canopy spectral information and relative water content(RWC)data of the entire growth period and important growth period(reviving,jointing,booting and flowering stage)of winter wheat,we established winter wheat water monitoring models based on the entire growth period and different growth periods respectively.We also assessed the impact of shaded leaves on detecting water stress.We chose winter oilseed rape as the research object for nutrient stress study.13 field experiments of winter oilseed rape were conducted over 2013 to 2019 at different nitrogen(N)application rates,phosphorous(P)application rates and potassium(K)application rates.We combined the canopy hyperspectral data with physiological and biochemical data to explore the feasibility and accuracy of hyperspectral technology in the nutrient diagnosis of winter rape.The main conclusions from the study were as followed:(1)We established models to detect winter wheat water stress.NDVI,WI,PWI and seven photochemical reflectance indexes(PRI570,PRI1,PRI2,PRI3,PRI4,PRI5and PRI6)were calculated from canopy hyperspectral images of winter wheat.We analyzed the relationships between the spectral indexes and the relative water content of winter wheat.PRI3 based on reflectance at 512 nm as the reference band is the optimal model vegetation index in the whole growth period(R2=0.38).During the growth of winter wheat,R~2 of PRI6 and PWI showed an increasing trend,however,R~2of the other eight spectral indexes increased first then decreased and R~2 is highest at booting stage.PRI3 is the optimal vegetation index in the greening period(R~2=0.31),PRI2 is the optimal vegetation index in the jointing period(R~2=0.57),PRI3 is the optimal vegetation index in the booting period(R~2=0.88)and the optimal vegetation index in the flowering period(R~2=0.76).(2)This study evaluated the impact of the varying shaded-leaf fractions on estimating RWC across growth stages of winter wheat using seven formulations of PRI(PRI570,PRI1,PRI2,PRI3,PRI4,PRI5 and PRI6).Results demonstrated that for the control treatment the mean PRI of sunlit leaves was slightly higher than those of shaded leaves,but the difference between PRI of sunlit and shaded leaves increased as water resources became more limiting.Despite the difference between PRI of sunlit and shaded leaves,the significance of the linear relationship between RWC and most studied formulations of PRI did not show obvious variations with shadow fractions,except for the 100%shaded-leaf condition.Among the studied formulations of PRI,PRI3 provided the most accurate estimates of RWC with varying shaded-leaf fractions,except for the 100%shaded-leaf condition.We then applied a uniform RWC prediction model to the data of varying shaded-leaf fractions,and found that the accuracy of RWC predictions was not significantly affected in the mixture of sunlit and shaded leaves.However,RWC estimated with PRI of the 100%shaded-leaf condition had the highest RMSE,implying that PRI of the pure shaded leaves may yield inaccurate estimates of plant water status.(3)This study established an integrated model for diagnosis of N,P and K nutrients in winter rape.The random forest model was used to integrate multiple subdatasets of samples to select spectral bands that were sensitive to different nutrient deficiency levels.Ten bands(630,640,650,660,670,680,690,2000,2020 and 2070nm)were selected as sensitive bands for N.Ten bands(680,690,760,810,910,1090,1120,1420,2000 and 2040nm)were selected as sensitive bands for P.Ten bands(530,650,680,2030,2040,2070,2080,2100,2260 and 2290nm)were selected as sensitive bands for K.Through the combination of multiple random forest models,the bands that sensitive to nutrient stress were converted into probabilistic features to diagnose nutrient deficiency.The new probability features greatly enhanced the differences among nutrient deficiency levels and the normal condition.The overall accuracy of the nutrient deficiency diagnosis achieved by the proposed framework reached80.09%,which was 16.74%more accurate than the results of the single RF model,18.91%more accurate than Support Vector Machine and 36.20%more accurate than Artificial Neural Networks.