Multispectral And Hyperspectral Sensing For Nitrogen Management In Agriculture

Abstract: In recent years, more and more researched for new techniques that would allow more efficient use of agronomic inputs in order to optimize yields while decreasing environmental impact. Precision farming can play a key role to fulfill these requirements. Precision farming uses the newest technologies to monitor within-field and between-field crop variability to support agronomic decisions. Moreover, precision agriculture adopts machines for site-specific distribution of agronomic inputs in order to optimize their efficiency. Among agronomic inputs, fertilizers represent a great cost for farmers and can be a source of environmental pollution if not properly managed. This is particularly true in Lombardy, a region characterized by a high risk of nitrate leaching into the groundwater. In this context, vegetation monitoring to support fertilization is very interesting. Researchers, in particular, focused on the application of remote sensing with optical sensors, because they are considered the most suitable for in-field applications. Thus, this research project began with a literature survey, whose results are presented in Chapter 1. concentration, LAI (leaf area index), above ground biomass, nitrogen uptake, grain yield, and optimal nitrogen rate. Maize was the target crop because it is the main crop cultivated in Lombardy. Ninety-one papers, published between 1992 and 2016, were identified. Relevant information describing the performance of various sensors was extracted from the papers. The performances of estimation were highly variable (R2 = 0.60-0.97). Moreover, each experiment produced specific regression equations for location, year, cultivar and development stage. This empiricism is the strongest limitation to the large-scale application of optical sensors for the estimation of nitrogen demands. The literature survey of Chapter 1 highlighted the successful local use of optical sensors to estimate crop related variables to nitrogen nutrition. However, it showed some limitations, irrespective of the studied crop. Limitations are in fact connected to the platforms the spatial resolution of the optical information obtained by the satellite sensors or the temporal and spectral resolution of the tractor-mounted sensors. Another limitation of multispectral sensors is their ability to acquire only a small number of broad spectral bands. At the same time, the literature is the solution to these problems: the use of unmanned aerial vehicles (UAVs) and the use of hyperspectral imaging sensors. The effects of combined stressors may be investigated. The effects of combined stresses may be investigated. Indeed, nitrogen stress is often combined with water stress in the Italian environment, they were not often studied in the literature. Chapter 2 reports the results of a greenhouse experiment to estimate nitrogen- and water-related variables of a model crop (Spinacia oleracea L.) using multivariate partial least squared regression models (PLS) on hyperspectral data. A completely randomized experimental design was arranged with two water levels x four nitrogen levels in two replicates. The reflectance of the canopy was acquired in 121 wavelengths, between 339 and 1094 nm, using a hyperspectral imaging system. For each pot, the average spectrum and the modified hyperspectrograms (a technique to compress the raw spectra, originally proposed in food science) were calculated and used as predictors of plant water content and plant nitrogen concentration. The best cross-validation performances were reached in the estimation of the water content, both from the average spectrum and the hyperspectrograms. The hyperspectrograms led to slightly better performance than the average spectra: R2cv (cross validation) = 0.82 and RMSECV (Root Mean Square Error in Cross Validation) = 0.86% FM for estimation of water content and R2cv = 0.57 DM for the estimation of the nitrogen concentration. The better performance in the estimation of water content (compared to nitrogen concentration). This result emphasizes the combined effect of multiple stressors on the structure and the reflectance of the canopy should be further studied. In conclusion, hyperspectral imaging proved to be an interesting technique as well as hyperspectrograms extraction, opening new opportunities for the in-field applications of this technique. Finally, knowing the great interest of UAV-based remote sensing applications, Chapters 3, 4 and 5 report results in two case studies in the field. The UAV-based optical monitoring was applied to the field of variability of winter wheat. The experimental field (Chapter 3) was monitored during two years (2014-2015) with a commercial digital camera (Canon® Powershot SX260 HS), modified to acquire reflectance in two visible channels (blue and green) and one near-infrared channel . Crop samples were taken at V6 and V9 (sixth and ninth unfolded leaves) phenological stages. These stages are adequate to carry out an diagnosis of the field. The BNDVI and GNDVI of the entire plots (soil + vegetation) were calculated from the optical images. The fractured cover. The regression equation built on two years of experimentation (V9 only) gained R2 = 0.87 and rRMSE (relative RMSE, ie the RMSE expressed as a percentage of the measured average) of 17%. The low cost digital camera led to very good performances in the estimation of the above ground biomass thanks to its high spatial resolution, compensated for the lack of adequate spectral resolution, presented in the Chapter 4). The experimental wheat field (Chapter 5) has identified the best time to make the UAV survey and to classify the field in homogeneous areas for nitrogen management. The camera used was a RedEdge MicaSense ™, which measures reflectance in five channels: blue, green, red, red-edge and near-infrared. Three vegetation indices were calculated from the aerial images (NDVI, GNDVI and NDRE). The NDRE was found to be the best estimator of grain yield (R2 = 0.76 to 0.91) and above ground biomass (R2 from 0.37 to 0.90), in all phenological stages. The most suitable time for crop monitoring was 31 BBCH. At this phenological stage, in fact, the crop monitoring guaranteed a satisfactory estimation of wheat above ground biomass Moreover, three homogeneous zones have been identified, based on the errors in biomass estimation. Finally, the average above ground level and nitrogen uptake were calculated for each homogeneous area. The experiments carried out during this process PhD project confirmed the reliability of optical sensors (multispectral and hyperspectral) The UAV was found to be a useful and reliable tool for in-field applications. Finally, it was found that, between two non-univocal relationships between optical properties and nitrogen-related crop variables, of soil and weather. Only in this way it would be possible to build a decision support system able to take into account agro-ecosystem complexity in order to provide accurate fertilization rate prescriptions.

read the paper

Authors: Martina Corti

Associations: Università degli Studi di Milano