Rapid Determination Of Growth Information And Yield Of Brassica Napus Based On Hyperspectral Imaging

Internet of things (IoT) in agriculture as a new direction of the development of modern agriculture is the key and kernel technology to improve resource utilization and productivity level, and accelerate the transition from traditional agriculture to modern agriculture. IoT in agriculture is mainly applied in the establishment of agricultural precise control and feedback systems for field and greenhouse environment. In order to achieve real-time and dynamic monitoring of the field, the precise management requires the information acquisition of the crop growth information rapidly and accurately. However, the traditional lab chemical measurements and field information monitoring methods are time-consuming and destructive, which cannot meet the needs of the development of IoT in agriculture. In this study, hyperspectral imaging technique combined with image process algorithms and chemometrics was used to detect the crop growth condition in situ in living plant samples for Brassica napus L. The content of nitrogen (N), phosphorus (P) and potassium (K) in oilseed rape leaves were determined rapidly and non-invasively. Image processing algorithms were developed for the visualization of macronutrients status of rape leaves in all pixels within an image to generate distribution maps of N, P and K content. A new method and system were also developed for rapid rapeseed yield estimation in an early growing stage. These results could provide information characterising plant growth to develop a programme of plant-specific application of fertiliser to improve agricultural profitability and minimize the impact on the environment. The main research contents and achievements are shown as follows:(1) Hyperspectral imaging technique was applied to determine and display the distribution maps of N content in rape leaves at different growing stages, including seedling, flowering and pod stage. Hyperspectral images of leaf samples were acquired in the Vis/NIR region (380-1030nm) and their spectral data were extracted from each stage. After the optimal spectral preprocessing and analysis of the PLS and LS-SVM models only based on the effective wavelengths (EWs) selected by regression coefficient (RC) and Successive projections algorithm (SPA), the best results were obtained with Rp of0.793,0.891,0.772and0.852for seedling, flowering, pod stage and the whole lifetime of seedling-flowering-pod, respectively. Besides, second-order statistic filtering algorithm was applied to extract the texture features (TFs) from hyperspectral images. Calibration models for N content detection were established by partial least squares (PLS) and least squares-support vector machine (LS-SVM) based on the combination of spectral features (SFs) and TFs. The optimal models got the best results with Rp of0.752,0.863and0.747for seedling, flowering and pod stage, respectively. Based on10EWs extracted by SPA and the developed SPA-PLS model of seeding-flowering-pod stage, the distribution maps were generated to visualize N content of rape leaves in three different stages.(2) The rapid and non-invasively detection models were developed and the visualized distribution maps were generated for P content in rape leaves. A complete comparison was first performed among raw spectra and different spectral preprocessing methods. RC, SPA and x-Loading Weights (x-LW) were proposed to select EWs. By comparison different modeling algorithms, SPA-BPNN model based on5EWs obtained the best prediction results (Rp=0.762, RMSEP=0.030). On the other hand, image TFs based on second-order statistic filtering algorithm were combined with SFs as the inputs of PLS, LS-SVM and BPNN models. The performance of BPNN model based on images at EWs got the best results with Rp of0.740and RMSEP of0.032. BPNN model also achieved the best results based on images of principal components (PCs) with Rp of0.757and RMSEP of0.032. The SPA-PLS model executed via5EWs was transferred to each pixel of hyperspectral image to predict P content in all spots of the leaf sample. The visualization of P distribution facilitated discovering the differences of P content within one sample as well as among the samples from different fertilized plots.(3) The rapid and non-invasively determination models of K content of rape leaf were established and the visualization of K content within a leaf was realized. Genetic algorithm (GA), RC and SPA were applied to acquire the EWs based on raw spectra in Vis/NIR region. By the comparison of PLS, LS-SVM and BPNN prediction models, RC-BPNN model obtained the optimal performance with Rp of0.759and RMSEP of0.158. Probability statistics filter and second-order probability statistic filtering algorithms were respectively conducted to extract the texture features from images at effective wavelengths. The prediction models for K content in leaves were developed by PLS, LS-SVM and BPNN based on the TFs combined with SFs. BPNN model was the best one with the TFs extracted by probability statistical filtering (Rp=0.730, RMSEP=0.171). Different models were also established using the combination of TFs from principal component images and SFs as input variables. The prediction accuracy achieved in BPNN model (Rp=0.726, RMSEP=0.179) were the most satisfactory one. Furthermore, the RC-PLS model built by using5EWs was applied to produce the distribution map at the EWs in each pixel of the reduced image. The final visualized maps of K content distribution could demonstrate how the nutrient content varied from sample to sample.(4) An early and rapid estimation of rapeseed yield was realized using hyperspectral imaging technique. The hyperspectral images of leaves were acquired in early growth periods (seedling, bolting, flowering and pod stage). By comparison of the performance of PLS models with the spectra extracted from different stages, the spectra data obtained from early flowering (on March25,2011) were considered as the optimal time for rapeseed yield prediction. The linear models (PLS and MLR) and nonlinear models (LS-SVM and BPNN) were developed using only6effective wavelengths selected by RC. The results showed that LS-SVM model (Rp=0.887, RMSEP=22.303) obtained a much more stable and consistent results for rapeseed yield estimation.