| China is the leading producer of lettuce in the world. The country contributes almost half of world’s total production. With the development of modern agriculture, an understanding into the rapid measurement of nitrogen (N), phosphorus (P), potassium (K) and water contents could promote the agriculture industrialization of the vegetable. In this paper, we established a single sensor and multiple sensors models by spectral and image technology, combined with chemometrics algorithms and image processing technology for the detection of lettuce canopy nitrogen, phosphorus and potassium and water content. The growth rhythm and its relationship with nutrient content, growth information and production were studied. Based on the models, the strategy of time-domain variable rate fertilization according to grow was proposed. The specific research contents and results are as follows:The raw spectral data were processed by Savitzky-Golay filt (SG), first-order derivative based Savitzky-Golay filt (SG+FD), logarithmic transformation based Savitzky-Golay filt (SG+Log(1/R)) and continuum-removed based Savitzky-Golay filt SG+CR. The SG+FD method was used for N, P and K spectra. The SG+Log(1/R) was used for water spectra. The samples set was subdivided into calibration set and prediction set by random sampling (RS), kennard-stone (KS) and sample set partitioning based on joint X-Y distances (SPXY). The results of SPXY were superior to RS and KS.Interval partial least squares (iPLS) algorithm, synergy interval partial least square (SiPLS) algorithm and backward interval partial least square (BiPLS) algorithm were used to select optimal spectral intervals, PLS models were established, the results showed that BiPLSN> SiPLSN> iPLSN, SiPLSP> BiPLSP> iPLSP, BiPLSK>SiPLSK>iPLSK, BiPLSw> SiPLSw> iPLSw. Genetic algorithm and successive projection algorithm wereused to select optimal spectral variable, the optimal spectral variable of N were 482,513,522,569,641,691,704 and 821nm; the optimal spectral variable of P were 675,680,972,991,1476,2016nm; the optimal spectral variable of K were 463,551,652,683,729,987,1041nm, the optimal spectral variable of water were 967,1170,1221,1406,1484,1942 and 1985nm.6-8 optimum spectral variables were selected from 2151 wavelengths using the number of algorithms.The models of N, P, K and water were established by multiple linear regression (MLR), radial basis function neural networks (RBFNN) and Extreme learning machine (ELM). The results of ELM were superior to MLR and RBFNN. The lowest root-mean-square error of prediction (RMSEP) of N model was 0.2842%, the correlation coefficient in the prediction set (Rp) was 0.9218; RMSEP of P model was 0.5164g/kg, Rp was 0.8462; RMSEP of K model was 0.1741%, Rp was 0.8749; RMSEP of water model was 197%, Rpwas 0.8243.(2) The images were processed by median filter algorithm, mean filter algorithm and wavelet denoising. The effect of wavelet denoising was superior to others. The top projected canopy area (TPCA), top projected canopy perimeter (TPCP) and plant height (PH) as lettuce morphological features; red (R), green (G), blue (B), hue (H), saturation (S), and intensity (I) values as color features; entropy (ENT), energy (ASM), contrast (CON), homogeneity (HOM) in color co-occurrence matrix as textural features. ELM algorithm was used to establish detection model, RMSEP=0.4651%, Rp=0.8217 for N model; RMSEP=0.6083g/kg, Rp=0.7649 for P model; RMSEP=0.2434%, Rp=0.8167 for K model; RMSEP=238%, Rp=0.7794 for water model. The software was developed for extraction TPCA、TPCP、PH、R、G、 B、H、S、I、ENT、ASM、CON and HOM.(3) The kernel principal component analysis (KPCA) and principal component analysis (PCA) were used for diminishing computational burden. The accumulation contribution rate of KPCA was higher than PCA. The multi-information fusion models were established by RBFNN and ELM, the results of ELM were superior to RBFNN. RMSEP=0.2531%, Rp=0.9464 for N model; RMSEP=0.3679g/kg, Rp=0.9034 for P model; RMSEP=0.1349%, Rp=0.9249 for K model; RMSEP=169%, Rp=0.8918 for water model. Analysis of models respective results indicated that the N and K models were superior to the P model; the model based on a fusion of both was superior to either model based on a single sensor modality.(4) The Gomportz, Logistic and grey Verhulst models of TPCA, PH and TPCP were established. The grey Verhulst model provided a good description of TPCA and TPCP, Logistic model provided a good description of PH. The theoretical maximums of TPCA, TPCP and PH were 692.8 cm2,25.1 cm and 123.3 cm. Fast-growing points of TPCA, TPCP and PH were 16.4,16.1 and 18. Rapid growth stage of TPCA, TPCP and PH were 8.2～24.7,4.0～28.2,4.2～25.8. The whole growth period was devided into three stages, which were logarithmic phase, linear phase and senescence phase.(5) The grey relational degree was calculated between nutrient content, growth information and yield. The influential order of nutrient content for growth information was N> P> K; the influential order of nutrient content for yield was N> K> P; the influential order of growth information for yield was TPCA> PH> TPCP. The TPCA, TPCP and PH prediction models based on environment temperature and nutrient content were established, the model about nutrient solution concentration and nutrient content was also established, the strategy of time-domain variable rate fertilization according to grow was proposed.The study provided important basis for rapid measurement of N, P, K and water content in lettuce canopy, at the same time, it is conducive to the fine management and high efficient production. |
PDF Full Text Request