Monitoring Of Crop Growth And Nutrition Status Based On Image And Spectral Analysis

Rapid acquisition of information about crop physiological activity is an important aspect of precise crop management in terms of monitoring growth,biomass yield,nutrition status,and quality assessment.Photosynthesis process provides essential nutrients for crop growth,and accumulation of these nutrients provides a material basis for crop biomass formation.Therefore,the key growth information such as photosynthetic trait parameters and biomass,as well as the main nutrition information such as nitrogen and protein contents,are the focus of crop production and management.Timely adjustment of crop production and management decisions based on growth and nutrition information is beneficial to avoid adverse impacts on crop yield and quality.At present,image and spectral analysis techniques have been widely applied to derive useful information for monitoring crop growth and nutrition status through canopy segmentation.However,current research involves a single crop growth environment,thus the accuracy and stability of canopy segmentation under different environments need further verification.Besides,there are few studies using image analysis technology to predict crop photosynthetic trait parameters at present.In addition,there are also issues with insufficient information utilization and data mining of image and spectra in crop biomass and nutrition status monitoring methods.Therefore,this study focused on designing crop canopy segmentation methods,established monitoring models for crop photosynthetic trait parameters,nutrition indices,and biomass based on image and spectral analysis technology and data fusion technology.Observation for the aforementioned study focus was made on forages and potatoes as test crops as they have distinct morphological development and planting patterns.The main contents and conclusions are as follows.（1）Crop canopy segmentation methods based on color component difference and transfer learning were designed by using image analysis technology for detecting forage and potato canopy regions under different planting environment and growth status.The RGB images of mixed clover and grass forage,the RGB images of potatoes in field,and the visible and thermal images of potatoes in a climate chamber were collected in this study.Three transfer learning methods,namely Deep Labv3+,Seg Net,and FCN-8s,were designed to detect the tiny clover objects in the mixed forage images.The G-R color component difference algorithm was introduced to segment the potato canopy in the field.The Mask R-CNN transfer learning model was established to extract the canopy region of potatoes from the visible and thermal images captured in a climate chamber.The results showed that Deep Labv3+could accurately segment the clover regions in the mixed forage images with the intersection over union（Io U）of 0.73.The G-R algorithm could obtain excellent accuracy for potato canopy segmentation in the field with Io U of 0.88.The Mask R-CNN had better performance for segmenting the potato canopy from the visible and thermal images captured in the climate chamber,and the Io U values all reached0.87.This study designed different canopy segmentation methods for crops in different environments,and effectively obtained crop canopy regions.（2）Prediction models of crop photosynthetic trait parameters,such as photosynthetic rate,transpiration rate,and stomatal conductivity,were established based on canopy image analysis,which achieved efficient monitoring of crop photosynthetic activity under a controlled environment.Visible and thermal images of two species of potatoes were collected in a climate chamber.Based on the potato canopy segmentation results from the Mask R-CNN transfer learning model,multi-modal image features of the potato canopy were extracted from both visible and thermal images,including statistical color components features and three-levels wavelet transformation texture features of visible images,as well as normalized canopy temperature features,gray level co-occurrence matrix texture features,and local binary patterns texture features of thermal images.The extracted features were subsequently combined to build partial least squares regression（PLSR）models for estimating transpiration rate,net photosynthetic rate,stomatal conductance（GSW）,electron transport rate,and maximum photochemical efficiency under Photosystem II（Fv’/Fm’）.The results showed that the PLSR using both visible and thermal image features as inputs had superior estimation performance.The Fv’/Fm’of D681 potato obtained the highest prediction accuracy,and the coefficient of determination（R²）,root mean square error（RMSE）,and relative root mean square error（RRMSE）were 0.86,0.77,and 14.85%,respectively.The prediction accuracy of GSW of Zhongshu 5 potato was the lowest with the R² of 0.66,RMSE of 0.09 mol m-2s-1,and RRMSE of 54.48%.This study realized rapid,accurate,and low-cost monitoring of the crop photosynthetic trait parameters by using visible and thermal image analysis-based remote sensing methods.（3）Prediction models of crop nutrition indices were constructed based on canopy spectal and image analysis,which realized the monitoring of the nutrition status of forages and potatoes and optimized the monitoring results of different crops.Canopy spectra data with the range of 350-2500 nm of mixed forages and the RGB images of potatoes were collected in this study.For the mixed clover-grass forages,PLSR and support vector machine（SVM）models were built by combining canopy spectral reflectance and plant heights to obtain in-vitro true digestibility,neutral detergent fiber content（NDF）,neutral detergent fiber digestibility,acid detergent fiber content,acid detergent lignin content,the crude protein（CP）,and crude protein accumulation.The results showed that the performance of PLSR was more excellent in terms of R² and RMSE.The prediction performance of PLSR model was significantly improved by introducing plant heights.Among the above mentioned 7 nutrition indices,the prediction accuracy of NDF was the highest with R² of 0.90,RMSE of 30.81 g kg^-1,and RRMSE of 6.58%,while the prediction accuracy of CP was the lowest with R² of 0.68,RMSE of 16.70 g kg^-1,and RRMSE of10.25%.For the potatoes,5 vegetation indices and 10 color multi-order moments were selected from potato canopy images segmented by G-R algorithm,and 6 models was built for the prediction of plant nitrogen concentration,including PLSR,SVM,ridge regression,Lasso regression,random forest,and one-dimensional convolutional regression.The results showed that SVM obtained the highest prediction accuracy with R² of 0.77,RMSE of 0.43%,and RRMSE of10.23%.This study established prediction models for forage nutrition indices by fusing canopy spectral information and plant height agronomic parameters,which significantly improved the prediction performance of forage quality.In addition,the plant nitrogen content of potato was predicted through image analysis methods,providing alternative solutions for rapid and non-destructive assessment of different crop nutrition status.（4）Estimation models of crop biomass were built based on image analysis,and improved the accuracy of biomass estimation models and achieved precise monitoring of forage and potato biomass by integrating canopy coverage,plant height,and texture features.The RGB images of the mixed clover-grass forages and potatoes were captured in this study.For the mixed clover-grass forages,this study explored the feasibility of forage biomass prediction based on the canopy coverage obtained by the segmentation results of Deep Labv3+transfer learning model.The canopy coverage,together with the plant height,was used to establish multiple linear regression,back propagation neural network（BPNN）,and SVM modes for predicting the clover biomass,grass biomass,the total forage biomass,and clover biomass proportion.The results showed that BPNN had the best prediction performance.The R² of the four biomass prediction models were 0.86,0.81,0.82,and 0.94,respectively,the RMSE were143.5 kg ha^-1,440.5 kg ha^-1,390.0 kg ha^-1,and 6.4%,respectively,and the RRMSE were 44.80%,24.63%,18.25%,and 31.57%,respectively.There were linear correlations between the grass height and the grass biomass or the total forage biomass,while the clover canopy coverage had strong correlations with clover biomass and its proportion.For the potatoes,the PLSR model achieved good results in predicting biomass based on the potato canopy coverage(CC_potato)obtained by G-R color component difference algorithm and potato height(H_potato)with R² of 0.83,RMSE of 183.52 kg ha^-1,and RRMSE of 22.58%.Moreover,this study created a canopy structural parameter of CC_potato×H_potato.The parameter was linearly correlated with potato biomass,which solved the feature saturation of canopy coverage and plant height in the later stages of potato growth.Through introducing CC_potato×H_potato and gray level co-occurrence matrix texture features,the prediction performance of PLSR model was further improved,and R² increased by 0.04,RMSE decreased by 26.91 kg ha^-1,and RRMSE decreased to 19.27%.This study achieved accurate prediction of forage and potato biomass in field by fully utilizing and mining crop canopy information,constructing a crop canopy structural parameter,and introduced the image texture features.