Alka AroraTANUJ MISRA2020-06-202020-06-202019http://krishikosh.egranth.ac.in/handle/1/5810147803T-10277Quantification of phenotypic parameter is necessary to meet the future demand of agricultural production. Conventional measurements of these traits/parameters are timeconsuming, destructive and labour-intensive. In this study, new approaches have been proposed based on image analysis and machine learning technique to derive phenotypic traits like Leaf Fresh Weight (LFW) in rice plant and spike identification and counting in wheat plant. For this purpose, images have been taken from high-throughput plant phenotyping facility established at Nanaji Deshmukh Plant Phenomics Centre, ICAR-IARI, New Delhi. In this study, it is hypothesized that combined use of visual (VIS) and near infrared (NIR) image can compute LFW more precisely than VIS image only as NIR reflectance image is used to measure water content of the plant. Two image derived parameters i.e., Green Leaf Proportion (GLP) from VIS image and Mean Gray Intensity (NIR_MGI) from NIR images have been used for building Artificial Neural Network (ANN) model to estimate LFW. The proposed approach is named as VN_LFW. The proposed approach significantly enhanced the fresh biomass prediction as compared with the conventional regression technique in both train and test dataset with Root Mean Square Error (RMSE) and Mean Absolute Percentage Error (MAPE) as 0.15 and 9.55 in training dataset and 0.13 and 9.65 in testing dataset respectively. The algorithm of measuring GLP and NIR_MGI has been proposed and the macro has been developed using Matlab software. Another significant area for spike identification has been attempted with deep learning models of Artificial Intelligence. Two models have been developed for spike identification namely SpikeSegNet (Spike Segmentation Network) and LGspikeNet (Local patch extraction and Global mask refinement Spike detection Network) based on convolutional encoder-decoder deep learning technique. For counting number of spikes per plant, “analyse particles” function of imageJ which implements flood-fill image analysis technique has been applied on the output image (binary/mask image containing spike regions only) of the developed model. For spike identification, precision, accuracy and robustness (F1 score) of the proposed SpikeSegNet model has been found as 94.56, 94.66 and 94.88% respectively whereas for LGspikeNet it has been as 99.95, 99.96 and 99.96% respectively. In spike counting using LGspikeNet, the metric values are 99, 94 and 92% respectively. Online software for identification and counting of wheat spikes has also been developed by using the proposed LGspikeNet network model. Keywords: deep learning, encoder-decoder deep network, GLP, image analysis, LFW, moisture content, NIR, NIR_MGI, non-destructive plant phenotyping, rice, VIS, wheat spikes identification and count.en-USnullThesis