الفهرس 
Chapter 1 IntroductionOnly 14 pages are availabe for public view
 |  | from | 157 |  |  |
| | | from | 157 | | |
Abstract
Currently, increasing water scarcity, one of the greatest challenges is negatively affecting crop production worldwide. It is convenient to apply modern remote sensing technologies with irrigation systems to ensure that scarce water resources are used effectively to irrigate fields. Also, crop water status must be observed in real time to accurately and quickly determine water needs. This inquiry establishes the groundwork for precision irrigation decision-making and management. Commonly, the use of ground-based sensors enabled the detection of plant water status that is not visible to the human eye. The oven drying method was a destructive tool, but it was the most accurate criteria for estimating the moisture content of plants. This study examined the water status of rice and wheat crops at different growth stages using robust and nondestructive methods such as hyperspectral, thermal and digital imaging. As well, several environmental parameters were measured using IoT-based sensors, including air temperature (°C), relative humidity (%), soil moisture (%), and wind speed (km/h). The entire data were processed and analyzed to adopt a superior classification and regression model for identifying the water status in rice and wheat crops. The main results were achieved as follows : 1. Prediction of canopy water content in rice by integrating feature selection approaches and regression algorithms based on hyperspectral data A canopy water content (CWC) assessment is the most important management decision for irrigation. Machine learning and hyperspectral imaging have combined to create a potentially valuable tool for precisely measuring the water content of plants. However, the tools are hindered by the selection of features and its advanced model. Thus, this work proposes an effective prediction model and compares three feature selection methods: model-based features (MF), vegetative indices (VI), and principal component analysis (PCA). Partial least square regression (PLSR), back-propagation neural network (BPNN), and random forest (RF) were used to train the samples with lowest loss on a cross-validation set using the specified features. The hyperspectral images were acquired on rice crops cultivated under various levels of water stress. To assess our suggested methodology, we used a total of 128 images. The findings showed that combining PCA and MF approaches may provide more significant feature selection for the prediction model. The bands of 1467, 1456, and 1106 nm were the three super-variables of CWC expectation. To improve foretelling accuracy, these above-mentioned characteristics were integrated with an optimized BPNN model. The advanced BPNN-PCA-MF model has an RMSE of 0.252% and an accuracy and correlation coefficient close to 1. At last, for researchers and policy-makers, this study supportively contributes to the prediction of plant water content so that effective and advanced steps are taken for precision irrigation. 2. Incorporating visible and thermal imaging for superior expectation of rice canopy water content via an artificial neural network Under different levels of water stress, a total of 120 rice plant samples were scanned by visible and thermal proximity sensing systems to assess the canopy water content (CWC). The oven-drying manner was used to determine the canopy’s water content. This CWC is immensely important for irrigation management decisions. The designed scheme employs an artificial neural network to incorporate visible and thermal imaging data as a useful prospective tool for precisely measuring the plant’s water content. The RGB-based characteristics comprised 6 gray level co-occurrence matrix-based texture features (GLCMF) and 20 color vegetation indices (VI). Thermal imaging characteristics included two thermal indicators (T), crop water stress index (CWSI) and normalized relative canopy temperature (NRCT), which were deduced from plant temperatures. These features were applied in conjunction with a back-propagation neural network (BPNN) to train the samples on a cross-validation set with minimum loss. Filtering high-level features and adjusting the model’s hyperparameters had an effect on the model’s behavior. The outcomes exhibited that feature-based modelling from both thermal and visible images performed better than features from a single thermal or visible image. The superior predictive features were 21 variables: 5GLCMF, 14VI, and 2T. The combination of color–texture–thermal characteristics significantly enhanced the precision of plant water content estimation (99.40%). It had the highest determination coefficient (R2=0.983) and the lowest RMSE of 0.599%. In summary, this work’s methodology can assist water managers and decision-makers in taking effective and timely decisions to ensure agricultural water sustainability. 3. Using IoT-based multimodal data and hybrid deep network to diagnosis water status in wheat crops Automatic detection of plant water status is a significant challenge in agriculture as it is a crucial regulator of growth, productivity, quality, and sustainability. As a result, accurate monitoring of the plant’s water condition has become imperative. Internet of Things (IoT) solutions based on specific sensor data acquisition and intelligent processing can assist water users for precise irrigation by providing accurate, consistent, and fast results. This paper aims to present a hybrid deep learning approach based on the combination of a convolutional neural network (CNN) and long short-term memory (LSTM) for automatically identifying the water state of wheat. The intended scheme used IoT-based data transmission devices such as a digital camera, soil moisture, wind speed, relative humidity, and air temperature. These environmental factors (EF) were recorded during the plant image capture. A total of 876 images of wheat plants were collected under different water deficit levels. A data augmentation technique was employed to expand the size of the training dataset to 5256 images. Various types of image color modes for example CMYK (cyan-magenta-yellow-black), HSV (hue-saturation-value), RGB (red-green-blue), and grayscale were evaluated with our proposed methods. The experimental results indicated that the combined CNNRGB-LSTMEF-CNNEF deep network based on features from both RGB images, climatic conditions, and soil moisture performed better than features from individual RGB images. Its outputs of validation accuracy, classification precision, recall, F-measure, and intersection over union are 100% with a loss of 0.0012. The proposed system behavior is very encouraging to develop our methodology with other crops in the future. The designed framework can serve the agricultural community to detect the water stress of plants before the critical level of growth and make timely management decisions. 4. A prominent expectation of wheat irrigation frequency and water demand in IoT environment using multimodal data-driven pre-trained deep networks Investigation of the plant’s water demand is a critical restraint to accomplish attractive agricultural productivity during the growing phase. The use of IoT-based multimodal data technology in agricultural proximity sensing applications to monitor crop water demands and manage site-specific images has sparked considerable attention. In addition, transfer learning has recently proven the capacity to transfer the learned feature detectors of a pre-trained convolutional neural network to a fresh dataset. This research revealed a potentially intelligent approach related to pre-trained deep networks like VGG16, VGG19, ResNet50, ResNet101, and MobileNet to extract distinctive features and track wheat irrigation frequency and water needs. The proposed scheme employed two deep learning networks such as convolutional neural network (CNN) and deep neural network (DNN) to train multiple features retrieved from pre-trained deep networks. The dataset was gathered by IoT-based data transmission devices, for example a digital camera, soil moisture, air temperature, wind speed, and relative humidity. While the plant image was captured, environmental factors (EF) were recorded. Extensive use of data augmentation with transfer learning has also been applied, the dataset of 876 samples was expanded to 5,256 samples. The results indicated that the fusion of VGG16–EF features with CNN greatly improved the expectation precision of irrigation frequency and plant water status (96.2% for validation). These features outperformed other transfer learning features with DNN in this study. Moreover, the hybrid model of CNNVGG19, CNNEF, and DNNEF obtained the highest validation performance (97.9%), whereas the outputs for precision, recall, F-measure, and intersection over union were 98%, 97.9%, 97.9%, 95.9%, respectively. The proposed framework outlines a roadmap toward the automated detection of irrigation frequency and water status during the plant’s life cycle. In the near future, the composite methodology will play a critical role in analyzing the growth traits of crops for precision cultivation and agricultural irrigation management.