Geforce gtx 1660 ti gpu

Two different approaches (1) feature extraction-based (DL models: AlexNet, GoogLeNet, Inception V3, MobileNet V2, and ResNet50) and (2) function approximation-based ML models (Artificial neural network (ANN), K-nearest neighbors (KNN), Support vector machine (SVM), and Logistic regression (LR)); and a DL model (DL-LSTM) were adopted for crop water stress classification. Feature extraction-based models were trained on thermal as well as RGB imagery. Function approximation-based models were trained on ambient weather and soil parameters, and T_c inputs from thermal imagery.
Deep CNNs typically have complex architecture and some may require significant computational resources. All CNN model training and validation processes were performed on a desktop computer (Intel Core I7 Processor with base frequency 2.60 GHz, 16 GB RAM, 6 GB NVIDIA GeForce GTX 1660 Ti GPU) with Windows 10 operating system (64 bits). CNN models were developed in MATLAB 2019b using the deep learning and machine learning toolbox. All the models are detailed in the following sub-sections.

As shown in Fig. 1 (green rectangle), the input dataset for the U-Net model training comprised spatiotemporal slice images ( $T\times Y$ ) obtained from the raw data and the corresponding GTs. Network training was conducted using 6144 augmented slice images with dimensions of $16\times 256$ , which were obtained from six sets of raw data. Spatiotemporal slice images from the other four sets of raw data were used for the blood vessel prediction. The training process lasted 50 epochs; 10% of the training data were selected at random for periodic training. In addition, binary cross-entropy was used as the loss function, and the stochastic gradient descent with momentum optimization algorithm was employed. We ran the U-Net model on a PC with an Intel i7-7700 CPU, NVidia GeForce GTX 1660Ti GPU, and 16 GB RAM.
To obtain the ST-AFI images, the trained model was used to predict spatiotemporal slice images not used for training. As shown in Fig. 1 (blue rectangle), $X$ -frame spatiotemporal slice images from a new set of raw data were input to the trained U-Net model to obtain $X$ -frame prediction probability maps ( $T\times Y$ ) of the blood flows; $P_{\mathrm{RBC}}(x,y,t)$ is the prediction probability. The reconstructed ST-AFI image can be expressed as $$P(x,y)=\mathrm{avr}{[P_{\mathrm{RBC}}(x,y,t)]}_{t\in [i,j]},$$ where $\mathrm{avr}{[]}_{t\in [i.j]}$ represents the averaging of $P_{\mathrm{RBC}}(x,y)$ in the region $t\in [i,j]$ .

Four DL datasets were trained separately using independent networks in the environment of TensorFlow (v2.0.0rc, Python3.7) on an Nvidia GeForce GTX 1660 Ti GPU with a batch size of 1 and epoch of 1,000. In the training phase, we stored the optimal weight configuration by optimizing the loss function to the minimum value, which resulted in four trained networks for the DL prediction at the testing stage. For the testing phase, the hemodynamic results of fluid points in P₂ were predicted by only importing the spatial coordinated information of the cavity point cloud in P₁ and the spatial coordinate information of the fluid point cloud in P₂ using the stored optimal configuration.