The Convolutional neural network (CNN), a most established deep learning algorithm, is developed to estimate fresh almond yield with multi-spectral aerial images as inputs. CNN has a unique ability to automatically and adaptively learn spatial hierarchies of important features that summarize the presence of detected features in the input image for a particular predictive modeling problem (LeCun et al., 2015 (link)). The extreme efficiency in dimensionality reduction of the CNN model makes it unnecessary to conduct any feature extraction work, which increases computation efficiency and improves estimation accuracy. A surge of interest in CNN deep learning has emerged in recent years due to its superior performance in various fields (Lobell et al., 2015 (link); Yamashita et al., 2018 (link); Kattenborn et al., 2021 (link); Li et al., 2021 (link)).
A CNN is typically composed of a stacking of three types of layers, i.e., convolution, pooling, and fully connected layers (LeCun et al., 2015 (link)). The first two perform feature extraction, whereas the third maps the extracted features into final output, such as yield. As a fundamental component of the CNN architecture, a convolutional layer typically consists of a combination of linear and nonlinear operations, i.e., convolution operation and activation function. A convolution is a simple application of a spatial filter (or kernel) to an input image that results in an activation. Repeated application of the same filter to an input result in a map of activations called a feature map. A small grid of parameters called kernel, an optimizable feature extractor, is applied at each image position, which makes CNNs highly efficient for image processing. The kernel values are optimized during the model training process to extract features from input data based on the model’s task. The outputs of a linear operation such as convolution are then passed through a nonlinear activation function, e.g., the most commonly used rectified linear unit (ReLU). Batch normalization can also be applied as an optimization strategy to increase the model training efficiency, although it is not a solid requirement of the CNN model. To reduce the dimensionality of the extracted feature maps, a pooling layer provides a down-sampling operation by aggregating the adjacent values with a selected aggregation function, such as taking maximum value within the predefined window size. Similar to convolution operations, hyperparameters including filter size, stride, and padding are set in pooling operations. As one layer feeds its output into the next layer, extracted features can hierarchically and progressively become more complex.
To improve CNN model’s overall performance, the spatial attention module is recently introduced into the CNN architecture by combining a global average pooling layer and the following dense layers (Woo et al., 2018 ; Sun et al., 2022 (link); Zhang et al., 2022 (link)). Global average pooling layer is usually applied once to downscale the feature maps into 1-D array by averaging all the elements in each feature map, while retaining the depth of the feature maps. Dense layer then connects the final feature maps to the final output of the model with learnable weights via model training. The combination of a global average pooling layer and the following dense layers helps the CNN model focus more on the relevant features and thus improves.
A CNN is typically composed of a stacking of three types of layers, i.e., convolution, pooling, and fully connected layers (LeCun et al., 2015 (link)). The first two perform feature extraction, whereas the third maps the extracted features into final output, such as yield. As a fundamental component of the CNN architecture, a convolutional layer typically consists of a combination of linear and nonlinear operations, i.e., convolution operation and activation function. A convolution is a simple application of a spatial filter (or kernel) to an input image that results in an activation. Repeated application of the same filter to an input result in a map of activations called a feature map. A small grid of parameters called kernel, an optimizable feature extractor, is applied at each image position, which makes CNNs highly efficient for image processing. The kernel values are optimized during the model training process to extract features from input data based on the model’s task. The outputs of a linear operation such as convolution are then passed through a nonlinear activation function, e.g., the most commonly used rectified linear unit (ReLU). Batch normalization can also be applied as an optimization strategy to increase the model training efficiency, although it is not a solid requirement of the CNN model. To reduce the dimensionality of the extracted feature maps, a pooling layer provides a down-sampling operation by aggregating the adjacent values with a selected aggregation function, such as taking maximum value within the predefined window size. Similar to convolution operations, hyperparameters including filter size, stride, and padding are set in pooling operations. As one layer feeds its output into the next layer, extracted features can hierarchically and progressively become more complex.
To improve CNN model’s overall performance, the spatial attention module is recently introduced into the CNN architecture by combining a global average pooling layer and the following dense layers (Woo et al., 2018 ; Sun et al., 2022 (link); Zhang et al., 2022 (link)). Global average pooling layer is usually applied once to downscale the feature maps into 1-D array by averaging all the elements in each feature map, while retaining the depth of the feature maps. Dense layer then connects the final feature maps to the final output of the model with learnable weights via model training. The combination of a global average pooling layer and the following dense layers helps the CNN model focus more on the relevant features and thus improves.
Free full text: Click here
