Figure 9
The diagram illustrates the structure of a convolutional neural network from left to right. On the far left, a yellow block labeled “input 2 by 500 by 500” represents a two-channel input image of size 500 by 500. An arrow leads to the first convolutional stage labeled “Convolutional layer 1, Block pooling”, shown as a stack of green feature maps labeled “256 filter”. This output passes through a R e L U activation and is annotated as “256 by 5 by 5”. The network then enters a grouped section labeled “Convolutional layers 2–6”. Within this section, multiple stacks of purple feature maps are shown, labeled sequentially with decreasing numbers of filters: “128 filter”, “64 filter”, “32 filter”, “16 filter”, and “8 filter”. A small orange square labeled “3 by 3 kernel” indicates the kernel size used in these convolutional layers. The final convolutional output is annotated as “8 by 5 by 5” followed by “R e L U”. After the convolutional stages, the feature maps are flattened, indicated by the label “Flatten”, and passed into a fully connected neural network arranged from left to right. On the left, a large group of circular nodes represents the first hidden layer, labeled with “R e L U” above it and annotated below as having “512 neurons”. Each node in this layer is connected by lines to every node in the next layer. The second hidden layer is shown to the right, also labeled “R e L U”, and is annotated below as having “256 neurons”. Dense connection lines link all neurons from the previous layer to this layer. The third hidden layer follows, again labeled “R e L U”, with “64 neurons” indicated below. As before, all neurons are fully connected to the next stage. On the far right, a small output layer contains two circular nodes, labeled “output 2 neurons”. Lines connect every neuron in the final hidden layer to both output neurons. Dotted vertical ellipses beneath each hidden layer indicate that the illustrated nodes represent larger groups of neurons rather than every individual unit being drawn.Architecture of the lightweight CNN model used for predicting local property ratios