Fig. 5: The diagram of model structures and interpretation methods. | npj Clean Water

Fig. 5: The diagram of model structures and interpretation methods.

From: Attention improvement for data-driven analyzing fluorescence excitation-emission matrix spectra via interpretable attention mechanism

Fig. 5

a The model structure of raw CNN classifier contained five convolutional layers (Conv1-Conv5), three max-pooling layers, two fully connected (FC) layers and one output layer. b Three convolutional block attention modules (CBAM) were embedded into the raw CNN classifier. Two types of input samples (i.e., 3D-EEM spectra with and without Rayleigh scattering) were utilized. c Gradient-weighted class activation mapping (Grad-CAM) method, guided Grad-CAM method, and structured attention graphs (SAGs) method were utilized to interpret the CNN classifier.

Back to article page