Contextualizing CNN attribution maps using predefined image-derived descriptors in medicinal plant seed classification

Deep learning classifiers can achieve high predictive performance in medicinal plant seed image classification, but prediction accuracy alone does not clarify which visual properties are linked to model responses. Here, we analyzed RGB images from five medicinal plant seed classes by comparing convolutional neural network (CNN) attribution maps with predefined image-derived descriptor maps within the seed foreground. ResNet50, EfficientNet-B0, and ConvNeXt-Small showed high classification performance under controlled imaging and preprocessing conditions. However, species-associated differences could not be disentangled from unrecorded acquisition-session effects. In ConvNeXt-Small, zero baseline absolute Integrated Gradients (IG) showed the largest associations with LAB L, Brightness, and fast Fourier transform (FFT) low pass, which formed a highly redundant intensity and lightness related cluster. Removing the explicit input baseline magnitude term reduced these associations by 46.3% to 49.2%, and additional baseline, signed IG, and Grad-CAM analyses showed variation in descriptor associations across attribution representations. Descriptor summary feature classification further showed that foreground level descriptor statistics contained class discriminative visual information under the controlled imaging conditions. These results support the use of predefined descriptors for contextualizing attribution patterns within image derived visual contexts, with interpretation focused on method dependent spatial associations.