Model architecture consisting of three parallel branches designed for multi-mode data integration: (1) a spectral branch (upper), which processes the photometric flux through 1D convolutional layers; (2) an image branch (middle), which analyses the galaxy image using 2D convolutions to extract morphological features; and (3) a defect-detection branch (lower), which processes localized image regions around the emission-line position to identify cosmetic artefacts. Each branch applies a combination of convolutional, pooling, and dense layers, with dropout regularization to prevent overfitting. The outputs from the active branches are concatenated and passed through a shared set of fully connected layers, followed by a final sigmoid activation that produces the output label. For the primary EELG classifier (P₀), only the spectral and image branches (upper and middle) are used. For the cosmetic-defect classifier (P₁), all three branches are active, including the additional defect-detection branch. A detailed description can be found in Appendix B.