Multimodal data is emerging from different sources and in large quantities, which has allowed researchers to train highly-performing intelligent models and agents. However, the computational and environmental costs of the current deep learning trends are against the sustainability goals set worldwide, including the environmental concerns about its carbon footprint. Spiking neural networks offer a bio-plausible alternative given their low computational needs, hence less carbon emissions. In this work, we analyze the performance of different spiking neurons, propose a new spiking layer implementation trained using surrogate gradients, and test against different feature extraction scenarios in fully connected spiking neural networks. We also introduce a new metric to compare in-model and cross-model for better decisions when designing and training spiking neural networks.