Multispectral remote sensing devices have the potential for large scale integration and can be successfully used in environmental monitoring; in particular optical sensors could embed microelectronic circuitry in order to perform signal preconditioning and more complex classification tasks for specialized applications. We propose a neural approach for supervised classification of the spectral signature of electrically sensed optical emissions, since the neural networks can be trained for the correct classification without any predetermined logic or programming burden. The software and hardware implementation of a such smart optical sensor for the spectral signature recognition is discussed and the classification results for fire detection are presented.