Passive Acoustic Monitoring provides an important tool for wildlife monitoring. Deep Learning and the use of convolutional neural networks have become the common approach to create bioacoustic classifiers. However, in order to perform these tasks in real-time, standard approaches suffer from high model complexity and computationally expensive pre-processing steps (such as downsampling). In this paper we introduce a genetic algorithm named ESO, designed to optimise the input spectrogram size by focusing on specific regions. This approach allows for significant reduction in model complexity (91%) and inference time (70%) with minimal impact on the F1-Score (4%). We furthermore develop a simple-to-use Graphical User Interface and Python package to run the algorithm.