This thesis explores the development of lightweight neural network models for predicting perceptual soundscape attributes and environmental sound sources, with the goal of enabling real-time, sensor-based soundscape monitoring in urban environments. Recognizing the limitations of conventional noise metrics, the study adopts a perceptual approach that accounts for how individuals experience urban soundscapes. The work leverages the Affective Responses to Augmented Urban Soundscapes (ARAUS) dataset to replicate and extend the training pipeline of the baseline AD_CNN model. Four parameter-reduction strategies are implemented - shrinking convolutional and dense layers, reducing temporal resolution, and increasing pooling - to produce compact models suitable for low-cost, resource-constrained sensors. Two optimized variants, AD_CNN_dense_layer and AD_CNN_hop_length, are combined into an ensemble model with only ~160K parameters, which outperforms the original AD_CNN in perceptual attribute prediction. A generalisation study evaluates model robustness on a separate urban parks dataset, revealing that while the lightweight ensemble performs well overall, it struggles with certain attributes such as pleasantness. Larger models like SoundAQnet generalise better but remain imperfect. The findings demonstrate the technical feasibility of lightweight perceptual models but also highlight the need for localised fine-tuning to account for subjective and context-dependent factors shaping soundscape perception. This work contributes to scalable, context-aware soundscape monitoring solutions for healthier urban environments.