Advancing Deep Ensembles and Knowledge Distillation for Efficient Local Climate Zone Mapping on Axelera Metis

Urban areas are growing rapidly worldwide, making it important to monitor how cities affect local climate. This thesis presents a complete system for classifying Local Climate Zones (LCZ) from satellite images, addressing both high accuracy and the ability to run on small, low-power devices. The work uses the So2Sat LCZ42 dataset, which contains over 400,000 paired radar and optical satellite images from 42 cities around the world. This research makes several important contributions to solve the challenge of urban climate monitoring. First, the system builds a data processing pipeline that efficiently converts raw satellite files into organized, ready-to-use training datasets. It then refines the satellite snapshots using super‑resolution methods sharpening the visuals and and capturing the details of urban structures. Additionally the framework pulls motifs, from the imagery via data analysis surfacing shape insights, about city layouts that standard methods often miss. Second, the classification system uses ensemble learning, where multiple neural networks combine their predictions to improve accuracy. The scheme trains models, DenseNet‑201 on one side, ResNet‑18 on the other, using blends of satellite inputs: a set runs on pure optical bands, another, mixes the two. Each group contains ten models trained on different channel selections, creating diversity that improves overall performance. The system fuses the models together yielding confidence scores that can be trusted alongside each prediction. Third, the framework addresses the challenge of deploying large models on resource-limited devices. Through knowledge distillation, the system transfers knowledge from large ensemble teachers into smaller student networks, maintaining high accuracy while significantly reducing model size. Quantization then converts the model from 32-bit floating-point to 8-bit integers, further reducing memory and increasing speed. Finally, the compressed models run on the Axelera Metis edge AI accelerator, delivering real-time results with low power consumption—exactly what continuous monitoring scenarios require. The pipeline combines MATLAB for data processing and a first attempt of proof, with Python for model training all the final arrangement, all running on high-performance computing clusters where resource usage is carefully managed. Reproducibility gets billing with a suite of documentation and modular code that anyone can pick up and expand. This research shows that cutting‑edge machine‑learning techniques can actually be brought into real‑world urban monitoring systems. By building an end‑to‑end pipeline that starts with raw satellite imagery and ends with models running on edge devices the study illustrates how to keep accuracy high while staying within the limits of speed, power consumption and memory. The resulting framework lays a foundation, for urban climate monitoring, a crucial element, for sustainable city planning and climate‑change adaptation.