The task centres on building and validating a deep-learning workflow that spots land-surface changes in time-separated satellite imagery, then visualises those changes and reports rigorous performance metrics. All modelling must be done in Python. I will supply paired, geo-referenced satellite scenes. The work starts with any preprocessing needed to align, normalise, and optionally augment this data. From there, the model—whether a UNet, CNN–LSTM, or another Keras-friendly architecture that you justify—should be trained to highlight pixel-level change. Please keep the code clear, reproducible, and GPU-ready. Visual output is essential: I need crisp change-detection masks overlaid on the original imagery plus colour-coded maps that make before/after differences intuitive to a non-technical audience. For evaluation, deliver the usual remote-sensing metrics (precision, recall, F1, IoU, and a confusion matrix). Include any additional indicators that reveal model bias or class imbalance. Deliverables • Clean, annotated Python notebook (or .py script) using Keras • Trained model weights and instructions to reload them • High-resolution visualisations of change maps in PNG and GeoTIFF • Brief report summarising methodology, hyperparameters, and all metrics The solution should run end-to-end on a standard workstation with TensorFlow/Keras installed; please note any external libraries or specific version requirements you introduce.