Geoinformatics Engineer | Applied Deep Learning to Earth Observation and Climate Systems
I bridge Environmental Engineering and Computer Science to build scalable Machine Learning models for Earth Observation. Currently pursuing an M.Sc. in Geoinformatics Engineering at Politecnico di Milano.
- Languages: Python, C, SQL, Matlab, R
- ML & DL: PyTorch, TensorFlow/Keras, Scikit-learn, TIMM, TerraTorch (PyTorch Lightning)
- Architecture: Transformer / ViT, CNNs, LSTMs, Encoder-Decoder, Foundational Model Fine-tuning, Ensemble Models
- Big Data & Systems: MongoDB, Cassandra, Elasticsearch, Neo4J, Redis, PostgreSQL, Flask, Dash
- Earth Observation: Google Earth Engine (GEE), GDAL, Rasterio, SNAP, QGIS, ArcGIS Pro, Copernicus/CAMS
- A complete PyTorch implementation of the Encoder-Decoder architecture (Vaswani et al.) without high-level libraries, featuring Multi-Head Attention, Sinusoidal Positional Encoding, and Pre-Layer Normalization.
- Replicated the original paper's optimization strategy using Mixed-Precision (FP16), Label Smoothing (ϵ=0.1), and a Learning Rate scheduler with Warmup and Inverse Square Root Decay.
- Integrated Beam Search decoding with length normalization and conducted rigorous quantitative evaluation using Corpus BLEU scores and Cross-Attention map visualizations.
- Link: https://github.com/astroedo/Transformer-nmt-en-it
- Technologies: PyTorch, Geospatial Foundation Models (TerraMind, AlphaEarth, Tessera, THOR), Mixed-Precision Training (AMP), Google Colab
- Details: Designed a multi-modal fusion network (SE-UNet, ASPP, cross-attention, dual bin-classification height heads) combining embeddings from 4 GFMs for joint land cover segmentation and nDSM height regression on 256×256 satellite patches.
- Results: Scored 0.4130 on the challenge leaderboard; diagnosed and fixed a silent FP16 overflow that blocked training, and built a crash-safe pipeline (RAM/SSD caching, auto-resume, TTA, multi-seed ensembling) for preemptible GPUs.
- Link: https://github.com/astroedo/ESA-GeoFM-Challenge
- Technologies: CNN, Random Forest, MLP, GEE, Google Colab
- Details: Developed and applied three ML models on 40 years of Landsat data (1,178 samples) for temporal glacier classification.
- Results: Achieved 99.1% accuracy classifying imbalanced geospatial data, comparing 1D-CNN, MLP, and Random Forest architectures.
- Link: https://github.com/astroedo/Rutor-Glacier-Melting.git
- Technologies: Python, QGIS, WebGIS
- Details: Processed 10+ years of pollution data, correlating
$\text{NO}_2/\text{PM}2.5/\text{PM}10$ with land cover/population - Results: Delivered an interactive WebGIS with bivariate mapping and zonal statistics for policy insights
- Link: https://astroedo.github.io/polimi-GIS2025/