Projects
A collection of my work spanning modeling, computer vision, and research engineering.
Filter by Track
Filter by Tags
Showing 8 of 8 projects
GPU-Accelerated Synthetic Data Generation for Crop Phenotyping
Physics-based sensor simulation using C++ ray-tracing to generate perfectly labeled training data. Dual virtual sensors (LiDAR + multispectral camera) provide structural point clouds and pixel-perfect segmentation masks for bean-wheat intercropping ML research.
Mechanistic Modeling of AMPA Persistence Across Soil Types
Meta-analysis using TESFO mechanistic modeling to evaluate glyphosate metabolite (AMPA) behavior across four soil types. Identified critical mobility-persistence trade-off requiring site-specific agricultural risk management.
GrowController: IoT Environmental Monitor
ESP32-based CO2, temperature, and humidity monitor with VPD calculation and Prometheus/Grafana integration.
Sweet Pepper Segmentation & Classification
Multi-algorithm ML pipeline for pepper detection comparing LogReg, KNN, SVM, GMM, and Gradient Boosting approaches.
Rapid Soil Water & Nitrogen Prediction via NIR Spectroscopy and ML
Non-invasive dual-model ML pipeline delivering soil predictions in <1 minute vs. traditional 24-48 hour lab assays. Combines SVR for water content (R²=0.844, MAE=1.55%) and Random Forest for nitrogen classification (98.1% balanced accuracy). Breakthrough: PCA reduced features 90% while increasing accuracy from 88.1% to 98.1%. Production-ready with OOD detection and confidence thresholding. Edge-optimized at 1.8 MB.
Nitrogen Loss Risk Spatial Analysis
Nationwide spatial risk mapping of German organic farms using Local Moran's I and Getis-Ord Gi* hotspot analysis. Cleaned 79,125 farm records, implemented IDW and Ordinary Kriging interpolation, and proved statistically significant clustering (p<0.05) linking sandy soils to high-risk zones.
Multispectral UAV Crop Classification Pipeline
End-to-end ML pipeline for temporal crop classification using 10-band UAV orthophotos. Built rigorous spatial zone-based data splits to prevent leakage, engineered 220 spectral features, and solved overfitting through PCA dimensionality reduction—achieving 0.96 F1-score with just 9 components.
ESP32 WiFi CSI Pepper 3D Mapping
Experimental project using ESP32 Channel State Information (WiFi CSI) to attempt non-contact 3D mapping of bell peppers.