Jose
Escobar

Standalone Python CLI that ingests any US DEM GeoTIFF and outputs an animated flood simulation, PDF drainage report, GIS-ready basin data, and per-building flood exposure — no manual GIS work. Pipeline: NHD/SSURGO/NOAA Atlas 14 data acquisition, stream-burned terrain conditioning, D8 flow routing, drainage-area-threshold sub-basin delineation, Green-Ampt infiltration (real SSURGO soil parameters, cross-validated against SCS Curve Number), Muskingum-Cunge kinematic wave routing through DEM-derived compound channels, HAND-based flood mapping. Found that repeated runs on an identical DEM produced different basin summaries; localized the fault by hashing every pipeline intermediate across runs to the terrain-conditioning stage. Root cause: two WhiteboxTools operations (depression breaching, flat resolution) break ties non-deterministically on flat terrain, shifting drainage divides between identical runs. Replaced conditioning with deterministic priority-flood filling (pysheds) and added a permanent reproducibility gate — two runs, byte-identical basin summaries, or the test fails. The fix exposed a second design flaw: the stream-delineation threshold was a fraction of the DEM's own peak flow accumulation, making basin count hostage to whatever the largest stream happened to carry. Redesigned it as an absolute drainage-area parameter (km², user-configurable), independent of conditioning method or DEM extent. Validated generalization on a second, physiographically opposite region — Arkansas Ozarks vs. coastal Houston — same command, zero code changes: the engine correctly resolved a region-specific 8.4in design storm (vs. 17.7in for the Gulf Coast) and the opposite drainage topology, one nested 33 sq mi watershed instead of 163 independent coastal catchments.

Standalone Python CLI engine that ingests any post-event airborne LiDAR point cloud (.laz), computes four geometric indicators per building footprint, ranks structures by anomaly score, and outputs a self-contained interactive HTML platform and an optimized GPX inspection route — with no internet connection required at runtime. Engine indicators: Vertical Distribution Entropy (40%) measuring debris-filled structural void collapse signature; Point Density Irregularity (30%) measuring chaotic debris scatter patterns; Roof Planarity RMSE (20%) measuring deformed roof geometry via least-squares plane fitting; Point Count Deficit (10%) measuring return shortfall vs expected intact-roof density. Percentile-based tier assignment — Critical, High, Moderate, Low — answers which buildings are most anomalous relative to neighbors in the scan. GPX inspection route uses nearest-neighbor optimization within each tier: Critical buildings first, minimum travel distance. Validated on two independent post-disaster datasets: Haiti 2010 M7.0 earthquake (17.9M points, EPSG:32618, informal concrete block construction) and Kumamoto 2016 M7.0 earthquake (8M points, EPSG:2444, Japanese wood frame construction). Same codebase, different countries, different CRS, different construction typologies. Published as open-source repository with research-level README documenting physics reasoning, known limitations, and full CLI reference.

Live agricultural monitoring platform for the Mekong Delta — one of the most intensively farmed river systems on earth. Exported 5 Sentinel-2 SR Harmonized composites from Google Earth Engine (Nov 2022–Apr 2023) with SCL cloud masking and fallback composite: True Color, NDVI, NDWI, NDRE, False Color NIR. Converted to Cloud Optimized GeoTIFF via GDAL, then to 23,245 XYZ tiles (zoom 7–13, JPEG) using gdal2tiles — served from GitHub Pages for instant band switching with no download wait. COG files hosted on AWS S3 (us-east-1). PostGIS schema on Supabase: 499 agricultural parcels across 13 Vietnamese provinces + Kandal and Prey Veng (Cambodia), 4,981 spectral observations across 4 rice seasons. Key finding: rice paddy flood cycle confirmed in NDWI data — December peak (0.464), May (0.324), September (0.185) matching the real Mekong three-crop calendar. Platform features: 5 Sentinel-2 band toggles, live parcel data from Supabase REST API, temporal NDVI/NDWI line chart per parcel click, land use and country filters, flood detection overlay, band color legends, click-to-detail sidebar.

Production-grade geospatial intelligence platform tracking real Amazon deforestation. Exported 1,121 real forest loss pixels from Google Earth Engine (Hansen GFC v1.11) across Para, Amazonas, Mato Grosso and Rondônia (2001–2023). Loss drivers derived from MODIS MCD12Q1 post-loss land cover — published proxy methodology used by INPE and Global Forest Watch, not synthetic classification: Forest Degradation (569), Cattle Ranching (541), Fire (5), Agriculture (4), Mining (2). Protected area boundaries from WDPA v2024 (370 registered areas: indigenous territories, national parks, extractive reserves). Schema: PostgreSQL 17 + PostGIS 3.6 with GIST spatial indexes, migrated to Supabase for live hosting. Key finding: 66 confirmed deforestation violations inside registered protected areas detected via ST_Within() spatial intersection. Deployed as a live interactive web map — Leaflet.js frontend fetching real PostGIS data from Supabase REST API. Filter by year, driver, violations. Click any point for full event details. Every data point verifiable against original public sources.

Fine-tuned YOLOv8n on the Solar Panel Fault Detection dataset (2,763 RGB images, 7 fault classes: Bird Drop, Defective, Dust, Dusty, Non Defective, Physical Damage, Snow). Transfer learning from COCO-pretrained weights on Tesla T4 GPU via Google Colab — 50 epochs, batch 16, imgsz 640. Overall mAP50: 0.321. Strongest class: Defective (0.640 mAP50). Weakest: Physical Damage (0.097) and Bird Drop (0.149) — attributed to high intra-class visual variability and limited per-class sample count. Training curves confirm clean convergence with no overfitting. Inference demonstrated on held-out test set with per-detection confidence scores. This sprint establishes the object localization layer missing from Sprint 05B — detecting and bounding individual panels in RGB imagery, complementing the thermal severity scoring pipeline.

Queried 462 real Sentinel-2 L2A images over Jalisco, México via Google Earth Engine Python API. Built a median composite from scenes with less than 10% cloud cover across 6 spectral bands (Blue, Green, Red, NIR, SWIR1, SWIR2). Defined 5 land cover classes — Water, Forest, Agriculture, Urban, Bare Soil — with 25 georeferenced training points over known locations (Lago de Chapala, Guadalajara, Sierra Madre, Los Altos, northern dry zones). Trained a Random Forest classifier (100 trees) directly on GEE infrastructure. Classification map confirms correct spatial distribution: Lago de Chapala isolated as water, Guadalajara metropolitan area as urban, Sierra Madre footprint as forest. Exported classified GeoTIFF and RGB composite to Google Drive. Interactive map rendered via geemap/ipyleaflet.

Processed a real classified LiDAR point cloud (2,515,216 points, EPSG:26912) sourced from USGS / OpenTopography. Separated ground (836,431 pts) from non-ground returns, then rasterized both to 1m/px grids via binned statistics: DTM (mean ground elevation) and DSM (max surface elevation). Filled DTM occlusion gaps beneath the structure footprint via nearest-neighbor interpolation, then computed a normalized height model (DSM − DTM) to isolate above-ground structures independent of terrain slope. Dual-threshold structure detection at 2m (vegetation + structures, 3.08 ha) and 10m (building-scale only, 1.64 ha) — capturing the Capitol dome at 75.9m height above local terrain. Exported DTM, DSM, and height-model as georeferenced GeoTIFFs. Same DSM/DTM-differencing method transfers directly to mining stockpile volumetrics and transmission-corridor clearance analysis.

Extended Sprint 05A toward a decision-ready triage tool. Built a per-panel severity scoring formula combining four signals: temperature delta above healthy baseline, hotspot area percentage, thermal gradient irregularity, and a domain-reasoned defect-class risk weight — output is a 0–100 severity score per image, bucketed into Critical / Moderate / Monitor / Healthy tiers. Applied across all 20,000 images: 403 Critical, 3,700 Moderate, 5,897 Monitor, 10,000 Healthy. Exported results to JSON and built a static interactive dashboard — ironbow thermal colorization, live filtering by tier and defect class, paginated card grid, detail modal with full stats on click, and CSV export of any active filter. Honest scope: severity classification demonstrated on a public benchmark dataset (InfraredSolarModules · ICLR 2020). Spatial/array-position layer pending real client-provided GPS or flight metadata — architecture for this is planned in the PostGIS sprint.

Ingested the InfraredSolarModules dataset (ICLR 2020) — 20,000 real thermal drone images of solar panels across 12 defect classes. Built a full inspection pipeline in Python: loaded and parsed JSON metadata, extracted thermal statistics (mean intensity, std deviation, thermal gradient) per image, and computed class-level profiles across all 20,000 samples. Hotspot detection implemented via 85th-percentile threshold with red overlay mask. Key findings: Hot-Spot-Multi ranked highest mean intensity (183.6 vs 150.9 baseline, +32.7 delta); Cracking scored highest thermal gradient (8.61) — most irregular heat distribution, consistent with structural fracture signature. Generated 4 publication-quality outputs: class distribution chart, 12-class sample grid, hotspot detection panels, and final dark-mode inspection map. Full pipeline documented via Jupyter notebook.

Connected to Microsoft Planetary Computer API (pystac-client + stackstac) and queried Sentinel-2 L2A imagery over an agricultural region in Jalisco, Mexico. Selected the cleanest scene available (0.002% cloud cover, 2024-05-26). Stacked 5 spectral bands (B03, B04, B05, B08, B11) at 10m/px across ~245 km² in EPSG:32614. Calculated four vegetation indices from real satellite data: NDVI (mean 0.186, max 0.837), NDWI water stress (mean −0.254), NDRE chlorophyll via red edge band (mean 0.117, max 0.761), and SAVI soil-adjusted (mean 0.279, L=0.5). 5.2% of valid pixels classified as healthy vegetation (NDVI > 0.4). Generated a 5-panel publication-quality map with false color composite (NIR/Red/Green). Full pipeline documented via Jupyter notebook.

Processed a real 77-image aerial dataset (Valley Parkway Connector, Ohio) through a full photogrammetry pipeline using WebODM + OpenDroneMap. Generated a georeferenced orthomosaic (EPSG:32617, 0.08 m/px, 4457×3355 px) covering 4.64 ha. Extracted GPS coordinates from EXIF metadata, plotted image capture positions and altitude profile. Computed NDVI from RGB+NIR bands via rasterio: mean NDVI 0.382, 73.4% healthy vegetation. All outputs documented via pyulog + Jupyter pipeline.

Executed a battery failsafe validation test in PX4 SITL / jMAVSim. Vehicle was running an autonomous mission when battery failure was injected at t=33.2s via failure battery off command. Flight controller triggered a 5-second loiter hold (nav_state=4) before transitioning to Auto Land (nav_state=18). Vehicle executed a controlled descent and landed at t=65.3s. Battery drained from 100% to 0%. Full nav_state timeline and battery drain extracted from log_0_2026-6-11-14-25-29.ulg and documented via pyulog + Jupyter pipeline.

Executed a GPS loss failsafe validation test in PX4 SITL / jMAVSim. Vehicle was running an autonomous mission when GPS failure was injected at t=30.0s via failure gps off command. EKF immediately lost valid position estimate, switching to fake_pos dead reckoning. Flight controller transitioned from nav_state=3 (Auto Mission) to nav_state=12 (Descend / Blind Land) with zero reaction lag. Vehicle executed a controlled descent and landed at t=72.2s. Full nav_state timeline and EKF behavior extracted from log_0_2026-6-11-13-47-44.ulg and documented via pyulog + Jupyter pipeline.

Executed a failsafe validation test in PX4 SITL / jMAVSim. Vehicle was running an autonomous mission (42 waypoints) when a simulated signal loss was introduced at t=268.5s. The flight controller correctly detected the loss, transitioned from nav_state=3 (Auto Mission) to nav_state=5 (RTL), and executed an autonomous return-to-launch without manual input. Signal restored at t=332.9s. Total flight time: 5.6 min. Failsafe trigger, RTL trajectory, and nav_state transitions extracted from log_0_2026-6-6-22-21-20.ulg and documented via pyulog + Jupyter pipeline.

Executed a full grid survey mission in PX4 SITL / jMAVSim. Mission: GRID_01 — autonomous photogrammetric mapping pattern at 80m AGL. Total distance covered: 1,115.2m. Max speed: ~5.0 m/s. Upon completing the final waypoint, the vehicle entered Safe Recovery autonomously, returned to Home, and executed a precision vertical landing without manual input. No GPS or compass errors logged. Telemetry extracted from log_0_2026-6-7-21-00-04.ulg and processed via pyulog + Jupyter pipeline.

Executed a dual-pass survey mission in PX4 SITL / jMAVSim — two overlapping transect patterns over the same area, simulating high-density photogrammetric coverage for infrastructure inspection. Total flight time: 12.6 min. Mission items: 58. Altitude held stable throughout both passes. Telemetry extracted from log_5_2026-6-5-18-26-47.ulg and processed via pyulog + Jupyter pipeline.

Installed and configured PX4 SITL from source on macOS arm64. Resolved Python dependency conflicts (kconfiglib, jsonschema, empy, lxml) inside an isolated venv. Fixed jMAVSim Java compilation error by installing OpenJDK 17 via Homebrew and forcing a clean rebuild. Connected QGroundControl via UDP port 14550. Validated the full stack with arm → takeoff → hover → land sequence. Log file generated and saved as evidence (.ulg).