Sourav Bhadra

Develop and deploy machine learning models that predict crop phenotypes for breeding R&D, working across remote-sensing imagery, climate grids, and sensor-based tabular data.

• Build end-to-end ML pipelines — from model development and validation to deployment and maintenance — using automated, cloud-native workflows.
• Prototype proof-of-concept solutions with cutting-edge sensors, the newest satellites, and modern model architectures to accelerate breeding cycles.
• Partner with cross-functional stakeholders to translate research and business objectives into deployable data science solutions.

• PyTorch
• TensorFlow
• Rasterio
• GDAL
• GeoPandas
• BigQuery
• AWS
• Airflow
• Docker

Led deep-learning research for digital agriculture, fusing multi-sensor UAV and satellite data to estimate crop traits and yield.

• Built a physics-informed transfer-learning framework (PROSAIL radiative-transfer model + deep neural networks), improving cross-environment generalization of crop-trait estimation by ~25%.
• Developed end-to-end 3D CNNs (ResNet/DenseNet) for plot-scale soybean yield prediction from multi-temporal UAV imagery, scaling training from a local GPU cluster to AWS SageMaker.
• Architected multimodal fusion CNNs integrating hyperspectral, thermal, and LiDAR data, outperforming single-sensor baselines for seed-composition estimation.
• Engineered a photogrammetric calibration pipeline (bundle block adjustment) achieving sub-centimeter multi-sensor co-registration.
• Published 4 first-author papers in top remote-sensing journals; presented at NAPPN, AGU, and AAG.

• Python
• PyTorch
• TensorFlow
• Rasterio
• GeoPandas
• AWS SageMaker
• Pix4D

Applied high-performance computing and deep learning to hydrologic connectivity and terrain analysis.

• Implemented hydrologic-connectivity models with TauDEM on an HPC cluster to quantify the impact of anthropogenic drainage structures.
• Trained a CNN to detect bridges and culverts from LiDAR-derived high-resolution DEMs.
• Automated stream-delineation pipelines from DEMs using ArcPy.

• Python
• ArcPy
• TensorFlow
• HPC
• ArcGIS

Built geospatial data-processing pipelines supporting large-scale flood-management and hydrological engineering projects in Bangladesh.

• Automated map-making and geospatial workflows with ArcPy, replacing manual processes for embankment and floodplain projects.
• Analyzed multi-temporal satellite imagery to detect riverbank change using eCognition, ERDAS Imagine, and Google Earth Engine.

• Python
• ArcPy
• Google Earth Engine
• eCognition
• ERDAS Imagine

Problem

Bringing satellite and drone imagery into a machine-learning workflow means hand-coding vegetation indices, stacking bands, and extracting zonal statistics — slow, error-prone boilerplate that demands deep geospatial expertise.

Solution

A Python library that automates the entire feature-extraction pipeline: stack rasters, compute 350+ vegetation indices, and extract statistics over shapefile geometries in a few lines of code, built on GDAL, Rasterio, and GeoPandas.

Impact

Adopted across the geospatial-ML community with 120+ GitHub stars and full ReadTheDocs documentation, lowering the barrier to remote-sensing ML for researchers and practitioners.

Problem

Travelers and leaf-peepers lack timely, location-specific forecasts of when fall foliage will peak — static seasonal calendars miss the spatial and year-to-year variation.

Solution

An interactive platform that predicts peak-foliage timing across the continental US from MODIS satellite time series using deep-learning transformers, served through a React/Leaflet map with a Gemini function-calling RAG chatbot for natural-language queries.

Impact

Turns satellite phenology science into a consumer-facing product — a full-stack showcase of generative AI, retrieval-augmented reasoning, and geospatial ML in one shipped application.

Problem

Over- and under-applying nitrogen costs farmers money and pollutes waterways; growers lack field-specific, in-season guidance on how much nitrogen to apply.

Solution

A decision-support web app that fuses PlanetScope satellite imagery with gridMET weather data to generate in-season nitrogen prescription maps for corn and sorghum at field scale.

Impact

Funded by a $15,000 USDA SARE grant and built with Illinois Corn Growers (ILCORN) and the Donald Danforth Plant Science Center to raise farmer profitability while cutting nitrogen loss.

Problem

Static maps can't show how a region changes over time — building animated geospatial timelapses from scratch takes substantial matplotlib and GeoPandas plumbing.

Solution

A Python library that turns a shapefile and a time-indexed dataset into animated choropleth or proportional-circle maps (GIF/MP4) with a single animate() call.

Impact

Packaged on PyPI with ReadTheDocs docs, giving data scientists a reusable one-liner for temporal map storytelling.

Problem

Understanding how US crop production shifts across counties and decades is hard to see in spreadsheets and static USDA reports.

Solution

A Streamlit web app that maps spatiotemporal trends for four major crops across US counties (1910–2021), combining USDA NASS statistics with Census TIGER boundaries and on-the-fly trend (slope) analysis.

Impact

A deployed, interactive tool that makes a century of agricultural data explorable for researchers and policymakers.

Problem

Estimating crop biophysical traits (leaf chlorophyll, leaf angle) from hyperspectral imagery with pure deep learning generalizes poorly across fields and seasons because labeled data is scarce.

Solution

A physics-informed, dual-stream neural network that embeds the PROSAIL radiative-transfer model into transfer learning, jointly estimating leaf chlorophyll and leaf angle from UAV hyperspectral images.

Impact

Improved cross-environment generalization of trait estimation by ~25%, published as first author in the ISPRS Journal of Photogrammetry & Remote Sensing.

Problem

Predicting soybean yield early and at plot scale requires capturing how the crop develops over the season, not just a single snapshot — a spatiotemporal problem classical features handle poorly.

Solution

An end-to-end 3D CNN (ResNet/DenseNet) trained on multi-temporal UAV RGB imagery, with the full training-and-inference pipeline scaled from a local GPU cluster onto AWS SageMaker.

Impact

Delivered accurate plot-scale yield prediction and demonstrated the value of temporal data over single-date imagery; published as first author in Precision Agriculture.