AI-driven spatial modeling, ensemble learning, and LLM-integrated decision systems for yard-scale pollinator plant recommendations.
This repository contains a public-facing code sample from one of several production-grade species distribution modeling (SDM) pipelines developed for the Pollinator Yards project.
Pollinator Yards is an end-to-end applied AI system that integrates spatial machine learning, large SDM ensembles, and large language models (LLMs) to generate parcel- and yard-scale native plant recommendations for pollinator conservation.
The complete system spans ecological modeling, data engineering, LLM-assisted knowledge extraction, and downstream web and mobile applications. This repository focuses specifically on the SDM layer, which produces high-resolution habitat suitability surfaces used in decision-making. This repository focuses specifically on the SDM component, which provides high-resolution habitat suitability estimates that drive parcel-level recommendations.
The full Pollinator Yards workflow consists of multiple coordinated pipelines:
-
Species Distribution Modeling (this repository)
- Ensemble SDMs trained for ~150 plant species as a proof of concept
- Focused on portions of Maryland and Pennsylvania
- Each species modeled using hundreds to thousands of fitted models under varying algorithms, seeds, and background configurations
- Outputs produced at high spatial resolution suitable for parcel-scale inference
-
Post-processing and Spatial Aggregation
- Aggregation of ensemble outputs across all models per species
- Derivation of suitability summaries and thresholds
- Transformation of gridded predictions into parcel-level suitability metrics
-
LLM-assisted Ecological and Botanical Knowledge Pipeline (separate repository)
- Automated collection of species traits, care requirements, phenology, and pollinator relationships
- Use of multiple LLMs with cross-validation to extract structured ecological and botanical knowledge from PDFs, databases, and web sources
- Image collection and verification using vision-capable models
-
Decision Layer and Applications
- Integration of SDM-derived suitability with LLM-derived species knowledge
- Generation of customized planting plans at the yard or parcel level
- Deployment via a web application and mobile app for end users
This repository represents one critical modeling layer within that broader system.
To make this repository suitable for public release and GitHub hosting:
- Large input data assets (TB-scale raster covariates, model outputs, and species occurrence datasets) are intentionally excluded.
- Environment-specific configuration, credentials, and large-scale orchestration logic have been removed or simplified.
- Companion pipelines (LLM workflows, parcel aggregation, and application backends) are not included.
As a result, the code here is not expected to run end-to-end as-is without the full data assets and supporting infrastructure from the original research environment.
This repository is published as a technical portfolio artifact, not as a standalone or turnkey SDM package.
The included R script(s) are intended to showcase:
-
Large-scale SDM experiment design
Multiple modeling algorithms, repeated background sampling, ensemble construction, and evaluation strategies. -
Reproducible workflow organization
Clear separation of inputs, intermediate artifacts, and outputs; deterministic ordering; explicit configuration blocks. -
Spatial data handling at scale
Raster stacking, masking, extraction, transformation, and aggregation using modern R spatial libraries. -
Ensemble post-processing
Aggregation and summarization of large numbers of fitted models per species to produce stable suitability estimates. -
Software engineering discipline
Defensive programming, explicit assertions, structured control flow, and extensive inline documentation.
In the full (non-public) version of the Pollinator Yards modeling framework, multiple layers of testing and validation are implemented, including:
- Data integrity and schema checks (coordinate validity, missing values, covariate alignment)
- Deterministic seeding and reproducibility controls
- Unit-style checks on intermediate artifacts (dimensions, value ranges, CRS consistency)
- Model evaluation diagnostics across individual models and ensembles
- Post-processing sanity checks on spatial predictions and parcel-level summaries
Elements of these practices are visible in the published code through explicit assertions, validation blocks, and defensive checks embedded throughout the workflow.
This repository focuses on spatial modeling and SDM experimentation.
Other components of the Pollinator Yards project—such as LLM-driven ecological and botanical knowledge extraction, image pipelines, and application-layer code—are maintained in separate repositories and are referenced here conceptually but not included.
Specifically:
- species-knowledge-base
Automated collection and validation of plant traits, care requirements, phenology, pollinator interactions, and imagery using multiple LLMs with cross-model validation.
Repository: https://github.com/sahilkharel7/species-knowledge-base
You are welcome to inspect and learn from the structure and patterns demonstrated here.
Meaningful reuse would require:
- Access to equivalent spatial covariates and occurrence data
- Adaptation of configuration and execution logic
- Integration with downstream aggregation and application layers
- Kumar Mainali — Project lead; SDM methodology, modeling pipeline design, large-scale experimentation, ensemble post-processing, and integration with downstream decision systems.
- Arjun Adhikari — Species occurrence data collection and curation; assistance with machine learning model development.
- Sahil Kharel — Contribution to LLM-based ecological, botanical and distributional knowledge extraction and validation workflows used within the Pollinator Yards system.