Idea
A scalable deep learning platform mapping individual trees globally from satellite imagery for forestry, conservation, and environmental monitoring.
Research Paper
Core Innovation
This paper introduces a novel method modeling tree crowns as Gaussian kernels to detect individual trees from medium-resolution satellite imagery. It leverages massive airborne lidar datasets for training, enabling accurate tree detection both inside and outside forests. This approach surpasses existing tree cover maps in precision and scalability and can be fine-tuned with manual labels for enhanced performance.
Market Size (TAM)
$2–10B TAM, $1–2B SAM; assumption: global forestry, environmental monitoring, and urban planning require scalable tree mapping solutions.
Potential Customers & Pain Points
- Forestry companies needing precise tree inventories
- Environmental NGOs monitoring deforestation
- Governments tracking forest health
- Agricultural firms managing agroforestry
- Urban planners mapping green spaces
Business Model
Subscription-based API and platform access for tree mapping data and analytics with tiered pricing for different user needs.
Competitive Landscape
- Global Forest Watch
- Planet Labs
- Descartes Labs
Implementation Challenges
- Access to high-resolution satellite imagery at scale
- Integration with existing forestry and environmental data systems
- Adapting model to diverse ecosystems and tree species
Validation Strategy
- Pilot with forestry companies to validate tree detection accuracy
- Partner with environmental NGOs for real-world monitoring use cases
- Iterate model fine-tuning using manual labels from diverse regions
Research Paper Overview
Trees as Gaussians: Large-Scale Individual Tree Mapping
Summary
This paper presents a deep learning method to detect large individual trees globally using 3-m resolution PlanetScope imagery. It models tree crowns as Gaussian kernels to extract crown centers and generate binary tree cover maps. Training uses billions of points from airborne lidar data, enabling accurate tree identification inside and outside forests. The approach outperforms existing tree cover maps and can be fine-tuned with manual labels for improved detection, offering a scalable framework for high-resolution global tree monitoring adaptable to future satellite imagery.