All projects

Tree Crown Detection

CNN and transformer models for estimating tree crown cover from PlanetScope satellite imagery.

CNNsTransformersComputer VisionRemote SensingPyTorch

Imagery

PlanetScope (~3m/px)

Models

CNN + Transformer

Lab

Duke Ecology

Overview

Tree crown cover is a key variable for carbon accounting, biodiversity monitoring, and land-use research — but field surveys cannot scale to whole regions. As an ML researcher in Duke’s Ecology Department, I build computer vision models that estimate crown cover directly from PlanetScope satellite imagery.

The work spans the full ML lifecycle: geospatial data engineering, model architecture comparison between convolutional and transformer-based approaches, and evaluation across heterogeneous terrain and canopy conditions.

Technical approach

  1. 01

    Geospatial preprocessing pipeline

    Raw PlanetScope scenes are tiled, co-registered with label data, normalized across acquisition dates, and filtered for cloud contamination before they ever reach a model.

  2. 02

    CNN baselines

    Convolutional encoder-decoder architectures establish strong baselines for per-pixel crown cover estimation, exploiting the local texture cues that distinguish canopy from ground.

  3. 03

    Transformer models

    Vision transformer variants capture longer-range spatial context — useful where canopy structure is ambiguous at the 3-meter resolution of PlanetScope imagery.

  4. 04

    Robustness-focused training

    Augmentation and sampling strategies are designed around generalization across terrain types, seasons, and canopy densities rather than leaderboard accuracy on a single region.

Architecture

  1. 01

    PlanetScope Imagery

    Multispectral scenes

  2. 02

    Preprocessing

    Tiling, normalization, QA

  3. 03

    CNN / Transformer

    Parallel model tracks

  4. 04

    Evaluation

    Cross-terrain validation

  5. 05

    Crown Cover Maps

    Per-pixel estimates

Challenges

Resolution limits

At ~3 meters per pixel, individual small crowns blur together. Models must estimate fractional cover from mixed pixels rather than detect crisp object boundaries.

Label noise

Reference data comes from higher-resolution sources with their own errors and registration offsets, so training has to tolerate imperfect supervision.

Domain shift

A model trained on one ecoregion degrades on another. Much of the engineering effort went into pipelines and augmentation that hold accuracy across varying terrain and canopy conditions.

Results

  • Working CNN and transformer crown cover models trained on large-scale PlanetScope datasets.

  • A reusable preprocessing and training pipeline that improved model robustness across terrain and canopy conditions.

  • Architecture comparison evidence informing the lab’s choice of models for region-scale mapping.

Media & links