PhenoEmbed Self-Supervised Multispectral UAV Time-Series Embeddings for Individual Tree Crown Phenology

Jul 16, 2026

Upload by Ezra

Tree crowns are unstable targets for resilient AI because their multispectral appearance, texture, translucency, and boundary cues vary across the growing season. PhenoEmbed introduces a crown-centric self-supervised temporal embedding model trained on contrastive and masked reconstruction objectives using HeideBench, an 18-date UAV multispectral time-series benchmark for forest crown phenology in Dölauer Heide. Seasonal crown dynamics are modeled as phenological appearance changes driven by leaf emergence, canopy closure, senescence, and leaf-off conditions. Retained crown polygons anchor aligned crops through time, enabling a 256-dimensional seasonal appearance vector per tree. Embeddings on 5,885 crop-safe crowns exhibit structured low-dimensional organization, and nearest-neighbor retrieval reaches a median top-1 cosine similarity of 0.946. Ablations show the contrastive loss, masked reconstruction loss, and explicit seasonal time features each shape the embedding space, supporting reusable forest crown representation learning.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages19

Reading Time48 min

Introduction

Motivation: seasonally dynamic tree crowns

Proposed direction: phenological embeddings for resilient analysis

PhenoEmbed (method overview)

Self-supervised temporal embeddings from multispectral UAV sequences

Objectives and training benchmark

arXiv :2607 . 10231v1 [ cs .CV] 11 Jul 2026 PhenoEmbed: Self-Supervised Multispectral UAVTime-Series Embeddings for Individual Tree Crown Phenology Taimur Khan 1 Abstract: Tree crowns are a challenging target for resilient AI because they are not static objects: their spectral response, internal texture, translucency, and apparent boundaries change substantially across the growing season. We develop PhenoEmbed, a self-supervised crown-centric temporal embedding model trained with contrastive and masked reconstruction objectives on HeideBench, an 18-date UAV multispectral time-series benchmark for forest crown phenology in Dölauer Heide. The model treats seasonal crown dynamics as phenological appearance change driven by leaf emergence, canopy closure, senescence, and leaf-off conditions. Segmented tree crown polygons are retained as object anchors to extract aligned crown-centered crops through time, allowing one 256-dimensional vector summarizing seasonal crown appearance to be learned per tree. On 5,885 crop-safe crowns, the exported embeddings show structured low-dimensional organization, with the first two principal components explaining 25.1% of variance and nearest-neighbor retrieval producing a median top-1 cosine similarity of 0.946 . Compared with handcrafted temporal features and a learned mean-pooling baseline, PhenoEmbed yields substantially more compact nearest-neighbor structure, while ablations show that the contrastive loss, masked reconstruction loss, and explicit seasonal time features each affect the structure of the learned embedding space. These results support PhenoEmbed as a reusable forest crown representation learner and motivate future downstream tests of whether such features improve tree-level models under seasonal change. Code: [https://github.com/Helmholtz-UFZ/PhenoEmbed](https://github.com/Helmholtz-UFZ/PhenoEmbed) Model: [https://doi.org/10.57967/hf/9558](https://doi.org/10.57967/hf/9558) Data: [https://doi.pangaea.de/10.1594/PANGAEA.993969](https://doi.pangaea.de/10.1594/PANGAEA.993969) Keywords: tree phenology, UAV remote sensing, multispectral, individual tree crowns, self-supervised learning 1 Introduction Tree monitoring models often assume that the object of interest has a reasonably stable visual signature. Individual tree crowns violate that assumption. Across a single growing season, the same crown can move from leaf emergence to canopy closure, from peak vigor to discoloration and senescence, and finally toward partially leaf-off conditions (Fig. 1) . Those transitions alter not only reflectance but also translucency, within-crown texture, shadowing, branch visibility, and even the apparent extent of the crown in an orthomosaic. In practice, this means that a crown segmentation or crown analysis model trained on one part of the year can become brittle when applied to another. 1 Helmholtz Centre for Environmental Research-UFZ, Community Ecology, Theodor-Lieser-Str. 4, 06120 Halle, Germany, [taimur.khan@ufz.de](taimur.khan@ufz.de), https://orcid.org/0000-0001-7833-5474 Fig. 1: Example crown-centered crop rendered at four points in the 2025 UAV time series. The same crown anchor is shown in early spring, green-up, late summer, and late autumn using synthetic RGB from the canonical 􀀧 , 􀀜, and 􀀧􀀚 bands. The panel illustrates why crowns are treated as temporally dynamic objects rather than static image patches. That brittleness is a concrete resilient AI problem in Earth observation. Season-varying remote sensing is well known to confound change analysis, because genuine structural change is entangled with phenology-induced spectral variation [Ko20] . At the same time, recent individual-tree studies show that UAV imagery is rich enough to resolve phenologyat the crown scale [Ka26; Kl24 ; KVK23 ; Pa19] . Embeddings are a way to compress large amounts of information into a smaller set of features that represent meaningful semantics. Therefore, the opportunity is clear:

FAQs

Why is individual tree crown analysis difficult across the growing season?

Crown appearance changes over time, including spectral response, internal texture, translucency, shadowing, and even apparent crown extent. Models trained on one seasonal state can therefore become brittle when applied to another.

What is PhenoEmbed and how is it trained?

PhenoEmbed is a crown-centric self-supervised temporal embedding model. It is trained with contrastive and masked reconstruction objectives on an 18-date UAV multispectral benchmark for forest crown phenology.

How are tree crowns represented for embeddings and what performance is reported?

Segmented crown polygons are kept as object anchors to extract aligned crown-centered crops across dates. For each tree, the model learns a 256-dimensional vector summarizing seasonal crown appearance, achieving a median top-1 cosine similarity of 0.946 in nearest-neighbor retrieval.

Jul 16, 2026

Upload by Ezra