papers/zoo-species-classification/zoo-species-classification.tex

\documentclass[11pt,a4paper]{article}
\usepackage{shared/zoocover}

% ── Packages ──────────────────────────────────────────────────────────────────
\usepackage[utf8]{inputenc}
\usepackage[T1]{fontenc}
\usepackage{lmodern}
% geometry loaded by zoocover
\usepackage{amsmath,amssymb,amsfonts}
\usepackage{graphicx}
\usepackage{booktabs}
\usepackage{hyperref}
\usepackage{cleveref}
\usepackage{natbib}
\usepackage{algorithm}
\usepackage{algorithmic}
\usepackage{xcolor}
\usepackage{tikz}
\usetikzlibrary{arrows.meta,positioning,shapes.geometric}
\usepackage{float}
\usepackage{subcaption}
\usepackage{enumitem}
\usepackage{microtype}
\usepackage{fancyhdr}
\usepackage{abstract}
\usepackage{multirow}

% ── Page Style ────────────────────────────────────────────────────────────────
\pagestyle{fancy}
\fancyhf{}
\fancyhead[L]{\small Zoo Labs Foundation Research}
\fancyhead[R]{\small Few-Shot Visual Recognition}
\fancyfoot[C]{\thepage}
\renewcommand{\headrulewidth}{0.4pt}

% ── Metadata ──────────────────────────────────────────────────────────────────
\title{%
  \textbf{Few-Shot Visual Recognition for\\Biodiversity Monitoring}\\[0.5em]
  \large Zoo Labs Foundation Paper ZOO-CLS-2026
}

\author{
  Zoo Labs Foundation Research\\
  \texttt{[email protected]}\\[0.5em]
  \small With contributions from the Zoo Open Science Network
}

\date{March 2023}

% ── Macros ────────────────────────────────────────────────────────────────────
\newcommand{\R}{\mathbb{R}}
\newcommand{\E}{\mathbb{E}}
\newcommand{\Loss}{\mathcal{L}}
\newcommand{\Support}{\mathcal{S}}
\newcommand{\Query}{\mathcal{Q}}

\begin{document}

\zoocoverpage

% ── Abstract ──────────────────────────────────────────────────────────────────
\begin{abstract}
\noindent
Automated species identification is fundamental to scalable biodiversity
monitoring, yet most deep learning classifiers require hundreds or
thousands of labeled images per species---a prohibitive requirement for
the vast majority of Earth's estimated 8.7 million species. We present
\textbf{TaxoNet}, a few-shot visual recognition framework designed for
fine-grained species classification in ecological monitoring contexts.
TaxoNet combines a vision transformer backbone pre-trained on large-scale
ecological imagery with a taxonomy-aware meta-learning strategy that
exploits the hierarchical structure of biological classification (kingdom
through species). A novel \emph{taxonomic prototypical network}
constructs class prototypes at multiple taxonomic ranks and fuses them
through learned rank-attention weights, enabling robust identification
from as few as 1--5 reference images per species. We evaluate TaxoNet on
four biodiversity datasets spanning birds, insects, plants, and marine
organisms, achieving state-of-the-art 1-shot accuracy of 78.4\% and
5-shot accuracy of 91.2\% on held-out species. We further demonstrate
integration with citizen science platforms, where TaxoNet assists
non-expert observers in identifying species with 94.7\% accuracy when
combined with geographic and phenological priors. All models and datasets
are released through the Zoo Open Science Network.

\medskip
\noindent\textbf{Keywords:} few-shot learning, fine-grained
classification, biodiversity monitoring, taxonomic hierarchy, citizen
science, species identification
\end{abstract}

\newpage
\tableofcontents
\newpage

% ══════════════════════════════════════════════════════════════════════════════

\input{sections/01-introduction}
\input{sections/02-related-work}
\input{sections/03-problem-formulation}
\input{sections/04-taxonet-architecture}
\input{sections/05-domain-adaptive-pre-training}
\input{sections/06-hierarchical-meta-learning}
\input{sections/07-evaluation}
\input{sections/08-system-design-for-citizen-science}
\input{sections/09-discussion}
\input{sections/10-implementation-details}
\input{sections/11-future-work}
\input{sections/13-supplementary-per-taxon-analysis}
\input{sections/14-supplementary-prototype-visualization}
\input{sections/99-conclusion}

\end{document}