서브메뉴
검색
Biodiversity Monitoring at Scale with Foundation Models
Biodiversity Monitoring at Scale with Foundation Models
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20260202104737
- ISBN
- 9798290649597
- DDC
- 500
- 서명/저자
- Biodiversity Monitoring at Scale with Foundation Models
- 발행사항
- [Sl] : Stanford University, 2025
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2025
- 형태사항
- 237 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 87-01, Section: B.
- 주기사항
- Advisor: Exposito-Alonso, Moises;Noah, Noah.
- 학위논문주기
- Thesis (Ph.D.)--Stanford University, 2025.
- 초록/해제
- 요약Climate and land use change have radically altered biodiversity worldwide, spurring the need for new high-resolution methods to monitor biodiversity at a global scale. Meanwhile, the recent advent of foundation models have driven rapid advancements in the life and climate sciences. This dissertation develops foundation models towards understanding the world's biodiversity and informing real-world conservation decision-making across a heterogeneous world. To do so, first I introduce a new architecture, loss function, and pre-training dataset for large-scale species distribution modeling from citizen science observation labels paired with remote sensing and climate data. This approach shows significant improvements in detecting local biodiversity and its temporal change, and can be adapted to develop diverse ecological maps from limited examples. Next, I develop a new ecologically-inspired contrastive learning technique for model pre-training from heterogeneous biodiversity data (remote sensing and ground-level pictures) without the need for extensive expert labeling. This multimodal approach shows significant improvement over single-view and fully-supervised baselines across four diverse ecological tasks. Lastly, I demonstrate how biases underlying many large-scale biodiversity datasets can influence model behavior, such as how data from human-modified regions fail to improve rare species performance in wilderness areas. I further introduce a new framework for quantifying the predictive effect of these biases, and discuss their knock-on effects on downstream conservation tasks. This work showcases how foundation models can advance our understanding of the natural world and its myriad changes.
- 일반주제명
- Nature
- 일반주제명
- Vegetation
- 일반주제명
- Taxonomy
- 일반주제명
- Remote sensing
- 일반주제명
- Sociopolitical factors
- 일반주제명
- Biodiversity
- 기타저자
- Stanford University.
- 기본자료저록
- Dissertations Abstracts International. 87-01B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008260126s2025 us c eng d■001000017358684
■00520260202104737
■006m o d
■007cr#unu||||||||
■020 ▼a9798290649597
■035 ▼a(MiAaPQ)AAI32149660
■035 ▼a(MiAaPQ)Stanfordgz526zj7191
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a500
■1001 ▼aGillespie, Lauren Estelle.
■24510▼aBiodiversity Monitoring at Scale with Foundation Models
■260 ▼a[Sl]▼bStanford University▼c2025
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2025
■300 ▼a237 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 87-01, Section: B.
■500 ▼aAdvisor: Exposito-Alonso, Moises;Noah, Noah.
■5021 ▼aThesis (Ph.D.)--Stanford University, 2025.
■520 ▼aClimate and land use change have radically altered biodiversity worldwide, spurring the need for new high-resolution methods to monitor biodiversity at a global scale. Meanwhile, the recent advent of foundation models have driven rapid advancements in the life and climate sciences. This dissertation develops foundation models towards understanding the world's biodiversity and informing real-world conservation decision-making across a heterogeneous world. To do so, first I introduce a new architecture, loss function, and pre-training dataset for large-scale species distribution modeling from citizen science observation labels paired with remote sensing and climate data. This approach shows significant improvements in detecting local biodiversity and its temporal change, and can be adapted to develop diverse ecological maps from limited examples. Next, I develop a new ecologically-inspired contrastive learning technique for model pre-training from heterogeneous biodiversity data (remote sensing and ground-level pictures) without the need for extensive expert labeling. This multimodal approach shows significant improvement over single-view and fully-supervised baselines across four diverse ecological tasks. Lastly, I demonstrate how biases underlying many large-scale biodiversity datasets can influence model behavior, such as how data from human-modified regions fail to improve rare species performance in wilderness areas. I further introduce a new framework for quantifying the predictive effect of these biases, and discuss their knock-on effects on downstream conservation tasks. This work showcases how foundation models can advance our understanding of the natural world and its myriad changes.
■590 ▼aSchool code: 0212.
■650 4▼aNature
■650 4▼aVegetation
■650 4▼aTaxonomy
■650 4▼aRemote sensing
■650 4▼aSociopolitical factors
■650 4▼aBiodiversity
■690 ▼a0799
■71020▼aStanford University.
■7730 ▼tDissertations Abstracts International▼g87-01B.
■790 ▼a0212
■791 ▼aPh.D.
■792 ▼a2025
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17358684▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.
