東華大學圖書館 |

語系: 繁體中文

說明(常見問題)

回圖書館首頁

手機版館藏查詢

登入

回首頁

切換: 標籤 | MARC模式 | ISBD

FindBook

Google Book

Amazon

博客來

Machine Learning for Satellite Imagery when Labels Are Scarce.

紀錄類型:	書目-電子資源 : Monograph/item
正題名/作者:	Machine Learning for Satellite Imagery when Labels Are Scarce./
作者:	Wang, Sherrie.
出版者:	Ann Arbor : ProQuest Dissertations & Theses, : 2021,
面頁冊數:	209 p.
附註:	Source: Dissertations Abstracts International, Volume: 83-05, Section: B.
Contained By:	Dissertations Abstracts International83-05B.
標題:	Crops. -
電子資源:	http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=28688400
ISBN:	9798544204381

Machine Learning for Satellite Imagery when Labels Are Scarce.
Wang, Sherrie.

Machine Learning for Satellite Imagery when Labels Are Scarce. - Ann Arbor : ProQuest Dissertations & Theses, 2021 - 209 p.

Source: Dissertations Abstracts International, Volume: 83-05, Section: B.

Thesis (Ph.D.)--Stanford University, 2021.

This item must not be sold to any third party vendors.

As the world aims to achieve the UN Sustainable Development Goals by 2030, data gaps in the developing world make it difficult to measure progress and target interventions. Recent rapid advances in computer vision and satellite imagery acquisition offer opportunities for automatically extracting knowledge about our planet from space. Compared to field surveys, which are the traditional source of knowledge about human activities and natural ecosystems, satellites offer global coverage at low marginal cost. However, many regions of interest around the world lack ground truth labels on which to train machine learning models. This dissertation will cover two strategies for mapping our planet when labels are scarce: (1) learning better features on satellite imagery to maximize label use efficiency and (2) using non-traditional data sets as ground truth. The application area of focus is agriculture, with a particular focus on cropland and crop type mapping, but the methods and data modes considered can be applied to any domain that uses remotely sensed data.The first part of the dissertation focuses on methodology that learns compressed features from satellite imagery to improve performance on downstream tasks with few or no labels. When no labels are available in the region of interest, we evaluate clustering algorithms and find that clustering on Fourier transform features recovers original classes in simple cropping settings. Our second method, Tile2Vec, leverages innate geographic structure in satellite imagery to learn unsupervised representations. We evaluate Tile2Vec on a variety of remote sensing data sets and demonstrate consistent improvement upon end-to-end supervised learning up to large label quantities. Lastly, we consider the case when high-resource regions in the world have plentiful labeled data but lowresource regions do not. We construct a meta-learning framework for remote sensing tasks and show that model-agnostic meta-learning performs well in previously unseen low-resource regions.The second part of the dissertation explores the use of non-traditional labels to supervise learning. Motivated by the fact that point-level and image-level labels are easier to acquire and more abundant in the wild than densely segmented labels, we show that such labels can be used to segment satellite imagery to high accuracy. Finally, we obtain millions of crowdsourced crop type labels from the Plantix mobile app across southeast India and demonstrate that, with the right data curation steps, noisy labels can be used to train and evaluate a crop type classifier. We produce a 10-meter resolution crop type map in southeast India, the first of its kind in a developing country.

ISBN: 9798544204381Subjects--Topical Terms:

672677
Crops.

Machine Learning for Satellite Imagery when Labels Are Scarce.
LDR:03741nmm a2200325 4500 001 2343561
005 20220512065838.5
008 241004s2021 ||||||||||||||||| ||eng d
020 $a 9798544204381
035 $a (MiAaPQ)AAI28688400
035 $a (MiAaPQ)STANFORDzw179sw4070
035 $a AAI28688400
040 $a MiAaPQ $c MiAaPQ
100 1 $a Wang, Sherrie. $3 3682150
245 1 0 $a Machine Learning for Satellite Imagery when Labels Are Scarce.
260 1 $a Ann Arbor : $b ProQuest Dissertations & Theses, $c 2021
300 $a 209 p.
500 $a Source: Dissertations Abstracts International, Volume: 83-05, Section: B.
500 $a Advisor: Lobell, David.
502 $a Thesis (Ph.D.)--Stanford University, 2021.
506 $a This item must not be sold to any third party vendors.
520 $a As the world aims to achieve the UN Sustainable Development Goals by 2030, data gaps in the developing world make it difficult to measure progress and target interventions. Recent rapid advances in computer vision and satellite imagery acquisition offer opportunities for automatically extracting knowledge about our planet from space. Compared to field surveys, which are the traditional source of knowledge about human activities and natural ecosystems, satellites offer global coverage at low marginal cost. However, many regions of interest around the world lack ground truth labels on which to train machine learning models. This dissertation will cover two strategies for mapping our planet when labels are scarce: (1) learning better features on satellite imagery to maximize label use efficiency and (2) using non-traditional data sets as ground truth. The application area of focus is agriculture, with a particular focus on cropland and crop type mapping, but the methods and data modes considered can be applied to any domain that uses remotely sensed data.The first part of the dissertation focuses on methodology that learns compressed features from satellite imagery to improve performance on downstream tasks with few or no labels. When no labels are available in the region of interest, we evaluate clustering algorithms and find that clustering on Fourier transform features recovers original classes in simple cropping settings. Our second method, Tile2Vec, leverages innate geographic structure in satellite imagery to learn unsupervised representations. We evaluate Tile2Vec on a variety of remote sensing data sets and demonstrate consistent improvement upon end-to-end supervised learning up to large label quantities. Lastly, we consider the case when high-resource regions in the world have plentiful labeled data but lowresource regions do not. We construct a meta-learning framework for remote sensing tasks and show that model-agnostic meta-learning performs well in previously unseen low-resource regions.The second part of the dissertation explores the use of non-traditional labels to supervise learning. Motivated by the fact that point-level and image-level labels are easier to acquire and more abundant in the wild than densely segmented labels, we show that such labels can be used to segment satellite imagery to high accuracy. Finally, we obtain millions of crowdsourced crop type labels from the Plantix mobile app across southeast India and demonstrate that, with the right data curation steps, noisy labels can be used to train and evaluate a crop type classifier. We produce a 10-meter resolution crop type map in southeast India, the first of its kind in a developing country.
590 $a School code: 0212.
650 4 $a Crops. $3 672677
650 4 $a Remote sensing. $3 535394
650 4 $a Time series. $3 3561811
650 4 $a Crowdsourcing. $3 3377825
650 4 $a Agronomy. $3 2122783
650 4 $a Artificial intelligence. $3 516317
690 $a 0799
690 $a 0285
690 $a 0800
710 2 $a Stanford University. $3 754827
773 0 $t Dissertations Abstracts International $g 83-05B.
790 $a 0212
791 $a Ph.D.
792 $a 2021
793 $a English
856 4 0 $u http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=28688400