Deep learning and process understanding for data-driven Earth system science

Nature, Journal Year: 2019, Volume and Issue: 566(7743), P. 195 - 204

Published: Feb. 1, 2019

Language: Английский

---

Deep Learning in Remote Sensing: A Comprehensive Review and List of Resources

IEEE Geoscience and Remote Sensing Magazine, Journal Year: 2017, Volume and Issue: 5(4), P. 8 - 36

Published: Dec. 1, 2017

Standing at the paradigm shift towards data-intensive science, machine learning techniques are becoming increasingly important. In particular, as a major breakthrough in field, deep has proven an extremely powerful tool many fields. Shall we embrace key to all? Or, should resist 'black-box' solution? There controversial opinions remote sensing community. this article, analyze challenges of using for data analysis, review recent advances, and provide resources make ridiculously simple start with. More importantly, advocate scientists bring their expertise into learning, use it implicit general model tackle unprecedented large-scale influential challenges, such climate change urbanization.

Language: Английский

---

Deep Feature Extraction and Classification of Hyperspectral Images Based on Convolutional Neural Networks

IEEE Transactions on Geoscience and Remote Sensing, Journal Year: 2016, Volume and Issue: 54(10), P. 6232 - 6251

Published: July 19, 2016

Due to the advantages of deep learning, in this paper, a regularized feature extraction (FE) method is presented for hyperspectral image (HSI) classification using convolutional neural network (CNN). The proposed approach employs several and pooling layers extract features from HSIs, which are nonlinear, discriminant, invariant. These useful target detection. Furthermore, order address common issue imbalance between high dimensionality limited availability training samples HSI, few strategies such as L2 regularization dropout investigated avoid overfitting class data modeling. More importantly, we propose 3-D CNN-based FE model with combined effective spectral-spatial imagery. Finally, further improve performance, virtual sample enhanced proposed. approaches carried out on three widely used sets: Indian Pines, University Pavia, Kennedy Space Center. obtained results reveal that models sparse constraints provide competitive state-of-the-art methods. In addition, opens new window research.

Language: Английский

---

Remote Sensing Image Scene Classification: Benchmark and State of the Art

Proceedings of the IEEE, Journal Year: 2017, Volume and Issue: 105(10), P. 1865 - 1883

Published: April 3, 2017

Remote sensing image scene classification plays an important role in a wide range of applications and hence has been receiving remarkable attention. During the past years, significant efforts have made to develop various data sets or present variety approaches for from remote images. However, systematic review literature concerning methods is still lacking. In addition, almost all existing number limitations, including small scale classes numbers, lack variations diversity, saturation accuracy. These limitations severely limit development new especially deep learning-based methods. This paper first provides comprehensive recent progress. Then, we propose large-scale set, termed "NWPU-RESISC45," which publicly available benchmark REmote Sensing Image Scene Classification (RESISC), created by Northwestern Polytechnical University (NWPU). set contains 31 500 images, covering 45 with 700 images each class. The proposed NWPU-RESISC45 1) on total number; 2) holds big translation, spatial resolution, viewpoint, object pose, illumination, background, occlusion; 3) high within-class diversity between-class similarity. creation this will enable community evaluate data-driven algorithms. Finally, several representative are evaluated using results reported as useful baseline future research.

Language: Английский

---

Remote Sensing Image Scene Classification Meets Deep Learning: Challenges, Methods, Benchmarks, and Opportunities

IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, Journal Year: 2020, Volume and Issue: 13, P. 3735 - 3756

Published: Jan. 1, 2020

Remote sensing image scene classification, which aims at labeling remote images with a set of semantic categories based on their contents, has broad applications in range fields. Propelled by the powerful feature learning capabilities deep neural networks, classification driven drawn remarkable attention and achieved significant breakthroughs. However, to best our knowledge, comprehensive review recent achievements regarding for is still lacking. Considering rapid evolution this field, paper provides systematic survey methods covering more than 160 papers. To be specific, we discuss main challenges (1) Autoencoder-based methods, (2) Convolutional Neural Network-based (3) Generative Adversarial methods. In addition, introduce benchmarks used summarize performance two dozen representative algorithms three commonly-used benchmark data sets. Finally, promising opportunities further research.

Language: Английский

---

Advanced Spectral Classifiers for Hyperspectral Images: A review

IEEE Geoscience and Remote Sensing Magazine, Journal Year: 2017, Volume and Issue: 5(1), P. 8 - 32

Published: March 1, 2017

Hyperspectral image classification has been a vibrant area of research in recent years. Given set observations, i.e., pixel vectors hyperspectral image, approaches try to allocate unique label each vector. However, the images is challenging task for number reasons, such as presence redundant features, imbalance among limited available training samples, and high dimensionality data.

Language: Английский

---

Scene Classification With Recurrent Attention of VHR Remote Sensing Images

IEEE Transactions on Geoscience and Remote Sensing, Journal Year: 2018, Volume and Issue: 57(2), P. 1155 - 1167

Published: Sept. 5, 2018

Scene classification of remote sensing images has drawn great attention because its wide applications. In this paper, with the guidance human visual system (HVS), we explore mechanism and propose a novel end-to-end recurrent convolutional network (ARCNet) for scene classification. It can learn to focus selectively on some key regions or locations just process them at high-level features, thereby discarding noncritical information promoting performance. The contributions paper are threefold. First, design structure squeeze semantic spatial features into several simplex vectors reduction learning parameters. Second, an named ARCNet is proposed adaptively select series then generate powerful predictions by sequentially. Third, construct new data set OPTIMAL-31, which contains more categories than popular sets gives researchers extra platform validate their algorithms. experimental results demonstrate that our model makes promotion in comparison state-of-the-art approaches.

Language: Английский

---

Domain Adaptation for the Classification of Remote Sensing Data: An Overview of Recent Advances

IEEE Geoscience and Remote Sensing Magazine, Journal Year: 2016, Volume and Issue: 4(2), P. 41 - 57

Published: June 1, 2016

The success of the supervised classification remotely sensed images acquired over large geographical areas or at short time intervals strongly depends on representativity samples used to train algorithm and define model. When training are collected from an image a spatial region that is different one for mapping, spectral shifts between two distributions likely make model fail. Such generally due differences in acquisition atmospheric conditions changes nature object observed. To design methods robust data set shifts, recent remote sensing literature has considered solutions based domain adaptation (DA) approaches. Inspired by machine-learning literature, several DA have been proposed solve specific problems classification. This article provides critical review advances approaches presents overview divided into four categories: 1) invariant feature selection, 2) representation matching, 3) classifiers, 4) selective sampling. We provide methodologies, examples applications techniques real characterized very high resolution as well possible guidelines selection method use application scenarios.

Language: Английский

---

Remote Sensing and Cropping Practices: A Review

Remote Sensing, Journal Year: 2018, Volume and Issue: 10(1), P. 99 - 99

Published: Jan. 12, 2018

For agronomic, environmental, and economic reasons, the need for spatialized information about agricultural practices is expected to rapidly increase. In this context, we reviewed literature on remote sensing mapping cropping practices. The studies were grouped into three categories of practices: crop succession (crop rotation fallowing), pattern (single tree planting pattern, sequential cropping, intercropping/agroforestry), techniques (irrigation, soil tillage, harvest post-harvest practices, varieties, agro-ecological infrastructures). We observed that majority exploratory investigations, tested a local scale with high dependence ground data, used only one type sensor. Furthermore, be correctly implemented, most methods relied heavily knowledge management environment, biological material. These limitations point future research directions, such as use land stratification, multi-sensor data combination, expert knowledge-driven methods. Finally, new spatial technologies, particularly Sentinel constellation, are improve monitoring in challenging context food security better agro-environmental issues.

Language: Английский

---

Deep Hyperspectral Image Sharpening

IEEE Transactions on Neural Networks and Learning Systems, Journal Year: 2018, Volume and Issue: 29(11), P. 5345 - 5355

Published: Feb. 20, 2018

Hyperspectral image (HSI) sharpening, which aims at fusing an observable low spatial resolution (LR) HSI (LR-HSI) with a high (HR) multispectral (HR-MSI) of the same scene to acquire HR-HSI, has recently attracted much attention. Most recent sharpening approaches are based on priors modeling, usually sensitive parameters selection and time-consuming. This paper presents deep method (named DHSIS) for fusion LR-HSI HR-MSI, directly learns via convolutional neural network-based residual learning. The DHSIS incorporates learned into HR-MSI framework. Specifically, we first initialize HR-HSI from framework solving Sylvester equation. Then, map initialized reference learning learn priors. Finally, returned reconstruct final HR-HSI. Experimental results demonstrate superiority approach over existing state-of-the-art in terms reconstruction accuracy running time.

Language: Английский

---