卫星图像分析：聚类与分类的深度探索

需积分: 10 195 浏览量更新于2024-07-17 收藏 4.55MB PDF 举报

《卫星图像分析聚类与分类》是一本于2019年发布的新书，隶属于SpringerBriefs in Applied Sciences and Technology系列中的Computational Intelligence子系列。该系列由Janusz Kacprzyk教授编辑，他来自波兰科学院系统研究所在华沙的研究机构。该系列致力于快速且高质量地发布计算智能领域的最新发展和进步，涵盖工程、计算机科学、物理学和生命科学等多个领域，以及这些领域中计算智能的理论、应用和设计方法。本书的核心主题是卫星图像分析，具体探讨了聚类和分类技术在这一特殊领域的应用。卫星图像分析对于遥感科学、地理信息系统（GIS）、环境监测、农业、灾害管理等多个行业至关重要，因为它提供了对地球表面的高分辨率数据，能够帮助科学家们做出精确的地理和环境决策。在书中，作者Surekha Borra、Rohit Thanki和Nilanjan Dey可能讨论了以下几个关键知识点： 1. **卫星图像获取与预处理**：介绍了如何通过遥感卫星获取地球表面的多光谱图像，以及如何进行图像去噪、校准和几何纠正等预处理步骤，确保后续分析的准确性。 2. **特征提取与选择**：讲解了如何从海量卫星图像中提取有用的特征，如纹理、形状、色彩和光谱信息，用于区分不同的地理特征或目标。 3. **聚类算法**：深入探讨了诸如K-means、DBSCAN、谱聚类等在卫星图像上的应用，如何根据相似性度量将图像像素或地区划分为不同的类别。 4. **分类技术**：详细介绍了支持向量机（SVM）、决策树、随机森林、深度学习等机器学习和人工智能技术在卫星图像分类中的应用，例如区分城市、森林、农田或海洋等。 5. **混合模型与集成方法**：可能会提到结合传统统计方法和深度学习的策略，提高卫星图像分析的准确性和鲁棒性。 6. **案例研究与应用示例**：书中可能包含实际案例，展示如何在特定地理环境中（如气候变化监测、自然资源管理或灾害预警）成功运用聚类和分类技术。 7. **挑战与未来方向**：讨论了当前在卫星图像分析中的技术限制，如数据不确定性、模型解释性以及如何适应不断发展的遥感技术。《卫星图像分析聚类与分类》提供了一个深入理解这个领域的实用指南，对于从事遥感科学、地理信息科学以及任何依赖卫星数据的应用来说，都具有重要的参考价值。

Chapter 1

Introduction

1.1 Introduction

The satellite imaging system allows easy and effective remote collection of large

amounts of information about the objects of interest from even inaccessible areas

for a long time, making it possible the change detection analysis at different scales

over time. The satellite imaging system obtains and measures electromagnetic energy

(EM) emanating from or reﬂected from earth objects as shown in Fig. 1.1 [1]. The

electromagnetic spectrum is a combination of different wavelengths which includes

visible light, infrared, microwave, and ultraviolet. The main source of EM energy is

the sun. Every object on the earth has different reﬂection strengths at different wave-

lengths. The strength of the reﬂection or radiation helps to identify various objects

such as water, trees, building, forests, etc., in the satellite image. The electromagnetic

energy or radiation from the objects of interest can be expressed either in terms of

frequency or wavelength as follows:

E  hc f ,

E  hc/λ (1.1)

where λ is the wavelength, f is frequency, c is the speed of light, and h is Planck’s

constant.

Equation (1.1) indicates that short wavelengths result in high energy.

The quality of satellite imaging depends on the type of energy sources (active/

passive/uniform/nonuniform), sensor resolutions (spatial/temporal/spectral/

radiometric), atmospheric interference, real-time energy reﬂectance/absorption

values, and data handling systems.

S. Borra et al., Satellite Image Analysis: Clustering and Classiﬁcation,

SpringerBriefs in Computational Intelligence,

https://doi.org/10.1007/978-981-13-6424-2_1

1.3 Panchromatic and Multispectral Images 3

Passive Sensor

Earth

Sun

Active Sensor

Earth

Fig. 1.2 The concept of active and passive sensors

image is denoted as a matrix of digital numbers, which are stored in the form of bits,

the length of which is related to the maximum number of colors/intensities/gray

levels used to represent each image.

Satellite images are generated by various sensors and scanners that operate in

a variety of bands of the EM spectrum. Usage of more spectral bands provides

more information with details, but involves storage and transmission complexity.

The width of the spectral band determines the strength of the signal. Wide spectral

bands collect radiation/photons/energy over a range of wavelengths in that band,

summing up to more energy and hence the detected signal is strong. The strength of

the signal collected from narrow spectral bands is low, but provides unique details.

Different class of land use or land covers reﬂects or absorbs differently in different

spectral bands.

Panchromatic images are the images obtained in a wide band covering the entire

visible range. Spectral images are representations of measured data corresponding

to reﬂectance or radiation in various spectral bands of the electromagnetic spectrum.

These images can be combined in various ways using different indices or can be com-

posite color mapped (false-color/pseudo color) for efﬁcient extraction of features,

and for easy and accurate interpretation of data.

Multispectral/hyperspectral images are set of images acquired by multiple sensors

that operate at discrete/continuous narrow wavelength ranges (bands/spectrums).

Spectral images obtained from bands other than visible band provide additional

information, which human eye fails to capture about the scene. Examples include

the objects under low illumination (darkness), hazy environments (smoke or fog),

earth’s surface (landmines and minerals), and military targets such as ballistic

missiles and tactical vehicles. Analysis of multiple spectral satellite images can be

done independently and in different band combinations using spectral signatures.

The classiﬁcation of pixels of multispectral images is a complex task and demands

validation of training samples with ground truth apart from expert’s knowledge.

Software such as Gerbil, Multispec, Opticks, Micro MSI are widely used for

visualization and analysis of multispectral images.

4 1 Introduction

The information extracted from satellite images belongs to a variety of land

use/land cover classes such as urbanization, vegetation, navigation, disaster man-

agement, minerals/resource estimation, valleys, mountains, deserts, rivers, plains,

streets, military areas, and so on.

1.4 Resolution in Satellite Images

Resolution determines the ability to detect, identify, and distinguish different objects

and their properties which vary with time. The resolution corresponds to the amount

of information or details that that is provided by the image. Four types of resolutions

with respect to satellite images are spatial, spectral, temporal, and radiometric reso-

lutions [1, 3]. Spectral resolution corresponds to the width, sensitivity, and location

of the spectral bands and their number. Higher the spectral resolution, higher the

distinguishing ability. Typical GIS satellites capture images in 3–8 spectral bands.

Spatial resolution corresponds to the smallest area covered by a single sensor at

any time. Larger the spatial resolution, smaller the pixels are, and hence higher the

ability to distinguish the s mallest details of the scene under acquisition. A typical

spatial resolution varies from a few km to cm. The objects closer than the resolution

appear as a single object in the satellite image. The system with more than 1 km spatial

resolution is considered as a low-resolution system. The system with 100 m to 1 km

spatial resolution is considered as a medium resolution system. The spatial resolution

of the high-resolution system is between 250 and 500 m. The satellite system with

approximately 5–100 m spatial resolution is referred to as a high-resolution system.

The very-high-resolution satellite systems have less than 5 m spatial resolution.

Temporal resolution corresponds to the time interval between image acquisitions

or observations of the same scene (revisit time/repetition rate). It depends on factors

such as viewing angle and satellite orbit. Radiometric resolution refers to the number

of gray levels (precision of digital numbers) used for displaying the scanners/sensor

data in the form of an image. It depends on the number of bits used to represent each

pixel value. With the high radiometric resolution, measurement of ﬁne variations in

reﬂectance is possible. However, the resolution depends on the type of sensors used.

1.5 Distortions in Satellite Images

Various factors such as the resolution of the image, irregularities in angular position

of the image object, the orientation of the subjects in the image, and occlusion of one

object by another object in the image [3–5] leads to distorted satellite images. The

geometric distortions in images mainly occur due to sensor characteristics, viewing

geometry, variations in aspect ratio, and motion of the satellite or object. The optical

distortions can be Pincushion distortion, Barrel panoramic effects, and mirror veloc-

ity variations. The distortions also occur due to motion and orientation aspects of

剩余109页未读，继续阅读

weixin_38290023

粉丝: 4
资源: 224

卫星图像分析：聚类与分类的深度探索

IMAGE ANALYSIS,CLASSIFICATION AND CHANGE DETECTION

Survey of Text Mining II Clustering Classification and Retrieval

survey of text mining: clustering,classification,and retrieval

image segmentation

Statistics and Machine Learning Toolbox

Unsupervised texture segmentation using feature distributions

deep adaptive image clustering

如何把自监督和聚类方法整合到nnUNet框架，请推荐近几年的top的会议工作

matrix pre post

基于模拟数据sklearn.datasets.make_classification 分别使用KMeans聚类、DBSCAN算法对其进行聚类

最新资源