This paper discusses the remote sensing satellite image processing and its development trend.

When the remote sensing satellite in the air obtains the digital image of the earth and sends it back to the ground, will the work be over? The answer is obviously no, on the contrary, this is the beginning of remote sensing digital image processing.

Remote sensing digital image (hereinafter referred to as "remote sensing image") is a digital remote sensing image. Different areas and objects on the earth's surface can reflect or radiate electromagnetic waves with different wavelengths. Using this feature, remote sensing system can produce different remote sensing digital images.

It differs from ordinary digital images, that is, electronic photos we usually take, in the imaging range and fineness of remote sensing images. The shooting area of remote sensing satellite is a macro dimension on the earth level, and each pixel in the image corresponds to several, one or a part of ground objects in the three-dimensional real world. Depending on the resolution of satellite imaging, a pixel may be a tree, a car or a window of a building.

Therefore, the brightness value (DN value, digital number) of each pixel in the image has important information significance. In order to obtain accurate information, users need to manage, transform, correct, enhance and extract pixels in satellite images according to their own application goals, which is convenient for further mining and business integration applications.

DN value (digital): the brightness value of pixels in remote sensing image, which records the gray value of ground objects. No unit is an integer value, which is related to the radiation resolution of the sensor, the emissivity of the ground object, the atmospheric transmittance and the scattering rate, and reflects the radiation brightness of the ground object.

We can go back to the "P-picture world" for analogy. In order to make our social media image more perfect, we have launched a Xiu Xiu software to whiten, slim, exfoliate and remove acne ... Of course, the data processing of remote sensing images is much more complicated and professional. To what extent? Can be written as a textbook—

Today, let's take a look at what "magic skills" are and how to use them. With the rapid development of remote sensing industry today, will the impact of high-frequency data output, algorithms and artificial intelligence change the traditional mode and underlying logic of these "god operations"?

0 1. What is remote sensing image processing?

Remote sensing image processing is a process of using computer image processing system to perform a series of operations on pixels in remote sensing images.

Remote sensing images contain a lot of information. Only after digitization (sampling, quantization and digital storage of imaging system) can information analysis and content extraction be effectively carried out. On this basis, image data can be processed and reprocessed, such as correcting the coordinates of graphic alignment and enhancing the contour of ground objects, which can greatly improve the accuracy of image processing and the efficiency of information extraction. This process can be called "remote sensing digital image processing".

As a basic and important link in the process of "Earth observation", remote sensing image processing plays an important role in the middle and lower reaches of the satellite application industry chain, connecting the past with the future. The front end undertakes satellite ground facilities, and the back end provides "ready-made" data services or tools for specific business applications in agriculture, forestry, meteorology, natural resources and other industries.

02. Why is remote sensing image processing the only way for application?

When we see neat and beautiful digital earth products like Google Earth, or remote sensing satellite thematic maps or interpretation maps applied to natural resource management, environmental protection, agriculture, meteorology and other fields, we need to be "baptized" in the middle of image processing.

Because the remote sensing satellite "works" at high altitude, its imaging environment is far more complicated than our daily photographing environment on the ground, and it will encounter geometric deformation, distortion, blur, noise and so on caused by system and non-system factors such as sensor instability, earth curvature, atmospheric conditions, illumination changes and terrain changes. Remote sensing data center carries out preliminary processing on the image, such as striping and geometric coarse correction. When the data reaches all end users, it needs to be further refined to make it closer to the real world's physical space environment and coordinates, and processed professionally according to its own business analysis objectives, so as to prepare for the next remote sensing image analysis, interpretation and business application.

Generally speaking, the main objectives of remote sensing image processing are as follows:

Image correction: restoring and restoring images. Before information extraction, the remote sensing image must be corrected so that the image can correctly reflect the actual ground information or physical process.

Image enhancement: suppress or eliminate image noise. In order to make the feature information contained in remote sensing images more readable, the objects of interest more prominent and easier to understand and explain, it is necessary to enhance the whole image or specific feature information.

Information extraction: according to the spectral characteristics and geometric characteristics of ground objects, the extraction rules of different ground object information are determined, and on this basis, various useful ground object information is extracted from corrected remote sensing data by using this rule.

03. What are the functions of remote sensing data processing?

The complete remote sensing digital image processing includes two parts: hardware system and software system. Remote sensing data storage is huge, which requires large-capacity digital storage equipment and software to cooperate with storage and processing. This paper mainly introduces the software processing part. The following figure is the interface of professional image processing software. Compared with common office software, the functions of image processing system are scattered and the links between menus are not tight.

In a sense, the image processing system is more like an image processing comprehensive toolbox. Because of different image processing objectives, users can call a function or a combination of several functions, and not all processes are selected. Some typical processing functions are summarized here, and the basic steps are introduced.

Digital storage and management

The storage capacity of remote sensing image itself is very large. The landsat remote sensing image in band 1 scene 7 is at least 200MB, while the hyperspectral image may reach 1 GB. Since entering the era of double heights of time and space, the high-frequency output and accumulation of data has also prompted remote sensing to enter the era of big data, making the trend of remote sensing cloud service, storage management and rapid distribution more and more obvious. Digital storage, online updating, management retrieval, publishing and browsing of remote sensing images based on private cloud and hybrid cloud have gradually become an inseparable and important foundation for remote sensing data processing, which will greatly improve the professional processing and business application efficiency of subsequent remote sensing images.

Image preprocessing

Radiation correction (radiation correction)

It refers to the process of correcting systematic and random radiation distortion or distortion caused by external factors and data acquisition and transmission system, and eliminating or correcting image distortion caused by radiation error.

Simply summarized, it is to remove the sensor or atmospheric "noise", more accurately represent the ground situation, and improve the "fidelity" of the image, mainly to recover the missing data, remove the mist, or prepare for mosaic and change monitoring.

The role of radiation correction in dynamic monitoring: In multi-temporal remote sensing images, the radiation value of invariant objects will be different in different temporal images besides the changes of objects. If it is necessary to use the spectral information of multi-temporal remote sensing images to dynamically monitor the changes of ground objects, we must first eliminate the differences of radiation values of invariant ground objects.

Through relative radiation correction, one image is taken as a reference (or reference) image, and the DN value of the other image is adjusted, so that the objects with the same name on the two phase images have the same DN value. This process is also called spectral normalization of multi-temporal remote sensing images. In this way, the change monitoring can be realized by analyzing the difference of radiation values on remote sensing images in different phases, so as to complete the remote sensing dynamic monitoring of the dynamic change of ground objects.

Geometric correction (geometric correction)

In the process of remote sensing imaging, due to the comprehensive influence of photosensitive material deformation, objective lens distortion, atmospheric refraction, earth curvature, earth rotation, terrain fluctuation and other factors, the geometric position, shape, size, scale, orientation and other characteristics of ground objects in the original image are often inconsistent with their corresponding characteristics. This phenomenon is called geometric deformation, also known as geometric distortion. Geometric correction is to correct and eliminate this geometric distortion through a series of mathematical models to make its positioning accurate.

The principle of geometric correction shows that the terrain in the real world is three-dimensional and uneven, but the remote sensing satellite sensor can only obtain two-dimensional pixels, which brings terrain distortion | Source: Network; Redrawing: overlooking time and space

Image enhancement

Image contrast enhancement (image contrast enhancement)

The random distribution obtained by counting the number of pixels of each brightness of each image is the histogram of the image. Generally speaking, the random distribution of pixel brightness should be normal in an image containing a large number of pixels. The histogram is non-normal, indicating that the brightness distribution of the image is too bright, too dark or too concentrated, and the contrast of the image is small. It is necessary to adjust the histogram to normal distribution to improve the image quality, and to distinguish the contour of ground objects and extract information.

color synthesis

In order to make full use of the advantages of color in remote sensing image interpretation and information extraction, multi-spectral images are often processed by color synthesis to obtain color images. As shown above, color images can be divided into true color images and false color images.

density slicing

Gray image is graded according to the gray value of pixels, and then graded with different colors, so that the original gray image becomes a pseudo-color image to achieve the purpose of image enhancement.

Image manipulation

After registering two or more single-band images in space, arithmetic operations can be performed to enhance the images. According to the gray difference of ground objects in different bands, new "bands" are generated by algebraic operations in different bands, such as addition, subtraction, ratio and synthesis, such as:

Subtraction operation: it can highlight the ground objects with large differences between infrared and red bands, and highlight vegetation information.

Ratio operation: it is often used to calculate vegetation index and eliminate terrain shadow.

Vegetation index: NDVI=(IR-R)/(IR+R)

Image fusion

Remote sensing image information fusion is an effective means to improve image resolution and information content. It is a process of generating a set of new information or synthesizing images from multi-source remote sensing data by using certain algorithms in a unified geographical coordinate system.

Different remote sensing data have different spatial resolution, spectral resolution and phase resolution. The image processing technology of resampling low-resolution multi-spectral images and high-resolution single-band images to generate high-resolution multi-spectral remote sensing images makes the processed images have high spatial resolution and multi-spectral characteristics at the same time.

Image editing

In the practical application of remote sensing, users may only be interested in the information in a specific range of remote sensing images, which requires narrowing the remote sensing images to the size of the research range. Common cropping methods include cropping by ROI (region of interest), cropping by file (according to the size of the specified image file) and cropping by map (according to the geographic coordinates or latitude and longitude range of the map).

Image mosaic

Also called image stitching, it is a technical process of stitching two or more digital images (possibly obtained under different photographic conditions) together to form a complete image. Usually, each image is geometrically corrected first, and they are planned into a unified coordinate system, then they are cut to remove the overlapping parts, and then the cut images are spliced to form a large-format image.

Mosaic uniform color

Several adjacent remote sensing images are merged into a unified new image by stitching and color equalization technology.

Information extraction

The characteristics of target objects in remote sensing images reflect the differences of electromagnetic radiation of objects in remote sensing images. According to the characteristics of ground objects in remote sensing images, the process of identifying the types, properties, spatial positions, shapes and sizes of ground objects is remote sensing information extraction.

Visual interpretation

It is also called manual interpretation, that is, the remote sensing image is interpreted by artificial naked eyes and experience, and the range of the target object on the remote sensing image is manually delineated to achieve the purpose of information extraction. Manual interpretation is a traditional information extraction method, but the efficiency of interpretation and analysis is relatively low under massive images.

Image classification

According to the spectral characteristics of ground objects, the discriminant function and corresponding discriminant criterion are determined, and all pixels of the image are divided into several categories according to their properties. The main methods are divided into supervised classification and unsupervised classification.

-supervised classification

Supervised classification means that people have prior knowledge of the category attributes of remote sensing image sample areas before classification, and then they can establish and train classifiers based on the characteristics of these sample categories (that is, establish discriminant functions), and then complete the classification of the whole image and merge each pixel into the corresponding category.

Supervised classification is also the most common application method of remote sensing AI at present, that is, using machine learning to classify, label or identify specific features through sample database.

-Unsupervised classification

Unsupervised classification, also known as cluster analysis, refers to "blind" classification based on data (the distribution law of spectral characteristics of remote sensing images), that is, the characteristics of natural clustering, without imposing any prior knowledge in the classification process; Based on clustering theory, it is a method of statistical analysis of images by computer, and it is also a method of pattern recognition. General algorithms include: regression analysis, trend analysis, equal mixed distance method, cluster analysis, principal component analysis and pattern recognition.

The difference between supervised classification and unsupervised classification: supervised classification must have training sets and test samples. Find out the rules in the training set and use the rules in the test samples; Unsupervised, there is no training set, only a set of data, in which to find rules.

04. What changes are taking place in remote sensing data processing?

Remote sensing data processing is more like "rough processing of raw materials" in manufacturing industry, and it is also the pre-means of intelligent application and business integration of remote sensing image data. From the introduction of the last article, its process is also more complicated and professional.

As an important part of earth observation and remote sensing industrialization, remote sensing data processing in the middle and lower reaches of the industry has also been impacted by the era of big data, and is changing in line with this trend, moving towards real-time, standardization, scale and automation.

In the digital transformation of enterprises, it is often said that all traditional industries are worth doing again with digitalization, as are traditional data production and information service industries, and their models and processes are worth doing again with algorithms and AI.

When algorithms and artificial intelligence gradually penetrate into remote sensing data processing, many problems in data production service in remote sensing industry can be solved, such as long data distribution cycle and links, many processing links, accuracy and consistency of massive data processing, etc. We can regard it as "automatic batch processing".

After the algorithm engine solves the problems of data service, data calculation efficiency and automation flow, there will be more refined application data products suitable for various vertical subdivision scenarios downstream. With the participation of AI and algorithm, there will be many efficient automation functions in the remote sensing image information extraction mentioned above, such as target recognition, feature extraction, feature classification, change detection and so on. This will gradually help mankind to improve interpretation efficiency and form an "intelligent information mining" mechanism in the downstream of remote sensing industry.

We can see that the efficiency of remote sensing data collection is closely related to the underlying data model from the source, data processing to terminal application. With the gradual formation of satellite Internet and Earth observation constellation, only by standardizing the process of data acquisition, processing and sharing can the large-scale, automated and streamlined remote sensing industry better exert its kinetic energy for the digital transformation of government and enterprises, and truly usher in the era of space-time big data.

reference data

Course of Remote Sensing Digital Image Processing edited by Wei Yuchun, Guoan Tang and Yang Xin.