Remote sensing image land classification method for multi-cloud environment

Abstract

The invention belongs to the technical field of satellite remote sensing image surveying and mapping, and particularly relates to a remote sensing image land classification method for a multi-cloud environment. A cloud removal algorithm parameter is optimized for a multi-cloud environment, and the problem of data voids formed by long-term coverage of a cloud layer is effectively solved. A land feature index system is established, the accuracy of feature extraction of different lands in different regions is improved, a sample training scheme is established based on the land features, and the accuracy of remote sensing image recognition and classification is enhanced.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of satellite remote sensing image surveying and mapping, and particularly relates to a remote sensing image land classification method for a cloudy environment. BACKGROUND

[0002] Land classification using satellite remote sensing images is a basic technology in geographic information science. The technology aims to assign corresponding land cover type labels (such as water, forest, farmland, urban buildings, etc.) to remote sensing images, so as to facilitate land resource management and research. The cloudy environment has a great impact on the accuracy of land classification. In order to improve the accuracy, the existing technology generally uses methods such as cloud removal processing provided by the remote sensing image, land feature extraction to enhance classification ability, or enhanced remote sensing image recognition technology.

[0003] Remote sensing image cloud removal processing is to use the quality evaluation band provided by the image or a special cloud detection algorithm to identify the pixels covered by clouds and cloud shadows, and mark them as invalid data (mask) to exclude the interference caused by cloud cover in the land area, so as to avoid classification errors caused by participation of these areas in subsequent analysis. However, there are problems such as unsatisfactory cloud removal effect, and if the area is long-term cloudy, there will still be a large number of invalid pixels after image processing, forming data holes.

[0004] Land feature extraction processing is to calculate a series of spectral indices in addition to directly using the original spectral bands of satellite images. These indices highlight or enhance the spectral response characteristics of target ground objects through specific band combination operations to enhance the accuracy of image recognition. However, there are problems such as difficulty in accurately extracting land features: the existing land feature system usually only includes conventional indices, and lacks feature indices in cloudy environments, etc.

[0005] Remote sensing image recognition and classification processing is to use the collected training sample points and their corresponding spectral features to train a supervised classification model, and apply the trained classification model to remote sensing image data to finally generate a complete land cover classification map. However, this method lacks a targeted model training scheme, which can easily cause confusion and misidentification of land use types. Therefore, it is necessary to develop a more accurate land classification model for remote sensing images in a cloudy environment. SUMMARY

[0006] Based on the above problems, the present application provides a remote sensing image land classification method for a cloudy environment. The method first constructs an optimized cloud removal algorithm based on the cloud removal algorithm and land feature index for the cloudy environment, effectively solving the problem of data holes formed by long-term cloud cover. A land feature index system is established to improve the accuracy of land feature extraction in different regions, and a sample training scheme is established based on the land feature index to enhance the accuracy of remote sensing image recognition and classification.

[0007] In order to achieve the above effects, the first technical solution of the application discloses a model construction method for land classification of remote sensing images in a cloudy environment, comprising the following steps:

[0008] S1. Cloudy image processing: acquiring Sentinel-2 satellite images of a target area, and analyzing the QA60 quality evaluation band thereof to generate a dynamic cloud mask layer through the binary bits representing clouds and cirrus in the band; performing cloud removal processing on the cloud mask layer to obtain a cloud-removed remote sensing image;

[0009] S2. Land feature index construction: constructing a land feature index of the cloud-removed remote sensing image of S1, and performing feature extraction;

[0010] S3. Image recognition and classification: based on the land feature index constructed in S2, collecting multi-element ground object samples to construct a training sample library for the cloudy environment; using a random forest algorithm and applying optimized random forest model parameters to perform model training to obtain a land classification model for remote sensing images in a cloudy environment.

[0011] Further, the cloud removal processing of S1 includes time sequence image synthesis: performing multi-temporal median synthesis on the image set after cloud masking, and filling in the cloud layer blocked area using time sequence.

[0012] Further, the feature extraction of S2 includes:

[0013] Vegetation features: extracted through normalized vegetation index NDVI, enhanced vegetation index EVI, and red edge position index REPI;

[0014] Water features: extracted through normalized water index NDWI and modified water index mNDWI;

[0015] Building features: extracted through building land index IBI;

[0016] Special surface features: extracted through bare soil index BSI;

[0017] Ice and snow features: extracted through normalized snow index NDSI.

[0018] Further, the multi-element ground object sample collection of S3 includes a 17-class ground object sample system: the multi-element ground object sample collection of S3 includes a 17-class ground object sample system: paddy field, dry land, greenhouse, arbor forest, shrub forest, sparse forest land, other forest land, grassland, river, tidal flat, snow land, urban construction land, rural construction land, other construction land, bare land, rock, and other types.

[0019] Further, the S3 optimized random forest model parameters are: 190 decision trees, 6 feature quantities, 6 minimum sample numbers of nodes, 0.7 feature sampling ratio, 50 maximum node numbers, and random seed 200, and the overall accuracy and KAPPA coefficient reach 0.80 after training.

[0020] Further, after the model is constructed, the overall accuracy and KAPPA coefficient of the confusion matrix are used to verify the model accuracy.

[0021] In addition, the land classification model for remote sensing images in a multi-cloud environment obtained by any of the above construction methods.

[0022] The second technical solution of the present application discloses a land classification method for remote sensing images in a multi-cloud environment, which identifies and classifies the land classification model for remote sensing images in a multi-cloud environment.

[0023] In addition, a land classification system for remote sensing images in a multi-cloud environment includes:

[0024] The multi-cloud image processing module is divided into a dynamic cloud mask algorithm module, a time series image synthesis module, and a terrain data fusion module, and is used for cloud removal processing of the target area in a multi-cloud environment;

[0025] The land feature index construction module is used to construct land features according to the terrain of the target area.

[0026] The image recognition and classification module is used for feature operation and outputs the land classification result.

[0027] In addition, a computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor is used to execute the land classification system for remote sensing images in a multi-cloud environment.

[0028] Advantages:

[0029] Effectively solve the influence of multi-cloud data in the subtropical region: through the dynamic cloud mask + time series median synthesis technology, the available rate of effective images in the multi-cloud area is improved by 35%, the data filling accuracy of the cloud-shielded area is 89.2%, and the long-existing image void problem in the multi-cloud area is solved.

[0030] Achieve high-precision classification: the scheme achieves good precision verification effect. Taking 2023 as an example, the overall accuracy of a certain multi-cloud area is 0.84, and the KAPPA coefficient is 0.80, both reaching high-precision level. BRIEF DESCRIPTION OF DRAWINGS

[0031] Figure 1 is the technical roadmap of the scheme of the present application;

[0032] Figure 2is a contrast chart before and after cloud removal;

[0033] Figure 3 is a contrast chart before and after feature extraction;

[0034] Figure 4 is a contrast chart between training and recognition of image classification. DETAILED DESCRIPTION

[0035] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0036] Unless otherwise specifically stated, the relative arrangements of parts and steps, numerical expressions, and numerical values set forth in the various examples herein are not limiting but are exemplary. Also, it is to be understood that the phraseology and terminology used herein is for the purpose of description and not of limitation. Any use of section headings is not intended to limit the scope of the application. Any section headings used herein are for organizational purposes only and are not to be used to construe the contents or meaning of the sections under the headings. Any use of the term “example” is intended to be an example of something that is not to be taken as limiting. Thus, other examples of the exemplary embodiments can have different values. It is noted that like references and descriptions set forth herein (throughout the several drawings) are directed to like items unless otherwise noted.

[0037] Unless otherwise specifically stated, the relative arrangements of parts and steps, numerical expressions, and numerical values set forth in the various examples herein are not limiting but are exemplary. Also, it is to be understood that the phraseology and terminology used herein is for the purpose of description and not of limitation. Any use of section headings is not intended to limit the scope of the application. Any section headings used herein are for organizational purposes only and are not to be used to construe the contents or meaning of the sections under the headings. Any use of the term “example” is intended to be an example of something that is not to be taken as limiting. Thus, other examples of the exemplary embodiments can have different values. It is noted that like references and descriptions set forth herein (throughout the several drawings) are directed to like items unless otherwise noted.

[0038] Embodiment 1

[0039] This embodiment takes the cloudy environment in Chengdu as an example, as shown in Figure 1 a specific method for land classification of remote sensing images in a cloudy environment is disclosed:

[0040] S1. Cloudy image processing: acquiring Sentinel-2 satellite images of a target area, and parsing the QA60 quality evaluation band thereof, generating a dynamic cloud mask layer through the binary bits representing clouds and cirrus in the band; performing cloud removal on the cloud mask layer to obtain a remote sensing image after cloud removal;

[0041] In the cloud removal on the mask layer, specifically:

[0042] Cloud classification and identification: identify the cloud and cirrus in the remote sensing image through the QA60 quality assessment band data (wherein the cloud mask parameter is 10, and the cirrus mask parameter is 11);

[0043] Clear mask extraction: perform a bitwise operation on the two kinds of clouds, and check whether the result is equal to 0. When the two kinds of clouds simultaneously meet the operation requirement, it indicates that the remote sensing image is in a clear state, and all clear regions are extracted to finally generate a cloud-free binary mask.

[0044] Time-series image synthesis: perform multi-temporal median synthesis on the clear mask image set, fill in the cloud-shielded area, perform surface reflectance conversion processing, and output the final processed image data set (wherein the time range screening parameter is from May 1 to December 31 of each year, and the cloud cover screening threshold is < 26%).

[0045] The cloud-removing before-and-after comparison chart is shown in Figure 2 , wherein a is the original cloud-containing image, and b is the result after dynamic cloud mask and time-series image synthesis processing. It can be seen that the data void area (white area) caused by cloud cover in the image a after cloud removal processing is effectively filled in, and the ground feature characteristics are revealed. Through the cloud classification and identification + clear mask extraction + time-series image synthesis technology, the effective image availability of the multi-cloud area is improved by 35%, the data filling accuracy of the cloud-shielded area is 89.2%, and the long-standing image void problem in the multi-cloud area is solved.

[0046] S2. Land feature index construction: constructing the land feature index of the cloud-removed remote sensing image in S1, and performing feature extraction;

[0047] Vegetation features: extracting through normalized vegetation index NDVI, enhanced vegetation index EVI and red edge position index REPI;

[0048] Water features: extracting through normalized water index NDWI and modified water index mNDWI;

[0049] Building features: extracting through building land index IBI;

[0050] Special surface features: extracting through bare soil index BSI;

[0051] Ice and snow features: extracting through normalized snow index NDSI.

[0052] As shown in Figure 3 , the comparison chart before and after feature extraction can be seen: the land feature index system constructed by the application can more clearly highlight the boundaries and spectral characteristics of different land classes such as vegetation, water, and buildings.

[0053] S3. Image recognition classification: based on the land feature index constructed in S2, multi-source ground object sample collection is performed to construct a training sample library for cloudy environment; a random forest algorithm is used and optimized random forest model parameters are applied for model training to obtain a land classification model for remote sensing images in cloudy environment.

[0054] The training sample library is constructed into a training set and a verification set at a ratio of 0.7:0.3, the training set is trained by using a random forest algorithm and applying optimized random forest model parameters to obtain a land classification model for remote sensing images in cloudy environment, specifically:

[0055] (1) Multi-source sample collection: 17 types of typical ground object samples (paddy field, dry land, greenhouse, arbor forest, shrub forest, sparse forest land, other forest land, grassland, river, beach, snow land, urban construction land, rural construction land, other construction land, bare land, rock, and other types) are collected to establish a geographic marker sample library;

[0056] (2) Spatial stratified sampling: the training set and the verification set are automatically divided at a ratio of 0.7:0.3;

[0057] (3) Adaptive weight distribution: the sample weight is dynamically adjusted according to the ground object distribution frequency, the principle is: the frequency of the ground object category in the training sample set is inversely weighted, the lower the frequency, the higher the weight of the sample in the training process;

[0058] (4) Model parameter optimization: the optimized random forest model parameters are: 190 decision trees, 6 feature quantities, 6 minimum sample numbers of nodes, 0.7 feature sampling ratio, 50 maximum node number, and 200 random seed, the overall accuracy and KAPPA coefficient are all 0.80 after training.

[0059] As shown in Figure 4 the training and recognition results of image classification can be seen: the sample label distribution and the land classification results have high consistency in spatial pattern.

[0060] Example 2 Precision verification

[0061] The model is verified by using the confusion matrix, the results show that the overall accuracy (Overall Accuracy) is 0.84 and the Kappa coefficient is 0.80, indicating that the model has high classification reliability.

[0062] To the extent not already already described herein, it will be apparent from this detailed description of embodiments of the application, which can be made and used without departing from the spirit and scope of the application, many other embodiments of the application will be readily ascertainable to those skilled in the art and the inventor contemplates the use of only the preferred embodiments of the application in the following claims.

Claims

1. A method for model construction for land classification of remote sensing images in a multi-cloud environment, characterized in that, It comprises the following steps: S1. Multi-cloud image processing: obtaining Sentinel-2 satellite images of the target area, and parsing the QA60 quality evaluation band thereof to generate a dynamic cloud mask layer through the binary bits representing clouds and cirrus in the band; performing cloud removal processing on the cloud mask layer to obtain a cloud-removed remote sensing image; S2. Land feature index construction: constructing a land feature index of the cloud-removed remote sensing image of S1, and performing feature extraction; S3. Image recognition and classification: based on the land feature index constructed in S2, collecting multi-element ground object samples to construct a training sample library for a multi-cloud environment; using a random forest algorithm and applying optimized random forest model parameters for model training to obtain a land classification model for remote sensing images in a multi-cloud environment.

2. The model building method of claim 1, wherein, The cloud removal processing of S1 comprises time-series image synthesis: performing multi-temporal median synthesis on the image set after cloud masking, and filling in the cloud-shielded area using time series.

3. The model building method of claim 1, wherein, The feature extraction of S2 comprises: Vegetation features: extracted by normalized difference vegetation index NDVI, enhanced vegetation index EVI, and red edge position index REPI; Water features: extracted by normalized water index NDWI and modified water index mNDWI; Building features: extracted by impervious building index IBI; Special surface features: extracted by bare soil index BSI; Ice and snow features: extracted by normalized difference snow index NDSI.

4. The model building method of claim 1, wherein, The multi-element ground object sample collection of S3 comprises a 17-class ground object sample system: paddy field, dry land, greenhouse, arbor forest, shrub forest, sparse forest land, other forest land, grassland, river, tidal flat, snow land, urban construction land, rural construction land, other construction land, bare land, rock, and other types.

5. The model building method of claim 1, wherein, The optimized random forest model parameters of S3 are: 190 decision trees, 6 feature numbers, 6 minimum sample numbers for nodes, 0.7 feature sampling ratio, 50 maximum node numbers, and random seed 200, and the overall accuracy and KAPPA coefficient reach 0.80 after training.

6. The model building method of claim 1, wherein, After the model is constructed, the overall accuracy and KAPPA coefficient of the confusion matrix are used to verify the model accuracy.

7. A land classification model for remote sensing images in a multi-cloud environment obtained by the construction method of any one of claims 1-6. 8.A method for classifying land in remote sensing images in a multi-cloud environment, characterized in that, The land classification model for remote sensing images in a multi-cloud environment of claim 7 is used for recognition and classification. 9.A land classification system for remote sensing images in a multi-cloud environment, characterized in that, It comprises: A multi-cloud image processing module: divided into a dynamic cloud mask algorithm module, a time-series image synthesis module, and a terrain data fusion module, which is used for cloud removal processing on a multi-cloud environment of a target area; A land feature index construction module: used for constructing land features according to the terrain of a target area; An image recognition and classification module: used for feature operation and output of land classification results.

10. A computer device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor is used to execute the land classification system for remote sensing images in a multi-cloud environment of claim 9.

Images