Remote sensing method for updating 1: 100,000 vegetation map
By integrating remote sensing, topographic, and climate information and employing a random forest machine learning model, the problems of blurred vegetation map boundaries and inaccurate classification in existing technologies have been solved, generating a high-precision 1:1,000,000 vegetation map that meets the needs of refined ecological management.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- 山西省地质环境监测和生态修复中心
- Filing Date
- 2025-12-31
- Publication Date
- 2026-05-01
AI Technical Summary
Existing technologies for preparing 1:1,000,000 vegetation maps suffer from subjective biases in boundary delineation, difficulty in accurately depicting complex vegetation transition zones, and low mapping efficiency, failing to meet the needs of refined ecological management. In particular, in areas with extremely complex topography and climate, there is a lack of clear schemes for the coordinated use of remote sensing, topography, and climate features, resulting in insufficient accuracy of vegetation maps.
By integrating remote sensing, topographic and climate information and employing a random forest machine learning model, a high-precision vegetation map is generated and the boundary is optimized through multi-dimensional feature set construction, feature selection and model training.
It achieves high-precision vegetation classification and boundary optimization, generating a highly reliable and accurate 1:1,000,000 vegetation map that can clearly identify small patches and vegetation transition boundaries, thus improving the objectivity and accuracy of mapping.
Smart Images
Figure CN121962897A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of remote sensing technology and vegetation mapping technology, and in particular to a remote sensing method for updating 1:1,000,000 vegetation maps. Background Technology
[0002] Vegetation maps are essential foundational geographic information for ecological research, environmental protection, and natural resource management. Currently, the "Vegetation Map of the People's Republic of China (1:1,000,000)" primarily relies on historical field surveys and manual delineation. This method suffers from subjective biases in boundary delineation, struggles to accurately depict complex vegetation transition zones, and is inefficient, failing to meet the demands of current refined ecological management. Satellite remote sensing technology, with its objective and efficient advantages, has become the mainstream method for vegetation mapping. Existing research generally relies on foreign satellite data such as Landsat, combined with machine learning algorithms like support vector machines and random forests for classification, which has improved automation to some extent. However, for special regions with extremely complex topography and climate, existing methods lack clear solutions on how to effectively coordinate and utilize multi-dimensional features such as remote sensing, topography, and climate, and optimize and filter them. This directly restricts further improvements in the accuracy of vegetation maps. Summary of the Invention
[0003] This invention aims to provide a remote sensing method for updating 1:1,000,000 vegetation maps. By fusing remote sensing, topographic and climate information and employing a random forest machine learning model, it achieves high-precision vegetation classification and boundary optimization, solving the problems of blurred boundaries and inaccurate classification in existing vegetation maps.
[0004] To achieve the above objectives, the present invention provides the following solution:
[0005] A remote sensing method for updating 1:1,000,000 vegetation maps includes:
[0006] Collect multi-source geospatial data of the study area; wherein, the multi-source geospatial data includes: remote sensing data, topographic data, and climate data;
[0007] Based on the multi-source geospatial data, samples are taken from the 1:1,000,000 vegetation map to be updated, according to vegetation type.
[0008] Construct a multidimensional feature set based on the sampled data;
[0009] A sampling recursive feature elimination method is used to filter features in the multidimensional feature set;
[0010] Based on the selected features, the random forest machine learning algorithm is used to train the model and evaluate it to obtain the optimal model.
[0011] The optimal model is used to predict the entire study area and generate an updated 1:1,000,000 vegetation map.
[0012] Optionally, the remote sensing data includes: several spectral bands;
[0013] The terrain data includes three types of terrain features: elevation, slope, and aspect.
[0014] The climate data refers to monthly data for the target period, including: minimum temperature, maximum temperature, precipitation, average temperature, potential evapotranspiration, annual extreme minimum temperature, annual extreme maximum temperature, annual precipitation, seasonal precipitation, annual average temperature, and seasonal average temperature.
[0015] Optionally, sampling by vegetation type from the 1:1,000,000 vegetation map to be updated includes:
[0016] Sampling points should avoid the boundary areas of vegetation types;
[0017] The total sample is randomly divided into training set and validation set according to the proportion.
[0018] Optionally, the multidimensional feature set includes: original reflectance features of remote sensing bands, vegetation index features, topographic features, and climate features.
[0019] Optionally, the vegetation index features include: normalized difference vegetation index, enhanced vegetation index, soil-regulated vegetation index, normalized difference water index, and bare soil index.
[0020] Optionally, the sampling recursive feature elimination method, which performs feature filtering on the multidimensional feature set, includes:
[0021] Using random forest as the base model, the multidimensional feature set is input into the model for training;
[0022] Based on the feature importance scores fed back by the model, the least important features are iteratively removed, and the effect of each removal is evaluated by the accuracy of the validation set, and finally an optimal feature subset is determined.
[0023] Optionally, model evaluation includes:
[0024] Using the validation set, the trained model is subjected to confusion matrix calculation to obtain multiple metrics for quantitative evaluation of each scheme; wherein, the multiple metrics include: overall accuracy, Kappa coefficient, producer accuracy, and user accuracy.
[0025] The beneficial effects of this invention are as follows:
[0026] This invention integrates remote sensing, topographic, and climate information, and employs a random forest machine learning model to achieve high-precision vegetation classification and boundary optimization, solving the problems of blurred boundaries and inaccurate classification in existing vegetation maps. It provides a complete technical solution for generating highly reliable and accurate 1:1,000,000 vegetation maps. Attached Figure Description
[0027] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0028] Figure 1 This is a schematic diagram of a remote sensing method for updating a 1:1,000,000 vegetation map according to an embodiment of the present invention.
[0029] Figure 2 This is an updated 1:1,000,000 vegetation map of the Qinghai-Tibet Plateau region, as per an embodiment of the present invention. Detailed Implementation
[0030] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0031] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0032] like Figure 1 As shown, this embodiment proposes a remote sensing method for updating 1:1,000,000 vegetation maps, including:
[0033] Collect multi-source geospatial data of the study area; wherein, the multi-source geospatial data includes: remote sensing data, topographic data, and climate data;
[0034] Based on the multi-source geospatial data, samples are taken from the 1:1,000,000 vegetation map to be updated, according to vegetation type.
[0035] Construct a multidimensional feature set based on the sampled data;
[0036] A sampling recursive feature elimination method is used to filter features in the multidimensional feature set;
[0037] Based on the selected features, the random forest machine learning algorithm is used to train the model and evaluate it to obtain the optimal model.
[0038] The optimal model is used to predict the entire study area and generate an updated 1:1,000,000 vegetation map.
[0039] Furthermore, in this embodiment, the multi-source geospatial data acquisition specifically includes: remote sensing data from Landsat 5 TM series images from 1980 to 1995, which contain six spectral bands: blue, green, red, near-infrared, shortwave infrared 1, and shortwave infrared 2, with a spatial resolution of 30 meters; topographic data originating from the digital elevation model of the Space Shuttle Radar Topographic Mapping Mission, from which three types of topographic features—elevation, slope, and aspect—are extracted; and climate data acquired is a monthly dataset covering the target period, including minimum temperature, maximum temperature, precipitation, average temperature, and potential evapotranspiration, and further calculations of 13 climate indicators, including annual extreme minimum temperature, annual extreme maximum temperature, annual precipitation, seasonal precipitation, annual average temperature, and seasonal average temperature.
[0040] Furthermore, in this embodiment, sample sampling specifically includes: systematically sampling from the 1:1,000,000 vegetation map to be updated, according to vegetation type, to obtain a sufficient number of sample points. To ensure sample purity, sampling points need to avoid the boundary areas of vegetation types. The total samples are randomly divided into training and validation sets proportionally.
[0041] Feature construction must be completed synchronously on both the training and validation sets, following the same rules. Feature selection involves training the model using different feature combinations based on the training set data. By comparing the overall accuracy of each feature combination on the validation set, the feature combination that optimizes the model's generalization ability is selected. The validation set serves only as a performance benchmark and is not involved in the feature selection process.
[0042] Furthermore, in this embodiment, feature construction specifically includes:
[0043] Based on the above original data sources (remote sensing, topography, and climate data), a multidimensional feature set containing 39 features is constructed, specifically including: 6 original reflectance features of remote sensing bands; 17 vegetation index features, including but not limited to normalized difference vegetation index, enhanced vegetation index, soil-regulated vegetation index, normalized difference water index, bare soil index, etc.; 3 topographic features; and 13 climate features.
[0044] Furthermore, in this embodiment, feature selection specifically includes:
[0045] The Recursive Feature Elimination (RFE) method is adopted, using Random Forest as the base model. Based on the feature weights or importance scores generated by it, multiple rounds of iterative training are performed, eliminating the least important features each time. Finally, the optimal feature subset for different models is determined based on the overall accuracy.
[0046] Furthermore, in this embodiment, model training specifically includes:
[0047] Using the same training set, the model was trained using a random forest machine learning algorithm based on features selected from remote sensing, terrain, and climate characteristics.
[0048] Furthermore, in this embodiment, model evaluation and selection specifically include:
[0049] Using the validation set, the confusion matrix is calculated to obtain multiple indicators such as overall accuracy, Kappa coefficient, producer accuracy, and user accuracy, and to quantitatively evaluate each scheme.
[0050] Furthermore, in this embodiment, vegetation map generation specifically includes:
[0051] The selected optimal model was applied to predict the entire study area, generating an updated 1:1,000,000 vegetation map. The new map was spatially overlaid and visually compared with the existing vegetation map on concurrent Landsat remote sensing imagery, focusing on verifying its improvement effects on small patch recognition, the naturalness of water and glacier boundaries, and the details of vegetation-ecological transition zones.
[0052] The following embodiment uses the Qinghai-Tibet Plateau as an example to explain the implementation steps in detail:
[0053] Step 1: Data Acquisition and Processing
[0054] This invention first systematically collects Landsat 5 TM remote sensing imagery from 1980 to 1995, and performs radiometric calibration, atmospheric correction, and geometric fine correction. Simultaneously, it acquires SRTM digital elevation model data, extracting three topographic factors: elevation, slope, and aspect. Furthermore, it collects monthly climate datasets including temperature, precipitation, and potential evapotranspiration, calculating 13 climate indicators, including annual extreme temperature, seasonal precipitation, and average temperature, based on data from 1986 to 1995. All data are unified to the same coordinate system and spatial resolution to ensure data consistency and provide a reliable foundation for subsequent analysis.
[0055] Step 2: Sample Data Preparation
[0056] From the original 1:1,000,000 vegetation map digitization results, a stratified random sampling method was used to collect 13,038 clean sample points while ensuring that the sample points were far from the vegetation type boundaries. Subsequently, all samples were randomly divided into a training set and a validation set in a 7:3 ratio. The training set contained 9,126 samples for model training, and the validation set contained 3,912 samples for accuracy evaluation.
[0057] Step 3: Feature Construction
[0058] Based on the prepared data, a multidimensional feature system comprising 39 features was constructed. This system integrates six Landsat raw band reflectances, 17 vegetation indices (including NDVI, EVI, SAVI, etc.), three topographic factors, and 13 climatic features. These features comprehensively cover the spectral characteristics of land features, topographic environment, and regional climate background, forming a feature space that can fully describe the distribution patterns of vegetation.
[0059] Step 4: Feature Filtering
[0060] A recursive feature elimination method is used for feature optimization. Specifically, a random forest is used as the base model, and all 39 features are input into the model for training. Based on the feature importance scores provided by the model, the least important features are iteratively eliminated. The effectiveness of each elimination is evaluated using the validation set accuracy, ultimately determining an optimal feature subset.
[0061] Step 5: Model Training
[0062] Using the training set samples prepared in step two, the random forest machine learning algorithm is used to train the model.
[0063] Step Six: Model Evaluation and Selection
[0064] Using the validation set reserved in step two, the model trained in step five is quantitatively evaluated. By calculating the confusion matrix, its overall accuracy is found to be 89.03%, and its Kappa coefficient is 0.88.
[0065] Step 7: Vegetation Map Generation and Verification
[0066] The optimal random forest model selected in step six is applied to the entire Qinghai-Tibet Plateau region to predict vegetation type for each pixel, thereby generating an updated 1:1,000,000 vegetation map. Results show that the vegetation map generated in this invention optimizes the 1:1,000,000 vegetation map to a certain extent, clearly identifying small patches, accurately extracting wavy water body boundaries, and accurately presenting mosaic-distributed vegetation transition boundaries. Optimized details include... Figure 2 As shown in the figure, this result is highly consistent with the actual surface conditions, providing more accurate and reliable map data support for vegetation research and ecological protection.
[0067] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.
Claims
1. A remote sensing method for updating 1:1,000,000 vegetation maps, characterized in that, include: Collect multi-source geospatial data of the study area; wherein, the multi-source geospatial data includes: remote sensing data, topographic data, and climate data; Based on the multi-source geospatial data, samples are taken from the 1:1,000,000 vegetation map to be updated, according to vegetation type. Construct a multidimensional feature set based on the sampled data; A sampling recursive feature elimination method is used to filter features in the multidimensional feature set; Based on the selected features, the random forest machine learning algorithm is used to train the model and evaluate it to obtain the optimal model. The optimal model is used to predict the entire study area and generate an updated 1:1,000,000 vegetation map.
2. The remote sensing method for updating a 1:1,000,000 vegetation map according to claim 1, characterized in that, The remote sensing data includes: several spectral bands; The terrain data includes three types of terrain features: elevation, slope, and aspect. The climate data refers to monthly data for the target period, including: minimum temperature, maximum temperature, precipitation, average temperature, potential evapotranspiration, annual extreme minimum temperature, annual extreme maximum temperature, annual precipitation, seasonal precipitation, annual average temperature, and seasonal average temperature.
3. The remote sensing method for updating a 1:1,000,000 vegetation map according to claim 1, characterized in that, Sampling by vegetation type from the 1:1,000,000 vegetation map to be updated includes: Sampling points should avoid the boundary areas of vegetation types; The total sample is randomly divided into training set and validation set according to the proportion.
4. The remote sensing method for updating a 1:1,000,000 vegetation map according to claim 1, characterized in that, The multidimensional feature set includes: original reflectance features of remote sensing bands, vegetation index features, topographic features, and climate features.
5. The remote sensing method for updating a 1:1,000,000 vegetation map according to claim 4, characterized in that, The vegetation index features include: normalized difference vegetation index, enhanced vegetation index, soil-regulated vegetation index, normalized difference water index, and bare soil index.
6. The remote sensing method for updating a 1:1,000,000 vegetation map according to claim 3, characterized in that, The sampling recursive feature elimination method, which performs feature filtering on the multidimensional feature set, includes: Using random forest as the base model, the multidimensional feature set is input into the model for training; Based on the feature importance scores fed back by the model, the least important features are iteratively removed, and the effect of each removal is evaluated by the accuracy of the validation set, and finally an optimal feature subset is determined.
7. The remote sensing method for updating a 1:1,000,000 vegetation map according to claim 3, characterized in that, Model evaluation includes: Using the validation set, the trained model is subjected to confusion matrix calculation to obtain multiple metrics for quantitative evaluation of each scheme; wherein, the multiple metrics include: overall accuracy, Kappa coefficient, producer accuracy, and user accuracy.