Grassland remote sensing resource automatic classification method, device and equipment and medium

Abstract

This invention relates to an automatic classification method, apparatus, equipment, and medium for grassland remote sensing resources, belonging to the fields of remote sensing image processing and machine learning. The method employs a three-layer classification framework, combining ground sample plots, remote sensing data, meteorological data, and DSM elevation data to progressively distinguish between grassland and non-grassland, coarsely classify grassland types (grassland, desert, shrubland, meadow, and marsh), and refine them into 18 grassland resource categories. Gaussian low-pass filtering is used for noise reduction, weighted pooling generates multi-scale feature maps, gated convolution fills in missing regions, and an attention mechanism enhances the spatial autocorrelation of geographical features. Random forest and convolutional neural networks are compared and optimized, and the optimal model is fused. Accuracy is evaluated based on UAV data and grassland resource maps, calculating the overall macro-average accuracy, user accuracy, producer accuracy, F1 index, and Kappa coefficient. This invention achieves high-precision grassland resource classification and is suitable for ecological monitoring and management of grassland resources.

Description

Technical Field

[0001] This invention relates to the fields of remote sensing image processing and machine learning technology, and in particular to a method, apparatus, device and medium for automatic classification of grassland remote sensing resources. Background Technology

[0002] Grasslands serve multiple functions, including windbreak, sand fixation, air purification, water conservation, prevention of soil erosion, and protection of the ecological environment. Grassland resources are an important component of the five essential elements of natural resources used in production activities to create material wealth. They are the material basis for human grassland animal husbandry, a crucial condition for national economic development and improving people's lives, and a significant source of social wealth. Grassland resources encompass the quantity of grassland resources, the size of grassland area, grassland yield, carrying capacity, primary productivity, quality of grassland forage, and the distribution pattern and composition of various grassland types in a three-dimensional space composed of latitude, longitude, and altitude.

[0003] With the development of remote sensing technology, existing grassland classification methods mainly rely on expert experience to construct remote sensing index thresholds or adopt shallow machine learning models.

[0004] However, this approach suffers from limitations in feature representation, insufficient utilization of spatial context information, and difficulties in fusing multi-source heterogeneous data. Existing deep learning methods still have significant shortcomings in handling multi-scale feature coupling under complex plateau terrain, long-distance spatial correlation modeling, and collaborative analysis of multi-temporal remote sensing and static environmental factors, resulting in low accuracy in classifying small land features, severe inter-class confusion, and weak model generalization ability. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a method, apparatus, equipment and medium for automatic classification of grassland remote sensing resources, aiming to solve at least one of the above-mentioned technical problems.

[0006] The technical solution of the present invention to solve the above-mentioned technical problems is as follows:

[0007] Firstly, this application provides an automatic classification method for grassland remote sensing resources, employing the following technical solution:

[0008] An automatic classification method for grassland remote sensing resources includes:

[0009] Acquire multi-source raw data of the target area. The multi-source raw data includes remote sensing data, meteorological data, ground quadrat data and elevation data. The ground quadrat data includes latitude and longitude, grassland type and dominant species information. The meteorological data includes annual average temperature and annual average precipitation.

[0010] Feature extraction is performed on the multi-source raw data to obtain target multi-source fusion feature data, which includes various remote sensing indices, topographic indices, and meteorological indices.

[0011] The target multi-source fusion feature data is input into a pre-constructed multi-level classification model to obtain a grassland resource type classification map of the target area;

[0012] The pre-built multi-level classification model is trained using a progressive three-level framework based on historical multi-source fusion feature data and corresponding grassland type labels. The pre-built multi-level classification model includes a first-level classification model, a second-level classification model, and a third-level classification model. The first-level classification model is used to distinguish between grassland and non-grassland areas and generate a grassland mask. The second-level classification model is used to perform primary grassland type classification within grassland areas. The third-level classification model is used to further subdivide secondary grassland types based on the primary grassland types.

[0013] The beneficial effects of this invention are: by acquiring multi-source raw data and extracting features, it can generate target multi-source fusion feature data containing multiple indices, providing a comprehensive and accurate data foundation for classification; by using a progressive three-level framework to train a multi-level classification model, and by distinguishing grassland from non-grassland through each layer of the model, performing first-level grassland type classification and second-level grassland type subdivision, it can reduce classification complexity, reduce confusion, improve the accuracy of grassland resource type classification, and finally obtain an accurate grassland resource type classification map of the target area.

[0014] Based on the above technical solution, the present invention can be further improved as follows.

[0015] Furthermore, the step of extracting features from the multi-source raw data to obtain target multi-source fused feature data includes:

[0016] Spatial registration and rasterization are performed on the acquired remote sensing data, meteorological data and elevation data to obtain the preprocessed first remote sensing data, first meteorological data and first elevation data;

[0017] The first remote sensing data is subjected to radiometric calibration, atmospheric correction and quality control to obtain the preprocessed second remote sensing data.

[0018] Based on ground quadrat data, areas with sampling points in the target area are marked as valid areas, and areas without sampling points are marked as invalid areas, thus generating a mask for the valid areas.

[0019] Based on the preprocessed second remote sensing data, a variety of remote sensing indices for the target area are calculated, including the normalized vegetation index, soil-regulated vegetation index, ratio vegetation index, humidity index, bare soil index, and normalized water index.

[0020] Based on the preprocessed first elevation data, the topographic index of the target area is extracted, and the topographic index includes elevation, slope and aspect.

[0021] Based on the preprocessed first meteorological data, the meteorological index of the target area is extracted. The meteorological index includes the annual average temperature, annual total precipitation, accumulated temperature index and growing season temperature-related days index.

[0022] The various remote sensing indices, topographic indices, and meteorological indices are superimposed on the effective area mask to generate initial multi-source fusion feature data;

[0023] The initial multi-source fusion feature data is denoised based on a preset Gaussian low-pass filter model to obtain preprocessed multi-source fusion feature data. The cutoff frequency parameter of the Gaussian low-pass filter model is set based on the classification level and grassland type.

[0024] Feature filtering is performed on the preprocessed multi-source fusion feature data based on recursive feature elimination, and optimization is performed based on weighted pooling operators to obtain the target multi-source fusion feature data.

[0025] The beneficial effects of adopting the above-mentioned further scheme are as follows: preprocessing of multi-source raw data, such as spatial registration, rasterization, radiometric calibration, atmospheric correction, and quality control, can improve data quality; generating effective region masks can eliminate prior noise in invalid regions; calculating various remote sensing indices, extracting topographic indices and meteorological indices, and superimposing them to generate initial multi-source fusion feature data can obtain more comprehensive feature information; using a Gaussian low-pass filter model for denoising can reduce data noise; feature selection based on recursive feature elimination and optimization of weight pooling operators can reduce feature dimensionality and improve data feature quality, thereby providing higher-quality input data for subsequent multi-level classification models and improving the accuracy of grassland resource type classification.

[0026] Furthermore, the step of inputting the multi-source fusion feature data into a pre-constructed multi-level classification model to obtain a grassland resource type classification map of the target area includes:

[0027] The multi-source fusion feature data is input into the first-layer classification model to obtain the first-layer classification result of the target region. The first-layer classification result includes grassland masking and non-grassland masking.

[0028] The grassland mask and the multi-source fusion feature data are input into the second-layer classification model to obtain the second-layer classification result of the target area. The second-layer classification result includes the primary grassland type, which includes grassland, desert, shrubland, meadow, swamp, cultivated land and terraces.

[0029] The second-layer classification result and the multi-source fusion feature data are input into the third-layer classification model to obtain the third-layer classification result. The third-layer classification result includes secondary grassland types, which include temperate meadow steppe, temperate steppe, temperate desert steppe, alpine meadow steppe, alpine steppe, alpine desert steppe, temperate steppe desert, temperate desert, alpine desert, tropical grassland, tropical shrub grassland, warm grassland, warm shrub grassland, lowland meadow, temperate mountain meadow, alpine meadow, and marsh.

[0030] The classification results of the secondary grassland types are merged into seven major categories: tropical grassland, warm grassland, meadow steppe, alpine meadow, typical steppe, alpine steppe and desert, generating a grassland resource type classification map of the target area.

[0031] The beneficial effects of adopting the above-mentioned further scheme are as follows: Multi-source fusion feature data is input into a multi-level classification model. First, the first-layer classification model is used to distinguish grassland and non-grassland areas, generating corresponding masks to narrow down the scope for subsequent classification. Then, based on the first-layer grassland mask, the second-layer classification model is used to classify the grassland areas into primary grassland types, obtaining results containing multiple primary grassland types. Next, based on the second-layer classification results, the third-layer classification model is used to further subdivide secondary grassland types. Finally, the secondary grassland types are merged into seven major categories to generate a grassland resource type classification map of the target area. This achieves multi-level classification and macro-level summarization of grassland resource types in the target area, meeting the needs of grassland resource analysis at different levels.

[0032] Furthermore, the method for constructing the pre-built multi-level classification model includes:

[0033] Obtain historical multi-source fusion feature data of the target area, and construct training and testing sets based on the historical multi-source fusion feature data;

[0034] For any level of classification model in a multi-level classification model, the training set and test set are cropped based on the mask output by the classification model of the previous level of the current level, respectively, to obtain the cropped training set and test set. The first level classification model crops the training set and test set based on the initial mask, which is a set mask.

[0035] For any layer of the multi-level classification model, feature filtering is performed on the pruned training and test sets through recursive feature elimination and random forest evaluator to obtain the feature-filtered training and test sets.

[0036] For any layer of the multi-level classification model, the pre-constructed random forest model is trained based on the training set after grid search, cross-validation optimization and feature selection, and the RF hyperparameters are optimized to obtain the trained random forest model.

[0037] For any layer of the multi-level classification model, multi-scale features of the cropped training set are generated based on weight pooling, and the CNN model is trained through encoder-decoder structure, gated convolution and attention mechanism to obtain the trained CNN model.

[0038] For any layer of the multi-level classification model, the trained random forest model and the trained CNN model are evaluated based on the pruned test set, and the optimal algorithm or algorithm fusion scheme is selected as the final classification model for the current level based on the predetermined evaluation index.

[0039] For any layer of the multi-level classification model, the grassland resources of the target area are predicted based on the final classification model of the current layer, the classification result of the current layer is generated, and it is transformed into the input mask required for the next layer, until the training of all layers of the model is completed, and the trained multi-level classification model is obtained.

[0040] The beneficial effects of adopting the above-mentioned further solutions are as follows: Obtaining historical multi-source fusion feature data of the target region and constructing training and test sets provides a data foundation for model training and evaluation; using the mask output by the previous-level classification model to prune the training and test sets reduces interference from irrelevant data and improves computational efficiency; feature selection through recursive feature elimination and random forest evaluators reduces feature dimensionality, eliminates redundant features, and improves model efficiency; training the random forest model based on grid search and cross-validation optimization avoids the blindness of manual parameter tuning and ensures generalization ability; generating multi-scale features based on weight pooling and training the CNN model through encoder-decoder structure, gated convolution, and attention mechanisms improves robustness to missing data and feature representation ability, reducing training difficulty; and evaluating the random forest model and CNN based on the test set... The model selects the optimal algorithm or algorithm fusion scheme to balance accuracy and efficiency and improve the robustness of the model. The classification results of the current level are transformed into the input mask required for the next level to realize hierarchical progressive training, which can reduce the number of categories in a single layer, reduce classification confusion, improve fine-grained accuracy, and finally obtain a multi-level classification model after training to accurately classify grassland resources.

[0041] Furthermore, the process of evaluating the trained random forest model and the trained CNN model based on the pruned test set, and selecting the optimal algorithm or algorithm fusion scheme as the final classification model for the current level based on predetermined evaluation metrics, includes:

[0042] Based on the test set, the macro-average evaluation index of the trained random forest model and the trained CNN model are calculated respectively. The macro-average evaluation index includes overall accuracy, user accuracy, producer accuracy, F1 index and Kappa coefficient.

[0043] If the F1 index of the random forest model is higher than that of the CNN model, and the difference in F1 index exceeds a first preset threshold, then the random forest model is selected as the final classification model for the current level.

[0044] If the F1 index of the CNN model is higher than that of the random forest model, and the difference in F1 index exceeds the first preset threshold, then the CNN model is selected as the final classification model for the current level.

[0045] If the difference in F1 index between the random forest model and the CNN model is less than or equal to the first preset threshold, a weighted fusion scheme is used to fuse the CNN model and the random forest model into the final classification model of the current level.

[0046] The probability calculation formula for weighted fusion in the weighted fusion scheme is as follows:

[0047] P=α×P RF +(1-α)×P CNN , where P RF P represents the predicted probability output by the random forest model. CNN The weight α is the predicted probability output by the convolutional neural network model, and it is determined by cross-validation optimization on the validation set.

[0048] The beneficial effects of adopting the above-mentioned further scheme are as follows: by calculating the macro-average evaluation index of the random forest model and the CNN model, and selecting a single model as the final classification model based on the F1 index difference and the first preset threshold, or by adopting a weighted fusion scheme to fuse the two, the advantages of the two models can be fully utilized, the classification accuracy and robustness of the model can be improved, and the generalization ability and accuracy of the model can be improved by optimizing the weights through cross-validation.

[0049] Furthermore, the pre-built random forest model is trained on the training set based on grid search, cross-validation optimization, and feature selection to optimize the RF hyperparameters and obtain the trained random forest model, including:

[0050] Based on the grid search strategy, all parameter combinations in the hyperparameter search space of the random forest model are traversed. The parameter combinations include the number of decision trees, the maximum depth of the tree, the minimum number of samples for splitting internal nodes, the minimum number of samples for leaf nodes, the maximum number of features considered when splitting, and whether to use bootstrap sampling.

[0051] For each hyperparameter combination, the model is trained and validated on the training set after feature selection based on the k-fold cross-validation algorithm to obtain the macro-average F1 score evaluation result of each hyperparameter combination under cross-validation.

[0052] The combination of hyperparameters that achieves the highest macro-average F1 score in cross-validation is selected as the optimal hyperparameters.

[0053] Based on the optimal hyperparameters and the entire training set after feature selection, the random forest model is retrained to obtain the trained random forest model.

[0054] The beneficial effects of adopting the above-mentioned further scheme are as follows: By using a grid search strategy to traverse all parameter combinations within the hyperparameter search space of the random forest model, and combining this with the k-fold cross-validation algorithm for training and validation on the feature-selected training set, the macro-average F1 score evaluation result of each hyperparameter combination under cross-validation can be obtained, thereby selecting the optimal hyperparameters. Retraining the random forest model based on the optimal hyperparameters and the entire training set avoids the blindness of manual parameter tuning, ensures the model's generalization ability, improves the performance of the random forest model, and enhances the accuracy of grassland resource remote sensing classification.

[0055] Furthermore, the multi-scale features of the training set after feature selection are generated based on weight pooling, and the CNN model is trained using an encoder-decoder structure, gated convolution, and attention mechanism to obtain the trained CNN model, including:

[0056] The cropped training set is downsampled based on the weighted pooling operator to generate feature maps at multiple scales. The weighted pooling operator is calculated by performing weighted average pooling on the effective regions of the image and setting the weight of the invalid regions to 0.

[0057] The encoder performs convolution operations on the feature maps at each scale to extract deep features from the multi-scale feature maps, resulting in an encoded multi-scale deep feature map. The encoder is composed of multiple convolutional layers, activation function layers, and pooling layers stacked together.

[0058] Calculate the gate value of the encoded multi-scale deep feature map, and dynamically select and fill the feature map based on the gate value to obtain the filled feature map;

[0059] Based on historical multi-source fusion feature data, the attention weights of geographic environmental features are calculated to obtain an attention weight map. The geographic environmental features include the geographic location, elevation, topographic relief, and classification trend of the pixels.

[0060] Based on the padded feature maps and attention weight maps, the decoder uses an upsampling operation to restore the multi-scale feature maps extracted by the encoder to their original resolution, and then fuses the outputs of the gated convolution and attention mechanism, outputting the trained CNN model through the Softmax activation function.

[0061] The beneficial effects of adopting the above-mentioned further scheme are as follows: By downsampling the cropped training set based on the weighted pooling operator to generate feature maps at multiple scales, noise and missing data in remote sensing images can be processed, improving the sensitivity to grassland resources; by performing convolution operations on the multi-scale feature maps based on the encoder to extract deep features, the trends and details of grassland resources can be captured; by calculating the gating values ​​of the encoded multi-scale deep feature maps and performing dynamic selection and filling, the robustness to missing data can be improved; by calculating the attention weights of the geographic environment features to obtain the attention weight map, the third law of geography can be reflected, noise interference can be reduced, and the feature expression ability can be improved; by using the decoder to restore the multi-scale feature maps to the original resolution and fusing the output of the gated convolution and attention mechanisms, low-level details and high-level semantics can be fused, nonlinear expression can be improved, gradient flow can be optimized, end-to-end training can be achieved, and finally the trained CNN model can be obtained, which is suitable for multi-scale grassland resource remote sensing classification and can be used for dynamic monitoring and decision support.

[0062] Furthermore, it also includes:

[0063] Obtain third-party verification data, including UAV imagery data and historical grassland resource map data;

[0064] The third-party verification data is preprocessed to obtain preprocessed third-party verification data. The preprocessed third-party verification data is consistent with the grassland resource type classification map output by the multi-level classification model in terms of spatial coordinate system and resolution.

[0065] Based on the UAV imagery data, the secondary grassland types of the grassland resource type classification map are locally verified to obtain the first verification result;

[0066] Based on the historical grassland resource map data, a macro-consistency verification of the merged types of the grassland resource type classification map is performed to obtain a second verification result;

[0067] Based on the first verification result and the second verification result, a multi-dimensional evaluation index of macro average of secondary grassland types and merged types is calculated. The evaluation index includes total accuracy, user accuracy, producer accuracy, F1 index and Kappa coefficient.

[0068] Based on the preset regional division rules, the spatial autocorrelation index of the grassland resource type classification map and the classification accuracy index within each zone are calculated to obtain the spatial assessment results and regional classification deviation.

[0069] The consistency between the grassland resource type classification map and the third-party verification data was analyzed based on statistical testing methods, and the test results were obtained.

[0070] Based on the first verification result, the second verification result, multi-dimensional evaluation indicators, spatial assessment results, regional classification bias, and test results, easily confused grassland type pairs are identified, the causes of classification errors are analyzed, and the analysis results are fed back into the training process of the multi-level classification model to optimize model parameters or feature selection.

[0071] The beneficial effects of adopting the above-mentioned further solutions are as follows: By acquiring and preprocessing third-party validation data, it can be ensured that the data is consistent with the classification map in terms of spatial coordinate system and resolution, providing a reliable foundation for subsequent validation; by using UAV imagery data for local validation and historical grassland resource map data for macro-consistency validation, the accuracy of the classification map can be tested from different scales; by calculating the multi-dimensional evaluation index of macro-average, the accuracy of the classification results can be comprehensively measured; by calculating the spatial autocorrelation index and the regional classification accuracy index, the spatial rationality and regional deviation of the classification map can be identified; by analyzing consistency through statistical tests, the degree of fit between the classification map and the third-party data can be determined; by identifying easily confused grassland type pairs and analyzing the causes of errors, the results can be fed back to the model training process, which can optimize model parameters or feature selection, and improve model performance and classification accuracy.

[0072] Secondly, this application provides an automatic classification device for grassland remote sensing resources, which adopts the following technical solution:

[0073] An automatic classification device for grassland remote sensing resources includes:

[0074] The acquisition module is used to acquire multi-source raw data of the target area. The multi-source raw data includes remote sensing data, meteorological data, ground quadrat data and elevation data. The ground quadrat data includes latitude and longitude, grassland type and dominant species information, and the meteorological data includes annual average temperature and annual average precipitation.

[0075] The feature extraction module is used to extract features from the multi-source raw data to obtain multi-source fused feature data, which includes spectral index, soil index, water body index, topographic index, and meteorological index.

[0076] The classification module is used to input the multi-source fusion feature data into a pre-constructed multi-level classification model to obtain a grassland resource type classification map of the target area;

[0077] The pre-built multi-level classification model is trained using a progressive three-level framework based on historical multi-source fusion feature data and corresponding grassland type labels. The pre-built multi-level classification model includes a first-level classification model, a second-level classification model, and a third-level classification model. The first-level classification model is used to distinguish between grassland and non-grassland areas and generate a grassland mask. The second-level classification model is used to perform primary grassland type classification within grassland areas. The third-level classification model is used to further subdivide secondary grassland types based on the primary grassland types.

[0078] Thirdly, this application provides an electronic device that adopts the following technical solution:

[0079] An electronic device includes a memory and a processor, wherein the memory stores a computer program capable of being loaded by the processor and executing the automatic classification method for grassland remote sensing resources as described in any of the first aspects.

[0080] Additional aspects and advantages of this application will be set forth in part in the description which follows, and will become apparent from the description or may be learned by practice of this application. Attached Figure Description

[0081] Figure 1 This is a flowchart illustrating an automatic classification method for grassland remote sensing resources according to an embodiment of the present invention.

[0082] Figure 2 This is a schematic diagram illustrating the process of inputting the multi-source fusion feature data into a pre-constructed multi-level classification model, as provided in one embodiment of the present invention.

[0083] Figure 3 A comparison chart of classification results provided in one embodiment of the present invention;

[0084] Figure 4 A schematic diagram of the structure of an automatic classification device for grassland remote sensing resources provided in an embodiment of the present invention;

[0085] Figure 5 This is a schematic diagram of the structure of an electronic device provided in one embodiment of the present invention. Detailed Implementation

[0086] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0087] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article, unless otherwise specified, generally indicates that the preceding and following related objects have an "or" relationship.

[0088] This application provides an automatic classification method for grassland remote sensing resources. This method can be executed by an electronic device, which can be a server or a mobile terminal device. The server can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides cloud computing services. The mobile terminal device can be a laptop computer, a desktop computer, etc., but is not limited to these.

[0089] like Figure 1 As shown, an automatic classification method for grass remote sensing resources includes:

[0090] S1, acquire multi-source raw data of the target area. The multi-source raw data includes remote sensing data, meteorological data, ground quadrat data and elevation data. The ground quadrat data includes latitude and longitude, grassland type and dominant species information. The meteorological data includes annual average temperature and annual average precipitation.

[0091] S2, extract features from the multi-source raw data to obtain target multi-source fusion feature data, which includes various remote sensing indices, terrain indices, and meteorological indices;

[0092] S3, input the target multi-source fusion feature data into a pre-constructed multi-level classification model to obtain a grassland resource type classification map of the target area;

[0093] The pre-built multi-level classification model is trained using a progressive three-level framework based on historical multi-source fusion feature data and corresponding grassland type labels. The pre-built multi-level classification model includes a first-level classification model, a second-level classification model, and a third-level classification model. The first-level classification model is used to distinguish between grassland and non-grassland areas and generate a grassland mask. The second-level classification model is used to perform primary grassland type classification within grassland areas. The third-level classification model is used to further subdivide secondary grassland types based on the primary grassland types.

[0094] In this embodiment, the multi-source raw data acquisition section includes acquiring remote sensing data, meteorological data, ground quadrat data, and elevation data. Remote sensing data can be multispectral satellite data, such as Landsat and Sentinel, which provide spectral information in different bands, reflecting the vegetation growth status of grasslands. Annual average temperature and annual average precipitation in meteorological data have a significant impact on grassland growth and distribution; different temperature and precipitation conditions are suitable for different types of grassland growth. Latitude and longitude information in ground quadrat data can determine the location of the quadrat, while grassland type and dominant species information provide practical reference standards for classification. Elevation data can be digital elevation model (DEM) data, which reflects the topographic relief; different topographic conditions affect grassland moisture, light, and other conditions, thus influencing grassland type. The remote sensing data can also be replaced by hyperspectral satellite data, which provides more refined spectral information; meteorological data can also incorporate more meteorological elements, such as wind speed and sunshine duration.

[0095] Optionally, feature extraction is performed on the multi-source raw data to obtain target multi-source fused feature data, including:

[0096] Spatial registration and rasterization are performed on the acquired remote sensing data, meteorological data and elevation data to obtain the preprocessed first remote sensing data, first meteorological data and first elevation data;

[0097] The first remote sensing data is subjected to radiometric calibration, atmospheric correction and quality control to obtain the preprocessed second remote sensing data.

[0098] Based on ground quadrat data, areas with sampling points in the target area are marked as valid areas, and areas without sampling points are marked as invalid areas, thus generating a mask for the valid areas.

[0099] Based on the preprocessed second remote sensing data, a variety of remote sensing indices for the target area are calculated, including the normalized vegetation index, soil-regulated vegetation index, ratio vegetation index, humidity index, bare soil index, and normalized water index.

[0100] Based on the preprocessed first elevation data, the topographic index of the target area is extracted, and the topographic index includes elevation, slope and aspect.

[0101] Based on the preprocessed first meteorological data, the meteorological index of the target area is extracted. The meteorological index includes the annual average temperature, annual total precipitation, accumulated temperature index and growing season temperature-related days index.

[0102] The various remote sensing indices, topographic indices, and meteorological indices are superimposed on the effective area mask to generate initial multi-source fusion feature data;

[0103] The initial multi-source fusion feature data is denoised based on a preset Gaussian low-pass filter model to obtain preprocessed multi-source fusion feature data. The cutoff frequency parameter of the Gaussian low-pass filter model is set based on the classification level and grassland type.

[0104] Feature filtering is performed on the preprocessed multi-source fusion feature data based on recursive feature elimination, and optimization is performed based on weighted pooling operators to obtain the target multi-source fusion feature data.

[0105] Optional, such as Figure 2 As shown, the multi-source fusion feature data is input into a pre-constructed multi-level classification model to obtain a grassland resource type classification map of the target area, including:

[0106] The multi-source fusion feature data is input into the first-layer classification model to obtain the first-layer classification result of the target region. The first-layer classification result includes grassland masking and non-grassland masking.

[0107] The grassland mask and the multi-source fusion feature data are input into the second-layer classification model to obtain the second-layer classification result of the target area. The second-layer classification result includes the primary grassland type, which includes grassland, desert, shrubland, meadow, swamp, cultivated land and terraces.

[0108] The second-layer classification result and the multi-source fusion feature data are input into the third-layer classification model to obtain the third-layer classification result. The third-layer classification result includes secondary grassland types, which include temperate meadow steppe, temperate steppe, temperate desert steppe, alpine meadow steppe, alpine steppe, alpine desert steppe, temperate steppe desert, temperate desert, alpine desert, tropical grassland, tropical shrub grassland, warm grassland, warm shrub grassland, lowland meadow, temperate mountain meadow, alpine meadow, and marsh.

[0109] The classification results of the secondary grassland types are merged into seven major categories: tropical grassland, warm grassland, meadow steppe, alpine meadow, typical steppe, alpine steppe and desert, generating a grassland resource type classification map of the target area.

[0110] In this embodiment, ground quadrat data, remote sensing data (Landsat5 and Landsat8), meteorological data (ERA5-Land), and elevation data (DSM) are collected. Ground quadrat data includes latitude and longitude, grassland type, and dominant species information; remote sensing data undergoes quality control and median synthesis to calculate features such as the Normalized Difference Vegetation Index (NDVI) and Soil-Adjusted Vegetation Index (SAVI); meteorological data includes annual mean temperature and annual total precipitation; and elevation data provides topographic information. All data are rasterized to generate masked data to identify valid areas (areas with ground quadrat points are marked as 1, and areas without quadrat points are marked as 0), eliminating prior noise in invalid areas.

[0111] Visible light data, annual mean temperature, annual precipitation, and DEM data for the years in which the field survey data were collected were extracted from the locations. The indices in Table 1 were calculated to obtain the dataset for training and validating the model. Landsat 5 reflectance data from sampling points before 2011 were corrected using a correction coefficient. A Gaussian low-pass model was used to filter data with lower noise in each category. The D0 value was set to 30 for both the first and second levels, and 50 and 10 for grassland and meadow in the third level, respectively. After the second-level filtering, some third-level categories were missing. Samples of the missing categories were then added to the third-level classification dataset by searching upwards through the next level.

[0112] The formula for the Gaussian low-pass model is as follows:

[0113] ;

[0114] After noise removal, each level type is divided into training and testing datasets in an 8:2 ratio. When the number of samples of a certain type in the same level is much smaller than that of other types, all samples of that type are used as both training and validation data.

[0115] Table 1. Description of Index Characteristics

[0116]

[0117] In this embodiment, a three-layer classification system is designed during the hierarchical random forest classification process. Before the classification system, the existing land use type dataset is simultaneously divided into built-up areas, cultivated land, water areas, and forest land (non-shrubland) to perform coarse masking on the non-grassland areas of the two periods of Qinghai-Tibet Plateau data. The first layer is to distinguish grassland areas from non-grassland areas. It should be noted that cultivated land and terraced fields are easily confused with grassland types in the Qinghai-Tibet Plateau region. Therefore, when distinguishing grassland from non-grassland, cultivated land and terraced fields that have been consistently present in multiple experiments, and forest, bare land, gravel, lake, glacier, and desert that appear at any one time, are used as masks for non-grassland areas.

[0118] The second layer performs a first-level fine classification of grassland areas, namely "meadows, grasslands, deserts, swamps, and secondary shrublands resulting from forest destruction." Since cultivated land and terraces within the non-grassland masks created in the first layer have consistently existed across multiple experiments, the possibility of localized cultivated land and terraces within grassland areas needs to be considered when performing the second-level random forest classification. Therefore, in addition to the five grassland vegetation types, training data for this layer needs to be supplemented with training data for cultivated land and terraces. The third layer, based on Table 2, supplements the easily confused second-level types included in the third-level classification when grouping and classifying the third-level types.

[0119] Table 2. Correlation between Secondary and Tertiary Grassland Resource Types and Easily Confused Information

[0120]

[0121] A progressive three-level classification is adopted. The first level extracts information on grassland, cultivated land, forest, bare land, gravel, lake, glacier, and desert types. The second level coarsely classifies grassland into steppe, desert, shrubland, meadow, and marsh. The third level finely classifies grassland into: temperate meadow steppe, temperate steppe, temperate desert steppe, alpine meadow steppe, alpine steppe, alpine desert steppe, temperate steppe-desert, temperate desert, alpine desert, tropical grassland, tropical shrubland, warm grassland, warm shrubland, lowland meadow, temperate mountain meadow, alpine meadow, and marsh.

[0122] For example, the classification results are as follows Figure 3 The comparison charts shown here are of the classification results for the Gonghe Basin. Figure a is a map of grassland resource types in the 1980s, Figure b is the classification result in 1980, and Figure c is the classification result in 2020.

[0123] Optionally, the method for constructing the pre-built multi-level classification model includes:

[0124] Obtain historical multi-source fusion feature data of the target area, and construct training and testing sets based on the historical multi-source fusion feature data;

[0125] For any level of classification model in a multi-level classification model, the training set and the test set are cropped based on the mask output by the classification model of the previous level of the current level, respectively, to obtain the cropped training set and the test set. The first level classification model crops the training set and the test set based on the initial mask, which is a set mask.

[0126] For any layer of the multi-level classification model, feature filtering is performed on the pruned training and test sets through recursive feature elimination and random forest evaluator to obtain the feature-filtered training and test sets.

[0127] For any layer of the multi-level classification model, the pre-constructed random forest model is trained based on the training set after grid search, cross-validation optimization and feature selection, and the RF hyperparameters are optimized to obtain the trained random forest model.

[0128] For any layer of the multi-level classification model, multi-scale features of the cropped training set are generated based on weight pooling, and the CNN model is trained through encoder-decoder structure, gated convolution and attention mechanism to obtain the trained CNN model.

[0129] For any layer of the multi-level classification model, the trained random forest model and the trained CNN model are evaluated based on the pruned test set, and the optimal algorithm or algorithm fusion scheme is selected as the final classification model for the current level based on the predetermined evaluation index.

[0130] For any layer of the multi-level classification model, the grassland resources of the target area are predicted based on the final classification model of the current layer, the classification result of the current layer is generated, and it is transformed into the input mask required for the next layer, until the training of all layers of the model is completed, and the trained multi-level classification model is obtained.

[0131] In this embodiment of the application, the step of training a pre-built random forest model based on the training set after grid search, cross-validation optimization, and feature selection, and optimizing the RF hyperparameters to obtain the trained random forest model includes:

[0132] Based on the grid search strategy, all parameter combinations in the hyperparameter search space of the random forest model are traversed. The parameter combinations include the number of decision trees, the maximum depth of the tree, the minimum number of samples for splitting internal nodes, the minimum number of samples for leaf nodes, the maximum number of features considered when splitting, and whether to use bootstrap sampling.

[0133] For each hyperparameter combination, the model is trained and validated on the training set after feature selection based on the k-fold cross-validation algorithm to obtain the macro-average F1 score evaluation result of each hyperparameter combination under cross-validation.

[0134] The combination of hyperparameters that achieves the highest macro-average F1 score in cross-validation is selected as the optimal hyperparameters.

[0135] Based on the optimal hyperparameters and the entire training set after feature selection, the random forest model is retrained to obtain the trained random forest model.

[0136] In the above implementation, grid search combined with k-fold cross-validation (k=5 or 10) is used to dynamically optimize the RF hyperparameters. The main parameters include:

[0137] n_estimators: Number of decision trees (range: 50-500, step size 50).

[0138] max_depth: The maximum depth of the tree (range: 5-30, or None means unlimited).

[0139] min_samples_split: The minimum number of samples for an internal node to split (range: 2-10).

[0140] min_samples_leaf: The minimum number of samples in a leaf node (range: 1-5).

[0141] max_features: The maximum number of features to consider during splitting (options: 'auto', 'sqrt', 'log2').

[0142] bootstrap: Whether to use self-sampling (True / False).

[0143] Optimization process: The system iterates through the parameter combinations, uses cross-validation scoring (CV Score, usually macro average F1 score) to evaluate the performance of each combination, and selects the parameter with the highest score as the optimal configuration.

[0144] The Randomization (RF) model is constructed using optimized parameters. Each decision tree is randomly sampled from the training set (with replacement, bootstrap sampling), and a subset of features is randomly selected when a node splits. The aggregation method is majority voting.

[0145] The first-level RF model classifies the overall land cover type; the second level trains / predicts only on the first-level "grassland" pixels; the third level further refines the second-level results. This progressive approach reduces classification confusion and improves fine-grained accuracy.

[0146] Grid search + CV avoids the blindness of manual parameter tuning and ensures generalization ability; random subspace enhances diversity and reduces variance; for large remote sensing samples, parallel computing (n_jobs=-1) is used to accelerate training.

[0147] In this embodiment of the application, the step of generating multi-scale features of the training set after feature selection based on weight pooling, and training the CNN model through an encoder-decoder structure, gated convolution, and attention mechanism to obtain the trained CNN model includes:

[0148] The cropped training set is downsampled based on the weighted pooling operator to generate feature maps at multiple scales. The weighted pooling operator is calculated by performing weighted average pooling on the effective regions of the image and setting the weight of the invalid regions to 0.

[0149] The encoder performs convolution operations on the feature maps at each scale to extract deep features from the multi-scale feature maps, resulting in an encoded multi-scale deep feature map. The encoder is composed of multiple convolutional layers, activation function layers, and pooling layers stacked together.

[0150] Calculate the gate value of the encoded multi-scale deep feature map, and dynamically select and fill the feature map based on the gate value to obtain the filled feature map;

[0151] Based on historical multi-source fusion feature data, the attention weights of geographic environmental features are calculated to obtain an attention weight map. The geographic environmental features include the geographic location, elevation, topographic relief, and classification trend of the pixels.

[0152] Based on the padded feature maps and attention weight maps, the decoder uses an upsampling operation to restore the multi-scale feature maps extracted by the encoder to their original resolution, and then fuses the outputs of the gated convolution and attention mechanism, outputting the trained CNN model through the Softmax activation function.

[0153] In the above implementation, preprocessing is performed on the original remote sensing images (such as multispectral satellite data), including atmospheric correction, geometric registration, and normalization. Initial feature maps are generated by combining the feature data (such as spectrum, texture, and terrain) in Table 1. For different resolutions (such as 500 meters and 1000 meters), the original images are downsampled using a weighted pooling operator to generate multi-scale feature maps.

[0154] Weighted average pooling is applied to the valid regions of the image, while invalid regions (such as those obscured by clouds or missing pixels) have a weight of 0. The formula is as follows:

[0155] ;

[0156] in, w i The weight is the reciprocal of the number of valid points within the valid region and 0 within the invalid region. x i These are pixel values.

[0157] Weighted pooling is applied layer by layer to generate feature maps at multiple scales (e.g., original resolution, 500 meters, 1000 meters, etc.) to capture features from local details to global trends.

[0158] Convolutional layers are stacked (e.g., ResNet or U-Net encoding paths) to perform convolutional operations on multi-scale feature maps. Features are extracted independently at each scale, including convolutional kernels (3x3 or 5x5), activation functions (ReLU), and pooling layers. The feature extraction process captures both trends (e.g., vegetation cover) and details (e.g., texture variations) of the grassland resources.

[0159] The standard convolution formula is:

[0160] ;

[0161] in, y j To output feature map channels j , x i For input channel i , w ij These are the kernel weights, where * indicates convolution. b j For bias, f This is the activation function.

[0162] Feature maps of different scales are processed through parallel branches, which reduces the difficulty of model training and improves stability.

[0163] Residual connections are used to avoid gradient vanishing and improve the training efficiency of deep networks; for large remote sensing images, depthwise separable convolution is used to reduce the number of parameters (reducing computational overhead by about 80%); feature extraction emphasizes multispectral channel fusion to capture the spatiotemporal changes of indices such as NDVI.

[0164] For invalid areas in remote sensing images (such as cloud cover), a gating unit is introduced to dynamically select features for trend filling.

[0165] Attention mechanism weight calculation: Integrate geographical environmental features (such as location, elevation, topographic relief, classification trend) to calculate attention weights and enhance spatial autocorrelation.

[0166] During the fusion process, attention weights are applied to multi-scale features and geographic features to achieve dynamic fusion.

[0167] Upsampling layers (such as transposed convolution or bilinear interpolation) are used to resample multi-scale feature maps to a uniform resolution. Weighted summation is then used for fusion.

[0168] ;

[0169] in, y For categorized output, x i For the input features, w i b is the weight, and b is the bias. f This is the LeakyReLU activation function.

[0170] The final layer uses Softmax activation to output a grassy fine-classification probability map.

[0171] In this embodiment of the application, the step of evaluating the trained random forest model and the trained CNN model based on the feature-selected test set, and selecting the optimal algorithm or algorithm fusion scheme as the final classification model for the current level based on a predetermined evaluation metric, includes:

[0172] Based on the test set, the macro-average evaluation index of the trained random forest model and the trained CNN model are calculated respectively. The macro-average evaluation index includes overall accuracy, user accuracy, producer accuracy, F1 index and Kappa coefficient.

[0173] If the F1 index of the random forest model is higher than that of the CNN model, and the difference in F1 index exceeds a first preset threshold, then the random forest model is selected as the final classification model for the current level.

[0174] If the F1 index of the CNN model is higher than that of the random forest model, and the difference in F1 index exceeds the first preset threshold, then the CNN model is selected as the final classification model for the current level.

[0175] If the difference in F1 index between the random forest model and the CNN model is less than or equal to the first preset threshold, a weighted fusion scheme is used to fuse the CNN model and the random forest model into the final classification model of the current level.

[0176] The probability calculation formula for weighted fusion in the weighted fusion scheme is as follows:

[0177] ;

[0178] Where P RF P represents the predicted probability output by the random forest model. CNN The weight α is the predicted probability output by the convolutional neural network model, and it is determined by cross-validation optimization on the validation set.

[0179] In the above implementation, a reserved test dataset (accounting for 15% of the total data) is used to independently evaluate the RF and CNN models to avoid training bias. The test set covers various terrains, seasons, and spectral variations to ensure generalization ability.

[0180] The model performance is quantified using multi-dimensional metrics, including:

[0181] Overall Accuracy (OA): The proportion of pixels that are correctly classified, expressed by the formula:

[0182] ;

[0183] TP, TN, FP, and FN represent true positive, true negative, false positive, and false negative, respectively.

[0184] User precision (UA): The proportion of samples that are truly positive among those classified as positive, expressed by the formula:

[0185] ;

[0186] Producer precision (PA, recall): the proportion of true positive samples that are correctly classified, expressed as follows:

[0187] ;

[0188] F1 score: the harmonic average of precision and recall, calculated using the following formula:

[0189] ;

[0190] Kappa coefficient: A classification consistency index that considers random consistency; the formula is:

[0191] ;

[0192] in, P e The expected consistency probability (based on the marginal distribution of the confusion matrix).

[0193] These indicators employ a macro-averaged approach to handle multiple categories, ensuring that subcategories (such as swamps) are not overlooked. Macro-averaging calculation: Indicators are calculated separately for the 18 subcategories and 7 major categories. The macro-averaging formula is as follows:

[0194] ;

[0195] Where N is the number of categories, Metric i Let be the index value for the i-th class.

[0196] Optionally, it may also include: obtaining third-party verification data, which includes UAV imagery data and historical grassland resource map data;

[0197] The third-party verification data is preprocessed to obtain preprocessed third-party verification data. The preprocessed third-party verification data is consistent with the grassland resource type classification map output by the multi-level classification model in terms of spatial coordinate system and resolution.

[0198] Based on the UAV imagery data, the secondary grassland types of the grassland resource type classification map are locally verified to obtain the first verification result;

[0199] Based on the historical grassland resource map data, a macro-consistency verification of the merged types of the grassland resource type classification map is performed to obtain a second verification result;

[0200] Based on the first verification result and the second verification result, a multi-dimensional evaluation index of macro average of secondary grassland types and merged types is calculated. The evaluation index includes total accuracy, user accuracy, producer accuracy, F1 index and Kappa coefficient.

[0201] Based on the preset regional division rules, the spatial autocorrelation index of the grassland resource type classification map and the classification accuracy index within each zone are calculated to obtain the spatial assessment results and regional classification deviation.

[0202] The consistency between the grassland resource type classification map and the third-party verification data was analyzed based on statistical testing methods, and the test results were obtained.

[0203] Based on the first verification result, the second verification result, multi-dimensional evaluation indicators, spatial assessment results, regional classification bias, and test results, easily confused grassland type pairs are identified, the causes of classification errors are analyzed, and the analysis results are fed back into the training process of the multi-level classification model to optimize model parameters or feature selection.

[0204] In this embodiment, the UAV imagery data provides high-resolution (centimeter-level) images covering typical grassland sample areas (such as alpine meadows and temperate grasslands). The UAV data includes multispectral imagery and ground-measured annotations, suitable for detailed local verification.

[0205] Perform geometric correction, radiometric calibration, and mosaicking on the UAV imagery data to ensure alignment with the model's output resolution (500 meters or 1000 meters). Use ground control points (GCPs) to improve registration accuracy (error < 1 pixel).

[0206] For historical grassland resource map data, the vector data is converted to raster format, and a unified projection (such as WGS84) and resolution are applied. To address temporal differences, potential biases (such as vegetation degradation) are corrected by combining contemporaneous NDVI and climate data.

[0207] The macro-averaged method is used to calculate the accuracy metrics and balance the importance of each category. The specific accuracy metrics include: overall accuracy, user accuracy, producer accuracy, and F1 index.

[0208] To assess the spatial continuity of classification results based on the spatial characteristics of remote sensing classification, Moran's I index was used. The Tibetan Plateau was divided into ecological zones (e.g., alpine and temperate zones), and indices were calculated for each zone to identify regional biases.

[0209] The model classification results were compared pixel-by-pixel with UAV data and historical grassland resource maps to calculate the consistency rate and areas of deviation. UAV data was used to validate fine-grained classifications (such as temperate meadow-steppe), while historical maps were used to validate the distribution of the seven major categories.

[0210] Based on the confusion matrix, easily confused categories (such as temperate grassland and temperate desert grassland) are identified, and the reasons (such as spectral similarity and terrain interference) are analyzed. For low-precision categories (such as swamps), feature selection or model parameters are examined.

[0211] The evaluation results are fed back to the algorithm optimization, adjusting RF or CNN parameters (such as adding decision trees or optimizing the learning rate) or fusing weights to improve model performance.

[0212] This method employs a three-layer classification framework, combining ground quadrat data, Landsat 5 / 8 remote sensing data, ERA5-Land meteorological data, and DSM elevation data to progressively distinguish between grassland and non-grassland, coarsely classify grassland types (steppe, desert, shrubland, meadow, and marsh), and refine them into 18 grassland resource categories. Gaussian low-pass filtering is used for noise reduction, weighted pooling generates multi-scale feature maps, gated convolution fills in missing regions, and an attention mechanism enhances the spatial autocorrelation of geographic environmental features. Random forest and convolutional neural networks are compared and optimized, and the best model is fused. Accuracy is evaluated based on UAV data and the 1980 Qinghai-Tibet Plateau grassland resource map, calculating the overall macro-average accuracy, user accuracy, producer accuracy, F1 index, and Kappa coefficient. This invention achieves high-precision grassland resource classification.

[0213] Figure 4 A schematic diagram of a grassland remote sensing resource automatic classification device 200 is shown.

[0214] like Figure 4 As shown, an automatic classification device 200 for grassland remote sensing resources mainly includes:

[0215] The acquisition module 201 is used to acquire multi-source raw data of the target area. The multi-source raw data includes remote sensing data, meteorological data, ground quadrat data, and elevation data. The ground quadrat data includes latitude and longitude, grassland type, and dominant species information. The meteorological data includes annual average temperature and annual average precipitation.

[0216] The feature extraction module 202 is used to extract features from the multi-source raw data to obtain multi-source fused feature data, which includes spectral index, soil index, water body index, topographic index, and meteorological index.

[0217] The classification module 203 is used to input the multi-source fusion feature data into a pre-constructed multi-level classification model to obtain a grassland resource type classification map of the target area;

[0218] The pre-built multi-level classification model is trained using a progressive three-level framework based on historical multi-source fusion feature data and corresponding grassland type labels. The pre-built multi-level classification model includes a first-level classification model, a second-level classification model, and a third-level classification model. The first-level classification model is used to distinguish between grassland and non-grassland areas and generate a grassland mask. The second-level classification model is used to perform primary grassland type classification within grassland areas. The third-level classification model is used to further subdivide secondary grassland types based on the primary grassland types.

[0219] In one example, the module in any of the above devices may be one or more integrated circuits configured to implement the above methods, such as one or more application-specific integrated circuits (ASICs), or one or more digital signal processors (DSPs), or one or more field-programmable gate arrays (FPGAs), or a combination of at least two of these integrated circuit forms.

[0220] For example, when modules in a device can be implemented via a processing element scheduler, the processing element can be a general-purpose processor, such as a central processing unit (CPU) or other processor capable of calling programs. Alternatively, these modules can be integrated together as a system-on-a-chip (SOC).

[0221] In this application, various objects such as messages / information / devices / network elements / systems / apparatus / actions / operations / processes / concepts may be named. It is understood that these specific names do not constitute a limitation on the relevant objects. The names may be changed depending on the scenario, context, or usage habits. The understanding of the technical meaning of the technical terms in this application should be mainly determined from their functions and technical effects embodied / performed in the technical solution.

[0222] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and modules described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0223] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0224] Figure 5 This is a structural block diagram of an electronic device 300 according to an embodiment of this application.

[0225] like Figure 5 As shown, the electronic device 300 includes a processor 301 and a memory 302, and may further include one or more of an information input / output (I / O) interface 303, a communication component 304, and a communication bus 305.

[0226] The processor 301 controls the overall operation of the electronic device 300 to complete all or part of the steps in the above-described automatic classification method for grassland remote sensing resources. The memory 302 stores various types of data to support the operation of the electronic device 300. This data may include, for example, instructions for any application or method operating on the electronic device 300, as well as application-related data. The memory 302 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as one or more of Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0227] I / O interface 303 provides an interface between processor 301 and other interface modules, such as keyboards, mice, and buttons. These buttons can be virtual or physical. Communication component 304 is used to test wired or wireless communication between electronic device 300 and other devices. Wireless communication includes Wi-Fi, Bluetooth, Near Field Communication (NFC), 2G, 3G, or 4G, or a combination thereof. Therefore, the corresponding communication component 304 may include a Wi-Fi component, a Bluetooth component, and an NFC component.

[0228] The communication bus 305 may include a path for transmitting information between the aforementioned components. The communication bus 305 may be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc. The communication bus 305 may be divided into an address bus, a data bus, a control bus, etc.

[0229] The electronic device 300 may be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to execute the automatic classification method for grassland remote sensing resources given in the above embodiments.

[0230] The terms “comprising,” “including,” or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0231] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the foregoing application concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions claimed in this application.

Claims

1. A method for automatic classification of grassland remote sensing resources, characterized in that, include: Acquire multi-source raw data of the target area. The multi-source raw data includes remote sensing data, meteorological data, ground quadrat data and elevation data. The ground quadrat data includes latitude and longitude, grassland type and dominant species information. The meteorological data includes annual average temperature and annual average precipitation. Feature extraction is performed on the multi-source raw data to obtain target multi-source fusion feature data. The target multi-source fusion feature data includes various remote sensing indices, topographic indices, and meteorological indices. The various remote sensing indices include normalized vegetation index, soil-regulated vegetation index, ratio vegetation index, humidity index, bare soil index, and normalized water index. The topographic indices include elevation, slope, and aspect. The meteorological indices include annual average temperature, annual total precipitation, accumulated temperature index, and growing season temperature-related days index. The target multi-source fusion feature data is input into a pre-constructed multi-level classification model to obtain a grassland resource type classification map of the target area; The pre-built multi-level classification model is trained using a progressive three-level framework based on historical multi-source fusion feature data and corresponding grassland type labels. The pre-built multi-level classification model includes a first-layer classification model, a second-layer classification model, and a third-layer classification model. The first-layer classification model is used to distinguish between grassland and non-grassland areas and generate a grassland mask. The second-layer classification model is used to perform primary grassland type classification within grassland areas. The third-layer classification model is used to further subdivide secondary grassland types based on the primary grassland types. The method for constructing the pre-built multi-level classification model includes: Obtain historical multi-source fusion feature data of the target area, and construct training and testing sets based on the historical multi-source fusion feature data; For any level of classification model in a multi-level classification model, the training set and test set are cropped based on the mask output by the classification model of the previous level of the current level, respectively, to obtain the cropped training set and test set. The first level classification model crops the training set and test set based on the initial mask, which is a set mask. For any layer of the multi-level classification model, feature filtering is performed on the pruned training and test sets through recursive feature elimination and random forest evaluator to obtain the feature-filtered training and test sets. For any layer of the multi-level classification model, the pre-constructed random forest model is trained based on the training set after grid search, cross-validation optimization and feature selection, and the RF hyperparameters are optimized to obtain the trained random forest model. For any layer of the multi-level classification model, multi-scale features of the cropped training set are generated based on weight pooling, and the CNN model is trained through encoder-decoder structure, gated convolution and attention mechanism to obtain the trained CNN model. For any layer of the multi-level classification model, the trained random forest model and the trained CNN model are evaluated based on the pruned test set, and the optimal algorithm or algorithm fusion scheme is selected as the final classification model for the current level based on the predetermined evaluation index. For any layer of the multi-level classification model, the grassland resources of the target area are predicted based on the final classification model of the current layer, the classification result of the current layer is generated, and it is transformed into the input mask required for the next layer, until the training of all layers of the model is completed, and the trained multi-level classification model is obtained.

2. The method for automatic classification of grassland remote sensing resources according to claim 1, characterized in that, The step of extracting features from the multi-source raw data to obtain target multi-source fused feature data includes: Spatial registration and rasterization are performed on the acquired remote sensing data, meteorological data and elevation data to obtain the preprocessed first remote sensing data, first meteorological data and first elevation data; The first remote sensing data is subjected to radiometric calibration, atmospheric correction and quality control to obtain the preprocessed second remote sensing data. Based on ground quadrat data, areas with sampling points in the target area are marked as valid areas, and areas without sampling points are marked as invalid areas, thus generating a mask for the valid areas. Based on the preprocessed second remote sensing data, various remote sensing indices for the target area are calculated. Based on the preprocessed first elevation data, the terrain index of the target area is extracted; Based on the preprocessed first meteorological data, the meteorological index of the target area is extracted; The various remote sensing indices, topographic indices, and meteorological indices are superimposed on the effective area mask to generate initial multi-source fusion feature data; The initial multi-source fusion feature data is denoised based on a preset Gaussian low-pass filter model to obtain preprocessed multi-source fusion feature data. The cutoff frequency parameter of the Gaussian low-pass filter model is set based on the classification level and grassland type. Feature filtering is performed on the preprocessed multi-source fusion feature data based on recursive feature elimination, and optimization is performed based on weighted pooling operators to obtain the target multi-source fusion feature data.

3. The method for automatic classification of grassland remote sensing resources according to claim 1, characterized in that, The step of inputting the multi-source fusion feature data into a pre-constructed multi-level classification model to obtain a grassland resource type classification map of the target area includes: The multi-source fusion feature data is input into the first-layer classification model to obtain the first-layer classification result of the target region. The first-layer classification result includes grassland masking and non-grassland masking. The grassland mask and the multi-source fusion feature data are input into the second-layer classification model to obtain the second-layer classification result of the target area. The second-layer classification result includes the primary grassland type, which includes grassland, desert, shrubland, meadow, swamp, cultivated land and terraces. The second-layer classification result and the multi-source fusion feature data are input into the third-layer classification model to obtain the third-layer classification result. The third-layer classification result includes secondary grassland types, which include temperate meadow steppe, temperate steppe, temperate desert steppe, alpine meadow steppe, alpine steppe, alpine desert steppe, temperate steppe desert, temperate desert, alpine desert, tropical grassland, tropical shrub grassland, warm grassland, warm shrub grassland, lowland meadow, temperate mountain meadow, alpine meadow, and marsh. The classification results of the secondary grassland types are merged into seven major categories: tropical grassland, warm grassland, meadow steppe, alpine meadow, typical steppe, alpine steppe and desert, generating a grassland resource type classification map of the target area.

4. The method for automatic classification of grassland remote sensing resources according to claim 1, characterized in that, The process of evaluating the trained random forest model and the trained CNN model based on the pruned test set, and selecting the optimal algorithm or algorithm fusion scheme as the final classification model for the current level based on predetermined evaluation metrics, includes: Based on the test set, the macro-average evaluation index of the trained random forest model and the trained CNN model are calculated respectively. The macro-average evaluation index includes overall accuracy, user accuracy, producer accuracy, F1 index and Kappa coefficient. If the F1 index of the random forest model is higher than that of the CNN model, and the difference in F1 index exceeds a first preset threshold, then the random forest model is selected as the final classification model for the current level. If the F1 index of the CNN model is higher than that of the random forest model, and the difference in F1 index exceeds the first preset threshold, then the CNN model is selected as the final classification model for the current level. If the difference in F1 index between the random forest model and the CNN model is less than or equal to the first preset threshold, a weighted fusion scheme is used to fuse the CNN model and the random forest model into the final classification model of the current level. The probability calculation formula for weighted fusion in the weighted fusion scheme is as follows: P=α×P RF +(1-α)×P CNN , where P RF P represents the predicted probability output by the random forest model. CNN The weight α is the predicted probability output by the convolutional neural network model, and it is determined by cross-validation optimization on the validation set.

5. The method for automatic classification of grassland remote sensing resources according to claim 1, characterized in that, The pre-built random forest model is trained using the training set obtained from grid search, cross-validation optimization, and feature selection. The RF hyperparameters are optimized to obtain the trained random forest model, including: Based on the grid search strategy, all parameter combinations in the hyperparameter search space of the random forest model are traversed. The parameter combinations include the number of decision trees, the maximum depth of the tree, the minimum number of samples for splitting internal nodes, the minimum number of samples for leaf nodes, the maximum number of features considered when splitting, and whether to use bootstrap sampling. For each hyperparameter combination, the model is trained and validated on the training set after feature selection based on the k-fold cross-validation algorithm to obtain the macro-average F1 score evaluation result of each hyperparameter combination under cross-validation. The combination of hyperparameters that achieves the highest macro-average F1 score in cross-validation is selected as the optimal hyperparameters. Based on the optimal hyperparameters and the entire training set after feature selection, the random forest model is retrained to obtain the trained random forest model.

6. The method for automatic classification of grassland remote sensing resources according to claim 1, characterized in that, The process involves generating multi-scale features of the training set after feature selection based on weight pooling, and training the CNN model using an encoder-decoder structure, gated convolution, and attention mechanism to obtain the trained CNN model, including: The cropped training set is downsampled based on the weighted pooling operator to generate feature maps at multiple scales. The weighted pooling operator is calculated by performing weighted average pooling on the effective regions of the image and setting the weight of the invalid regions to 0. The encoder performs convolution operations on the feature maps at each scale to extract deep features from the multi-scale feature maps, resulting in an encoded multi-scale deep feature map. The encoder is composed of multiple convolutional layers, activation function layers, and pooling layers stacked together. Calculate the gate value of the encoded multi-scale deep feature map, and dynamically select and fill the feature map based on the gate value to obtain the filled feature map; Based on historical multi-source fusion feature data, the attention weights of geographic environmental features are calculated to obtain an attention weight map. The geographic environmental features include the geographic location, elevation, topographic relief, and classification trend of the pixels. Based on the padded feature maps and attention weight maps, the decoder uses an upsampling operation to restore the multi-scale feature maps extracted by the encoder to their original resolution, and then fuses the outputs of the gated convolution and attention mechanism, outputting the trained CNN model through the Softmax activation function.

7. The method for automatic classification of grassland remote sensing resources according to claim 1, characterized in that, Also includes: Obtain third-party verification data, including UAV imagery data and historical grassland resource map data; The third-party verification data is preprocessed to obtain preprocessed third-party verification data. The preprocessed third-party verification data is consistent with the grassland resource type classification map output by the multi-level classification model in terms of spatial coordinate system and resolution. Based on the UAV imagery data, the secondary grassland types of the grassland resource type classification map are locally verified to obtain the first verification result; Based on the historical grassland resource map data, a macro-consistency verification of the merged types of the grassland resource type classification map is performed to obtain a second verification result; Based on the first verification result and the second verification result, a multi-dimensional evaluation index of macro average of secondary grassland types and merged types is calculated. The evaluation index includes total accuracy, user accuracy, producer accuracy, F1 index and Kappa coefficient. Based on the preset regional division rules, the spatial autocorrelation index of the grassland resource type classification map and the classification accuracy index within each zone are calculated to obtain the spatial assessment results and regional classification deviation. The consistency between the grassland resource type classification map and the third-party verification data was analyzed based on statistical testing methods, and the test results were obtained. Based on the first verification result, the second verification result, multi-dimensional evaluation indicators, spatial assessment results, regional classification bias, and test results, easily confused grassland type pairs are identified, the causes of classification errors are analyzed, and the analysis results are fed back into the training process of the multi-level classification model to optimize model parameters or feature selection.

8. An automatic classification device for grassland remote sensing resources, characterized in that, include: The acquisition module is used to acquire multi-source raw data of the target area. The multi-source raw data includes remote sensing data, meteorological data, ground quadrat data and elevation data. The ground quadrat data includes latitude and longitude, grassland type and dominant species information. The meteorological data includes annual average temperature and annual average precipitation. The feature extraction module is used to extract features from the multi-source raw data to obtain target multi-source fusion feature data. The target multi-source fusion feature data includes various remote sensing indices, topographic indices, and meteorological indices. The various remote sensing indices include normalized vegetation index, soil-regulated vegetation index, ratio vegetation index, humidity index, bare soil index, and normalized water index. The topographic indices include elevation, slope, and aspect. The meteorological indices include average annual temperature, total annual precipitation, accumulated temperature index, and growing season temperature-related days index. The classification module is used to input the multi-source fusion feature data into a pre-constructed multi-level classification model to obtain a grassland resource type classification map of the target area; The pre-built multi-level classification model is trained using a progressive three-level framework based on historical multi-source fusion feature data and corresponding grassland type labels. The pre-built multi-level classification model includes a first-layer classification model, a second-layer classification model, and a third-layer classification model. The first-layer classification model is used to distinguish between grassland and non-grassland areas and generate a grassland mask. The second-layer classification model is used to perform primary grassland type classification within grassland areas. The third-layer classification model is used to further subdivide secondary grassland types based on the primary grassland types. The classification module is specifically used for: Obtain historical multi-source fusion feature data of the target area, and construct training and testing sets based on the historical multi-source fusion feature data; For any level of classification model in a multi-level classification model, the training set and test set are cropped based on the mask output by the classification model of the previous level of the current level, respectively, to obtain the cropped training set and test set. The first level classification model crops the training set and test set based on the initial mask, which is a set mask. For any layer of the multi-level classification model, feature filtering is performed on the pruned training and test sets through recursive feature elimination and random forest evaluator to obtain the feature-filtered training and test sets. For any layer of the multi-level classification model, the pre-constructed random forest model is trained based on the training set after grid search, cross-validation optimization and feature selection, and the RF hyperparameters are optimized to obtain the trained random forest model. For any layer of the multi-level classification model, multi-scale features of the cropped training set are generated based on weight pooling, and the CNN model is trained through encoder-decoder structure, gated convolution and attention mechanism to obtain the trained CNN model. For any layer of the multi-level classification model, the trained random forest model and the trained CNN model are evaluated based on the pruned test set, and the optimal algorithm or algorithm fusion scheme is selected as the final classification model for the current level based on the predetermined evaluation index. For any layer of the multi-level classification model, the grassland resources of the target area are predicted based on the final classification model of the current layer, the classification result of the current layer is generated, and it is transformed into the input mask required for the next layer, until the training of all layers of the model is completed, and the trained multi-level classification model is obtained.

9. An electronic device, characterized in that, Includes a processor, which is coupled to a memory; The processor is configured to execute a computer program stored in the memory to cause the electronic device to perform the method as described in any one of claims 1 to 7.

Images