Reinforced AI computing power and timing traceability vegetation species identification method

Abstract

The application discloses a vegetation species identification method for strengthening AI computing power and time sequence tracing, comprising the following steps: collecting multi-source remote sensing data and performing pretreatment, constructing a vegetation classification dataset with spatiotemporal alignment and unified resolution; generating a vegetation mask based on the vegetation classification dataset, obtaining a standardized sample slice, and constructing a training dataset in combination with spectral characteristics; performing time-phase processing on the training dataset based on vegetation phenological characteristics, and outputting a preliminary vegetation classification result; performing object-level optimization on the preliminary vegetation classification result, and obtaining an optimized vegetation classification result; correcting the optimized vegetation classification result in combination with terrain data, generating a vegetation species classification map of a target year, and realizing vegetation dynamic change inversion in a specified time period through transfer learning. Therefore, the traditional resolution limit can be broken through, the classification error problem caused by independent use of multi-source data can be solved, the discrimination of complex vegetation types can be significantly improved, and historical vegetation dynamic backtracking analysis can be supported.

Description

Technical Field

[0001] This invention relates to the field of machine learning, specifically to a method for identifying vegetation species that enhances AI computing power and time-series tracing. Background Technology

[0002] Species-level vegetation classification of forests, shrubs and grasses is a core foundational task in fields such as ecological restoration and carbon sequestration assessment, but its accuracy and efficiency have long been limited by the inadequacy of existing technologies.

[0003] Currently, publicly available land use data (such as ESA World Cover and ESR I Land Cover) have a resolution of only 10-30 meters, which cannot distinguish fragmented vegetation patches with an area of ​​less than 0.1 hectares. Especially in the shrub-grass transition zone, it is difficult to meet the needs of species-level fine classification, which seriously restricts the accuracy of ecological management.

[0004] Meanwhile, multi-source data, including space-based (satellite), air-based (drone), and ground-based (field survey) data, have long been processed independently without an effective collaborative mechanism. This has resulted in the data value not being fully realized, and vegetation classification results are easily affected by factors such as cloud cover and temporal differences, leading to significant errors.

[0005] In terms of algorithmic models, traditional machine learning methods (such as SVM and single random forest) have low distinguishability for vegetation with complex spectral mixtures, with a shrub-grass confusion rate exceeding 40%. Single deep learning models (such as U-Net) do not optimize the association of multi-temporal features and are prone to misjudgment due to seasonal changes. Mainstream technical solutions either rely on single models such as SegFormer and random forest to process multispectral data, failing to integrate historical temporal data to capture vegetation succession patterns; or mechanically string together multi-source data for only pixel-level processing, with manual correction accounting for over 70%; or use independently running deep learning and machine learning models, failing to leverage their collaborative optimization potential. All of these approaches are insufficient to meet the needs of species-level vegetation identification and dynamic tracking. Summary of the Invention

[0006] The present invention aims to at least partially solve the technical problems in the above-mentioned technologies.

[0007] Therefore, this invention discloses a method for identifying vegetation species that enhances AI computing power and time-series tracing, comprising the following steps:

[0008] S1: Collect multi-source remote sensing data and preprocess it to construct a vegetation classification dataset with spatiotemporal alignment and uniform resolution. The multi-source remote sensing data includes space-based remote sensing data, airborne remote sensing data and ground-based spectral data.

[0009] S2: Generate a vegetation mask based on the vegetation classification dataset, and use the vegetation mask to cut the multi-source remote sensing data to obtain standardized sample slices, and combine them with spectral features to construct a training dataset;

[0010] S3: Based on the phenological characteristics of vegetation, the training dataset is processed by time phase, and an enhanced AI model is used to extract and classify the processed training dataset to output preliminary vegetation classification results.

[0011] S4: Perform object-level optimization on the preliminary vegetation classification results, eliminate abnormal pixels, and obtain optimized vegetation classification results;

[0012] S5: Correct the optimized vegetation classification results by combining terrain data, generate a vegetation species classification map for the target year, and realize the inversion of vegetation dynamic changes within a specified time period through transfer learning.

[0013] The vegetation species identification method with enhanced AI computing power and time-series tracing disclosed in this invention has at least the following beneficial effects:

[0014] (1) By integrating multi-source data and unifying resolution, the traditional resolution limitation has been broken through, and species-accurate vegetation classification at the 2-meter level has been achieved. This can effectively distinguish fragmented vegetation patches such as shrub-grass transition zones and meet the needs of fine classification of forest, shrub and grass species.

[0015] (2) A multi-dimensional monitoring data collaboration mechanism of "space-time spectrum" nested in the sky and ground was constructed, which solved the classification error problem caused by the independent use of multi-source data and improved the accuracy and stability of vegetation classification results.

[0016] (3) By combining the enhanced AI computing architecture (the residual degenerate neural network of transfer learning and the random forest ensemble model) and the time-phase processing strategy, the multi-temporal object features and spectral features are effectively integrated, which significantly improves the distinguishability of complex vegetation types (especially shrubs and grasses) and realizes species-level identification and classification.

[0017] (4) A multi-temporal object association model based on transfer learning was designed, and an algorithm framework was established that can invert the dynamic changes of vegetation every five years from 1985 to 2025, supporting the retrospective analysis of historical vegetation dynamics, and providing key technical support for ecological succession research, carbon sink assessment, etc.

[0018] In addition, the vegetation species identification method with enhanced AI computing power and time-series tracing disclosed in this invention may also have the following additional technical features:

[0019] In one embodiment of the present invention, in step S1:

[0020] The space-based remote sensing data includes WorldImagery tile map images, Gaofen-1 multispectral images, and historical satellite images; the air-based remote sensing data includes UAV visible light images; and the ground-based spectral data includes vegetation canopy spectral data collected by a portable hyperspectral instrument. The preprocessing includes resolution unification processing, georegistration, and coordinate system standardization, with the coordinate system standardization using the WGS84 coordinate system.

[0021] In one embodiment of the present invention, in step S1:

[0022] The resolution unification process specifically involves upscaling the Gaofen-1 multispectral imagery to 2-meter resolution using panchromatic multispectral band fusion technology, downscaling historical satellite imagery to 2-meter resolution using pyramid resampling technology, and resampling UAV visible light imagery into a 2-meter resolution grid after georegistration.

[0023] In one embodiment of the present invention, in step S2:

[0024] The vegetation mask generation process involves selecting vegetation types based on land use data at a preset resolution, and combining this with the NDVI index of multispectral imagery to generate a 2-meter resolution vegetation mask, wherein the NDVI index threshold is set to be greater than 0.2.

[0025] In one embodiment of the present invention, in step S2:

[0026] The spectral features include species spectral fingerprints collected by a handheld hyperspectral analyzer and satellite multispectral vegetation indices. The spectral features are constructed as a 24-dimensional feature set, and the spectral features of the shrub-grass transition zone are supplemented by manually labeled sample points, which include 20% on-site locations and 80% visual locations.

[0027] In one embodiment of the present invention, in step S3:

[0028] The time-phase processing specifically involves preparing forest land thematic slices in April when the leaf unfolding characteristics of trees are significant, and preparing shrub-grass composite slices in July and August when the spectral response of shrubs and grasses is strong; the enhanced AI model adopts a residual degradation neural network architecture based on transfer learning, loads ImageNet pre-trained weights, and performs fine-tuning of the fully connected layer using UAV image supervision samples.

[0029] In one embodiment of the present invention, in step S3:

[0030] The enhanced AI model outputs a species-level probability distribution matrix at a 2×2 meter pixel scale. The probability distribution matrix includes independent probability layers for woodland, shrubs, and herbs. The preliminary vegetation classification result is determined by extracting the maximum probability value from the probability distribution matrix.

[0031] In one embodiment of the present invention, in step S4:

[0032] The object-level optimization specifically involves generating homogeneous vegetation objects based on superpixel segmentation, integrating 24-dimensional spectral features into a random forest ensemble model, which contains hundreds of decision trees, and performing mode statistical correction on pixels within the object based on the GINI coefficient splitting criterion.

[0033] In one embodiment of the present invention, in step S5:

[0034] The target year is 2025, the specified time period is from 1985 to 2025, and the time interval for vegetation dynamic change inversion is once every five years.

[0035] In one embodiment of the present invention, in step S1:

[0036] When Gaofen-1 multispectral images are severely obscured by clouds and fog, Sentinel-2 images are used as an auxiliary verification source. They are manually identified through visual interpretation and cross-verified with ground survey samples.

[0037] Additional features and advantages of this invention will be set forth in the description which follows, or may be learned by practicing the invention. Attached Figure Description

[0038] The technical solution and beneficial effects of the present invention will become apparent and readily understood from the following description in conjunction with the accompanying drawings, wherein:

[0039] Figure 1 This is an architecture diagram of data acquisition and fusion for the vegetation species identification method that enhances AI computing power and time-series tracing according to the present invention;

[0040] Figure 2 This is an architecture diagram of the sample construction and preprocessing of the vegetation species identification method for enhancing AI computing power and time-series tracing in this invention;

[0041] Figure 3 This is an architecture diagram of the object association feature of the enhanced AI computing power driven framework of the vegetation species identification method for enhanced AI computing power and time-series tracing of the present invention.

[0042] Figure 4 This is an architecture diagram of the enhanced AI decision-making mechanism of the enhanced AI computing power driven framework of the vegetation species identification method with enhanced AI computing power and time-series tracing of the present invention.

[0043] Figure 5 This is an architecture diagram of the classification results of the vegetation species identification method with enhanced AI computing power and time-series tracing of the present invention. Detailed Implementation

[0044] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0045] The vegetation species identification method with enhanced AI computing power and time-series tracing disclosed in this invention will now be described with reference to the accompanying drawings.

[0046] like Figure 1 , Figure 2 , Figure 3 , Figure 4 and Figure 5 As shown, the method for identifying vegetation species by enhancing AI computing power and time-series tracing includes the following steps:

[0047] S1: Collect and preprocess multi-source remote sensing data to construct a spatiotemporally aligned and resolution-uniform vegetation classification dataset. The multi-source remote sensing data includes space-based remote sensing data, airborne remote sensing data, and ground-based spectral data.

[0048] S2: Generate a vegetation mask based on the vegetation classification dataset, cut the multi-source remote sensing data through the vegetation mask to obtain standardized sample slices, and construct a training dataset by combining spectral features.

[0049] S3: Based on vegetation phenological characteristics, the training dataset is processed by time phase, and an enhanced AI model is used to extract features and classify the processed training dataset, outputting preliminary vegetation classification results.

[0050] S4: Perform object-level optimization on the preliminary vegetation classification results, eliminate abnormal pixels, and obtain optimized vegetation classification results;

[0051] S5: Correct the optimized vegetation classification results by combining terrain data, generate a vegetation species classification map for the target year, and realize the inversion of vegetation dynamic changes within a specified time period through transfer learning.

[0052] In step S1:

[0053] Space-based remote sensing data includes WorldImagery tile map images, Gaofen-1 multispectral images, and historical satellite images; airborne remote sensing data includes UAV visible light images; and ground-based spectral data includes vegetation canopy spectral data collected by a portable hyperspectral instrument. Preprocessing includes resolution unification, georegistration, and coordinate system standardization, with the WGS84 coordinate system adopted for coordinate system standardization.

[0054] In step S1:

[0055] The resolution is unified by upscaling the Gaofen-1 multispectral imagery to 2-meter resolution using panchromatic multispectral band fusion technology, downscaling historical satellite imagery to 2-meter resolution using pyramid resampling technology, and resampling UAV visible light imagery into a 2-meter resolution grid after georegistration.

[0056] In step S2:

[0057] The vegetation mask is generated by selecting vegetation types based on land use data at a preset resolution and combining the NDVI index of multispectral imagery to generate a 2-meter resolution vegetation mask, wherein the NDVI index threshold is set to be greater than 0.2.

[0058] In step S2:

[0059] Spectral features include species spectral fingerprints collected by a handheld hyperspectral analyzer and satellite multispectral vegetation indices. The spectral features are constructed as a 24-dimensional feature set, and the spectral features of the shrub-grass transition zone are supplemented by manually labeled sample points, which include 20% on-site locations and 80% visual locations.

[0060] In step S3:

[0061] The process involved time-phase processing, specifically preparing thematic slices of forest land in April when the leaf unfolding characteristics of trees are significant, and preparing composite slices of shrubs and grasses in July and August when the spectral response of shrubs and grasses is strong. The AI ​​model was enhanced by adopting a residual degradation neural network architecture based on transfer learning, loading ImageNet pre-trained weights, and fine-tuning the fully connected layers using UAV image supervision samples.

[0062] In step S3:

[0063] The AI ​​model outputs a species-level probability distribution matrix at a 2×2 meter pixel scale. The probability distribution matrix includes independent probability layers for woodland, shrubs, and herbs. The preliminary vegetation classification results are determined by extracting the maximum probability value from the probability distribution matrix.

[0064] In step S4:

[0065] Object-level optimization specifically involves generating homogeneous vegetation objects based on superpixel segmentation, integrating 24-dimensional spectral features into a random forest ensemble model, which contains hundreds of decision trees, and performing mode statistical correction on pixels within the object based on the GINI coefficient splitting criterion.

[0066] In step S5:

[0067] The target year is 2025, the specified time period is from 1985 to 2025, and the vegetation dynamics inversion interval is once every five years.

[0068] In step S1:

[0069] When Gaofen-1 multispectral images are severely obscured by clouds and fog, Sentinel-2 images are used as an auxiliary verification source. They are manually identified through visual interpretation and cross-verified with ground survey samples.

[0070] Specifically:

[0071] Step S1 aims to acquire and preprocess multi-source remote sensing data to construct a vegetation classification dataset. This step integrates space-based, air-based, and ground-based multi-source data, and after standardization, forms a spatiotemporally aligned vegetation classification dataset with uniform resolution. The specific process is as follows:

[0072] For space-based data acquisition and processing, World Imagery tile map imagery (original spatial resolution 2 meters) and Gaofen-1 multispectral imagery (original resolution 8 meters) were used. The resolution of Gaofen-1 imagery was improved to 2 meters using panchromatic-multispectral band fusion technology in ENVI software. For long-term time series requirements from 1985 to 2025, Google high-resolution satellite imagery maps (maximum resolution 0.5 meters) were called through the Google Earth Engine platform. Pyramid resampling technology was used to uniformly downscale them to 2-meter resolution to construct a historical image sequence with full coverage.

[0073] For space-based data acquisition and processing, a UAV visible light system is deployed to acquire high-resolution images of local areas (original resolution 0.3 meters). After georegistration, the images are resampled to a 2-meter resolution grid to ensure consistency with the spatial scale of space-based data.

[0074] For ground-based data acquisition and processing, a portable hyperspectral instrument (wavelength range 400-2500nm, spectral resolution 5nm) was used to simultaneously collect vegetation canopy spectral data in the field plots, with positioning accuracy controlled within ±1 meter; the hyperspectral data was matched with UAV visible light images at the pixel level to construct a species spectral fingerprint database; for the shrub-grass transition zone, 10 additional narrow-band indices (such as PSRI and PRI) were extracted to supplement the feature dimensions of the multispectral vegetation indices.

[0075] For data standardization, all data are georeferenced to the WGS84 coordinate system to ensure spatial consistency. If the Gaofen-1 data is severely obscured by clouds and fog, Sentinel-2 imagery (10-meter resolution) is used as an auxiliary verification source. Visual interpretation is performed on the imagery for manual identification, and cross-verification is performed with ground survey samples to ultimately form a multi-scale vegetation classification dataset that is spatiotemporally aligned and has a unified resolution.

[0076] Step S2 aims to construct the training dataset. Based on the vegetation classification dataset constructed in S1, training samples and feature sets are generated. The specific process is as follows:

[0077] For vegetation mask generation, CLCD 30m resolution land use data was selected, and vegetation types such as woodland and grassland were manually screened. Combined with the NDVI index of Gaofen-1 multispectral image (with a set threshold of NDVI>0.2), a 2-meter resolution vegetation mask was generated to define the vegetation coverage area.

[0078] For standardized sample slices, the World Imagery tile map is cut using the above-mentioned vegetation mask, and only the vegetation area is processed to generate standardized slices of 224×224 pixels; at the same time, the geo-registered UAV imagery is sliced ​​into slices of the same size, which together serve as input samples to strengthen the AI ​​classification network.

[0079] For the construction of the spectral feature set, the species spectral fingerprints collected by handheld hyperspectral instruments are merged with satellite multispectral vegetation indices (such as NDVI, EVI, and SAVI) to construct a 24-dimensional spectral feature set. Feature information is supplemented by manually annotating sample points (20% of which are field survey points and 80% are visual interpretation points), with a focus on annotating the spectral features of the shrub-grass transition zone, ensuring that the samples cover all typical vegetation community zones, and forming a feature library that strengthens the AI ​​decision-making mechanism.

[0080] Step S3 aims to enhance AI classification and preliminary result generation. Based on vegetation phenological characteristics, samples are processed according to time phases. Features are extracted and classified by enhancing the AI ​​model. The specific process is as follows:

[0081] For the preparation of thematic slices based on time phases, according to the phenological differentiation patterns of vegetation, thematic slices of forest land are prepared in April when the leaf unfolding characteristics of trees are significant, and composite slices of shrubs and grasses are prepared in July and August when the spectral response of shrubs and grasses is strong, so as to improve the characteristic identification of different vegetation types.

[0082] To enhance AI model training and classification, a residual degenerate neural network architecture based on transfer learning is adopted, with ImageNet pre-trained weights loaded. The fully connected layers are fine-tuned using supervised samples of UAV images of the same size to achieve deep adaptation of phenological features. The model extracts multi-temporal object association features and outputs a species-level probability distribution matrix (including independent probability layers for woodland, shrubs, and herbs) at a 2×2 meter pixel scale. The maximum probability value of each pixel in the matrix is ​​extracted using the ArcGIS spatial analysis module to determine the dominant species and generate a preliminary vegetation classification map.

[0083] Step S4 aims to optimize at the object level, eliminate abnormal pixels, and optimize the preliminary classification results generated in S3. The specific process is as follows:

[0084] For the generation of homogeneous vegetation objects, a superpixel segmentation algorithm is used to aggregate pixels with similar spectral and texture features in the preliminary classification image into homogeneous vegetation objects, thereby reducing pixel-level noise interference.

[0085] For multi-feature ensemble optimization, an enhanced AI decision-making mechanism is activated, integrating the 24-dimensional spectral feature set (including hyperspectral narrow band index and multispectral index) constructed in S2, and inputting it into a random forest ensemble model with a scale of 100 decision trees; the model performs mode statistical correction on all pixels in each homogeneous object according to the GINI coefficient splitting criterion, eliminates abnormal pixels in the primary classification, and obtains the optimized vegetation classification results.

[0086] Step S5 aims at terrain correction and long-term dynamic inversion, correcting the optimized classification results and realizing the dynamic reconstruction of historical vegetation. The specific process is as follows:

[0087] For terrain correction, the optimized vegetation classification results are overlaid with the digital elevation model (DEM) to correct the terrain slope and aspect, generating a 2-meter resolution forest-shrub-grass object-level species classification map for 2025.

[0088] For long-term dynamic inversion, based on the historical image sequence in S1 and the enhanced AI model in S3, the vegetation spectral characteristics of different eras are adapted through transfer learning algorithms to realize the inversion of vegetation dynamic changes every five years from 1985 to 2025, and output the distribution and succession patterns of vegetation species at each time point.

[0089] In summary, the vegetation species identification method with enhanced AI computing power and temporal tracing disclosed in this invention can overcome the limitations of traditional resolution through multi-source data fusion and unified resolution processing, achieving species-accurate vegetation classification at the 2-meter level. It can effectively distinguish fragmented vegetation patches such as shrub-grass transition zones, meeting the needs of fine classification at the forest-shrub-grass species level. A three-dimensional monitoring data collaboration mechanism with nested sky-ground "spatiotemporal spectrum" multi-source information is constructed, solving the classification error problem caused by independent use of multi-source data and improving the accuracy and stability of vegetation classification results. Combining the enhanced AI computing power architecture (transfer learning residual degenerate neural network and random forest ensemble model) and temporal phase processing strategy, it effectively integrates multi-temporal object features and spectral features, significantly improving the distinguishability of complex vegetation types (especially shrubs and grasses), and achieving species-level identification and classification. A multi-temporal object association model based on transfer learning is designed, and an algorithm framework that can invert vegetation dynamic changes every five years from 1985 to 2025 is established, supporting historical vegetation dynamic retrospective analysis and providing key technical support for ecological succession research, carbon sink assessment, and other fields.

[0090] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for identifying vegetation species that enhances AI computing power and time-series tracing, characterized in that, Includes the following steps: S1: Collect multi-source remote sensing data and preprocess it to construct a vegetation classification dataset with spatiotemporal alignment and uniform resolution. The multi-source remote sensing data includes space-based remote sensing data, airborne remote sensing data and ground-based spectral data. S2: Based on land use data at a preset resolution, vegetation types are selected. Combined with the NDVI index of multispectral imagery, a 2-meter resolution vegetation mask is generated. The NDVI index threshold is set to be greater than 0.

2. The multi-source remote sensing data is then segmented using the vegetation mask to obtain standardized sample slices. A training dataset is constructed by combining these slices with spectral features. The spectral features include species spectral fingerprints collected by a handheld hyperspectral analyzer and satellite multispectral vegetation indices. The spectral features are constructed as a 24-dimensional feature set, and the spectral features of the shrub-grass transition zone are supplemented by manually labeled sample points, which include 20% on-site locations and 80% visual locations. S3: Based on vegetation phenological characteristics, the training dataset is processed by time phase division. An enhanced AI model is then used to extract and classify features from the processed training dataset, outputting preliminary vegetation classification results. The enhanced AI model adopts a ResNet50 residual degradation neural network architecture based on transfer learning, loads ImageNet pre-trained weights, and performs fine-tuning of the fully connected layers using drone image supervision samples. S4: Perform object-level optimization on the preliminary vegetation classification results to eliminate abnormal pixels and obtain optimized vegetation classification results, specifically: For the generation of homogeneous vegetation objects, a superpixel segmentation algorithm is used to aggregate pixels with similar spectral and texture features in the preliminary classification image into homogeneous vegetation objects, thereby reducing pixel-level noise interference. For multi-feature ensemble optimization, an enhanced AI decision-making mechanism is activated, integrating the 24-dimensional spectral feature set constructed in S2 and inputting it into a random forest ensemble model with a scale of hundreds of decision trees; The model performs mode statistical correction on all pixels within each homogeneous object based on the GINI coefficient splitting criterion, eliminating outlier pixels in the primary classification and obtaining optimized vegetation classification results. S5: Correct the optimized vegetation classification results by combining terrain data, generate a vegetation species classification map for the target year, and realize the inversion of vegetation dynamic changes within a specified time period through transfer learning.

2. The vegetation species identification method with enhanced AI computing power and time-series tracing as described in claim 1, characterized in that, In step S1: The space-based remote sensing data includes WorldImagery tile map images, Gaofen-1 multispectral images, and historical satellite images; the air-based remote sensing data includes UAV visible light images; and the ground-based spectral data includes vegetation canopy spectral data collected by a portable hyperspectral instrument. The preprocessing includes resolution unification processing, georegistration, and coordinate system standardization, with the coordinate system standardization using the WGS84 coordinate system.

3. The vegetation species identification method with enhanced AI computing power and time-series tracing as described in claim 2, characterized in that, In step S1: The resolution unification process specifically involves upscaling the Gaofen-1 multispectral imagery to 2-meter resolution using panchromatic multispectral band fusion technology, downscaling historical satellite imagery to 2-meter resolution using pyramid resampling technology, and resampling UAV visible light imagery into a 2-meter resolution grid after georegistration.

4. The vegetation species identification method with enhanced AI computing power and time-series tracing as described in claim 1, characterized in that, In step S3: The time-phase treatment specifically involves preparing forest land thematic slices in April when the leaf unfolding characteristics of trees are significant, and preparing shrub-grass composite slices in July and August when the spectral response of shrubs and grasses is strong.

5. The vegetation species identification method with enhanced AI computing power and time-series tracing as described in claim 1, characterized in that, In step S3: The enhanced AI model outputs a species-level probability distribution matrix at a 2×2 meter pixel scale. The probability distribution matrix includes independent probability layers for woodland, shrubs, and herbs. The preliminary vegetation classification result is determined by extracting the maximum probability value from the probability distribution matrix.

6. The vegetation species identification method with enhanced AI computing power and time-series tracing as described in claim 1, characterized in that, In step S5: The target year is 2025, the specified time period is from 1985 to 2025, and the time interval for vegetation dynamic change inversion is once every five years.

7. The vegetation species identification method with enhanced AI computing power and time-series tracing as described in claim 2, characterized in that, In step S1: When Gaofen-1 multispectral images are severely obscured by clouds and fog, Sentinel-2 images are used as an auxiliary verification source. They are manually identified through visual interpretation and cross-verified with ground survey samples.

Images