Ecosystem remote sensing assessment method based on depth image features

Abstract

The present invention relates to the technical field of ecosystem remote sensing evaluation, specifically to an ecosystem remote sensing evaluation method based on deep image features, comprising the following steps: ecosystem specific data collection and customized preprocessing, ecological scene simulation and data enhancement, starting the construction of a CNN basic feature extractor and performing ecological pre-training, generating ecological images through GAN and optimizing stability, screening ecological features, then performing image verification through GAN, performing multi-source feature fusion and constructing ecological indicators, performing comprehensive evaluation of ecosystem remote sensing data, and implementing a feedback iteration mechanism. The present invention integrates features extracted by CNN, image features generated by GAN, and meteorological data and terrain data to construct a comprehensive ecosystem feature representation and ecological indicators, providing a quantitative basis for evaluation, evaluating model performance by comparing real ecological data and prediction results, and performing feedback iteration according to the evaluation results, thereby improving the accuracy and reliability of ecosystem remote sensing evaluation.

Description

Technical Field

[0001] The present invention relates to the technical field of ecosystem remote sensing evaluation, and in particular to an ecosystem remote sensing evaluation method based on depth image features. Background Art

[0002] An ecosystem refers to a unified whole of interaction and interdependence formed between biological communities and their non-biological environment through material circulation, energy flow and information transfer within a certain space, including organisms and the inorganic environment in which they live.

[0003] Ecosystem assessment is an important means to evaluate the health status, functional integrity and response of ecosystems to human activities. It helps to understand the dynamic changes of ecosystems and provide a scientific basis for ecological protection, resource management and sustainable development. With the intensification of global environmental changes, the importance of ecosystem assessment has become increasingly prominent. Traditional ecosystem assessment methods mainly rely on ground-based measured data, such as vegetation sample surveys and soil moisture measurements. Although these methods are accurate, they are limited by spatial coverage and timeliness, and it is difficult to fully reflect the overall status of the ecosystem.

[0004] Therefore, remote sensing technology provides new perspectives and tools for ecosystem evaluation. Remote sensing images obtained by satellites or drones can achieve large-scale and continuous monitoring of ecosystems. However, traditional remote sensing evaluation methods often rely on simple spectral analysis or vegetation index calculations, which make it difficult to capture the complex characteristics and dynamic changes of ecosystems. In addition, the quality of remote sensing images is affected by cloud cover, shadows, and noise, requiring a complex preprocessing process.

[0005] The traditional steps of ecosystem remote sensing assessment include data collection, preprocessing, feature extraction, evaluation and analysis. Therefore, the feature extraction is single and it is difficult to fully reflect the characteristics of the ecosystem. It has high requirements for image quality and the preprocessing process is complex, making it difficult to capture the dynamic changes and subtle differences in the ecosystem.

[0006] In summary, it is necessary to propose an ecosystem remote sensing evaluation method based on deep image features to solve the above problems. Summary of the invention

[0007] The purpose of the present invention is to provide an ecosystem remote sensing evaluation method based on depth image features to solve the problems raised in the above background technology.

[0008] To achieve the above object, the present invention provides the following technical solutions: The ecosystem remote sensing assessment method based on deep image features includes the following steps: S1. Ecosystem-specific data collection and customized pre-processing; S2. Conduct ecological scenario simulation and data enhancement; S3. Start building the CNN basic feature extractor and perform ecological pre-training; S4. Generate ecological images through GAN and optimize stability; S5. Screen ecological features and then perform image verification through GAN; S6. Introduce attention mechanism to refine CNN features; S7. Conduct ecological transfer learning and model adaptation; S8. Fusion of multi-source features and construction of ecological indicators; S9. Conduct comprehensive evaluation of ecosystem remote sensing data and implement feedback iteration mechanisms.

[0009] Preferably, the implementation process of step S1 is: S1.1. Collect multispectral and hyperspectral remote sensing images from different satellites, drone aerial images, and ground-based data, including vegetation plot surveys and soil moisture measurements, to provide high-resolution details as validation data; S1.2. Use masking techniques to remove cloud cover and shadow interference factors and unify data from different sources to the same coordinate system, resolution and format; S1.3. Apply adaptive filtering methods to remove peculiar noise, use empirical linear correction methods to adjust the spectral response, and then use a deep super-resolution network to enhance image details.

[0010] Preferably, the implementation process of step S2 is: S2.1. Simulate the dynamic changes of ecosystems based on the DNDC ecological process model, input climate and soil type parameters, and generate simulation scenarios under different seasons and weather conditions; S2.2. Apply rotation, scaling, and flipping operations to increase data diversity, and use color jitter to simulate image changes under different lighting conditions; S2.3. Combine the simulated data with the real remote sensing images, use the hybrid strategy to generate mixed samples, and control the weights of the real and simulated data.

[0011] Preferably, the implementation process of step S3 is: S3.1. Construct a deep convolutional neural network to adapt to remote sensing image feature extraction, and capture the unique texture and shape characteristics of the ecosystem through the convolution layer; S3.2. Use ecosystem remote sensing image dataset for pre-training, introduce ecosystem vegetation type and soil moisture labels, and use cross entropy loss function for training; S3.3. Learning preliminary ecosystem characteristics through pre-trained models.

[0012] Preferably, the implementation process of step S4 is: S4.1. Design a generator-discriminator structure to generate remote sensing images with ecosystem characteristics, and introduce the ecological balance principle as an ecological constraint as a regularization term for the generator; S4.2. Use WGAN to reduce the risk of mode collapse, introduce gradient penalty terms, and stabilize the training process; S4.3. Define ecological indicators such as vegetation index or species diversity index to evaluate the quality of the generated image.

[0013] Preferably, the implementation process of step S5 is: S5.1. Use ecological trait screening methods to retain key ecological traits; S5.2. Visually interpret and verify the screened images and adjust the screening threshold based on expert or historical data feedback to ensure image quality.

[0014] Preferably, the implementation process of step S6 is: S6.1. Introduce channel attention or spatial attention ecological attention mechanism; S6.2. During training, the model is trained to focus on key areas related to ecosystem assessment and key areas are visualized using attention maps.

[0015] Preferably, the implementation process of step S7 is: S7.1. Based on the pre-trained CNN model, fine-tune it using the historical data of the ecosystem; S7.2. Adjust model parameters based on ecological knowledge to improve adaptability to new ecosystems, use cross-validation to evaluate model performance, and select the optimal parameter combination.

[0016] Preferably, the implementation process of step S8 is: S8.1. Use network feature fusion methods such as concatenation, weighting, or deep fusion to fuse features from CNN, GAN, and other data sources; S8.2. Construct comprehensive ecological indicators based on ecosystem assessment needs.

[0017] Preferably, the implementation process of step S9 is: S9.1. Use the fused feature representation and ecological indicators to classify and assess the quality of ecosystems, using SVM or random forest classifiers for classification and linear regression models for quality assessment; S9.2. Compare with real ecological data to evaluate model accuracy, recall rate, and F1 score indicators; S9.3. Based on the evaluation results, adjust model parameters, feature selection, and ecological indicators, and iterate training until the model performance is stable and meets the needs of ecosystem remote sensing evaluation.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: the ecosystem remote sensing evaluation method based on deep image features of the present invention combines deep learning technology and ecosystem remote sensing evaluation needs, integrates multi-source data from different satellites, drones and ground measurements, performs customized preprocessing such as noise removal, spectral correction and spatial resolution enhancement for specific ecosystems, uses CNN models to extract deep features of remote sensing images, simulates ecosystem scenes under different conditions through ecological models, combines data enhancement technology to construct a rich training set, improves the generalization ability of the model, improves the richness and accuracy of data, introduces an ecological attention mechanism, enables the model to focus on key areas related to ecosystem evaluation, improves the accuracy and pertinence of feature extraction, and simultaneously integrates features extracted by CNN, image features generated by GAN, and meteorological data and terrain data to construct a comprehensive ecosystem feature representation and ecological indicators, providing a quantitative basis for evaluation, and evaluating model performance by comparing real ecological data and prediction results. Feedback iteration is performed based on the evaluation results to continuously optimize model parameters and feature representation, thereby improving the accuracy and reliability of ecosystem remote sensing evaluation. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 A flow chart showing the ecosystem remote sensing evaluation method based on depth image features of the present invention. DETAILED DESCRIPTION

[0020] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0021] See also Figure 1 The present invention proposes an ecosystem remote sensing evaluation method based on deep image features, comprising the following steps: S1. Collect multi-source and multi-temporal remote sensing image data for specific ecosystems and conduct customized pre-processing, including noise removal, spectral correction, and spatial resolution enhancement specific to the ecosystem, to ensure that the data quality meets the requirements of subsequent analysis and enhances the correlation with ecosystem characteristics; S2. Use ecological models to simulate ecosystem scenarios under different seasons and weather conditions, generate diverse simulation data, and combine data rotation and scaling enhancement methods with simulated data to build a rich training set, reduce dependence on actual labeled data, and enhance the model's adaptability to ecosystem changes; It should be noted that the implementation process of step S2 is: S2.1. Simulate the dynamic changes of ecosystems based on the DNDC ecological process model, input climate and soil type parameters, and generate simulation scenarios under different seasons and weather conditions; S2.2. Apply rotation, scaling, and flipping operations to increase data diversity, and use color jitter to simulate image changes under different lighting conditions; S2.3. Combine simulated data with real remote sensing images, use a hybrid strategy to generate mixed samples, and control the weights of real and simulated data; S3. Build a CNN model for remote sensing images and use a large-scale, diverse ecosystem remote sensing image dataset for pre-training. At the same time, introduce ecosystem-specific vegetation types and soil moisture labels to enable the model to initially learn ecosystem feature representations, laying the foundation for subsequent tasks. It should be noted that the implementation process of step S3 is as follows: S3.1. Construct a deep convolutional neural network to adapt to remote sensing image feature extraction, and capture the unique texture and shape characteristics of the ecosystem through the convolution layer; S3.2. Use ecosystem remote sensing image dataset for pre-training, introduce ecosystem vegetation type and soil moisture labels, and use cross entropy loss function for training; S3.3. Learn preliminary ecosystem features through pre-trained models; S4. Train the GAN model to generate high-quality remote sensing images with ecosystem characteristics. During the training process, the ecological balance principle and species distribution law ecological constraints are introduced as regularization terms to stabilize the training process and reduce mode collapse. At the same time, ecological indicators are used to evaluate the quality of the generated images to ensure that they are consistent with the real ecosystem. It should be noted that the implementation process of step S4 is as follows: S4.1. Design a generator-discriminator structure to generate remote sensing images with ecosystem characteristics, and introduce the ecological balance principle as an ecological constraint as a regularization term for the generator; S4.2. Use WGAN to reduce the risk of mode collapse, introduce gradient penalty terms, and stabilize the training process; S4.3. Define ecological indicators such as vegetation index or species diversity index to evaluate the quality of generated images; S5. Screen the remote sensing images generated by GAN for ecological characteristics, retain the images that meet the characteristics of the ecosystem, and then verify the screened images to ensure that the image quality meets the requirements of subsequent analysis and enhances the relevance to the remote sensing evaluation of the ecosystem; It should be noted that the implementation process of step S5 is as follows: S5.1. Use ecological trait screening methods to retain key ecological traits; S5.2. Visually interpret and verify the screened images and adjust the screening threshold based on expert or historical data feedback to ensure image quality; S6. Introduce the ecological attention mechanism into the CNN model, so that the model can focus on the vegetation coverage and water pollution key areas related to ecosystem evaluation in remote sensing images. The guidance of the attention mechanism is used to provide refined feature representation for ecosystem remote sensing evaluation. It should be noted that the implementation process of step S6 is as follows: S6.1. Introduce channel attention or spatial attention ecological attention mechanism; S6.2. During training, the model is trained to focus on key areas related to ecosystem assessment and key areas are visualized using attention maps. S7. Use the pre-trained CNN model and remote sensing image data generated by GAN, combined with historical data of specific ecosystems for transfer learning, so that the model can quickly adapt to the remote sensing image evaluation task of specific ecosystems and enhance the sensitivity and accuracy of the model to ecosystem changes; It should be noted that the implementation process of step S7 is as follows: S7.1. Based on the pre-trained CNN model, fine-tune it using the historical data of the ecosystem; S7.2. Adjust model parameters based on ecological knowledge to improve adaptability to new ecosystems, use cross-validation to evaluate model performance, and select the optimal parameter combination; S8. The features extracted by CNN, the image features generated by GAN, and the meteorological data source and terrain data source unique to the ecosystem are integrated to construct a comprehensive ecosystem feature representation through the integration of multi-source features. Then, according to the needs of ecosystem evaluation, ecological indicators are constructed as the basis for quantitative evaluation. It should be noted that the implementation process of step S8 is as follows: S8.1. Use network feature fusion methods such as concatenation, weighted sum, or deep fusion to fuse features from CNN, GAN, and other data sources; S8.2. Construct comprehensive ecological indicators based on ecosystem assessment needs; S9. Use the fused feature representation and ecological indicators to conduct comprehensive remote sensing evaluation of ecosystem types and ecological quality assessment. Evaluate the performance of the model by comparing real ecological data and prediction results, and conduct feedback iteration based on the evaluation results to continuously optimize the model parameters and feature representation to improve the accuracy and reliability of remote sensing evaluation of ecosystems.

[0022] It should be noted that the implementation process of step S9 is as follows: S9.1. Use the fused feature representation and ecological indicators to classify and assess the quality of ecosystems, using SVM or random forest classifiers for classification and linear regression models for quality assessment; S9.2. Compare with real ecological data to evaluate model accuracy, recall rate, and F1 score indicators; S9.3. Based on the evaluation results, adjust model parameters, feature selection, and ecological indicators, and iterate training until the model performance is stable and meets the needs of ecosystem remote sensing evaluation.

[0023] Example

[0024] See also Figure 1 In the actual application process, the above-mentioned ecosystem remote sensing evaluation method based on deep image features specifically includes the following steps: (1) Ecosystem-specific data collection and customized preprocessing: (1.1) Multi-source data collection: Collect satellite multispectral and hyperspectral remote sensing images from Landsat and Sentinel-2, collect drone aerial images to provide high-resolution details, and collect ground-truth data, such as vegetation plot surveys and soil moisture measurements, as validation data; (1.2) Data cleaning and format unification: Remove cloud cover and shadow interference factors and use masking technology; Unify data from different sources to the same coordinate system, resolution and format; (1.3) Customized preprocessing: (1.3.1) Noise removal: Lee adaptive filtering algorithm is used to remove the unique noise, see formula (1): (1); In the formula, is the original pixel value, and are the mean and variance of the local noise, is a small constant that prevents division by zero; (1.3.2) Spectral correction: The spectral response is adjusted using the empirical linear correction (ELC) method, see formula (2): (2); In the formula, is the correction coefficient, obtained by fitting the ground measured data; (1.3.3) Spatial resolution enhancement: Use super-resolution reconstruction algorithms, such as the Deep Super-Resolution Network (DSRN), to improve image details; (2) Ecological scenario simulation and data enhancement: (2.1) Ecological model construction: Simulate ecosystem dynamics based on DNDC or SWAT ecological process models; Input climate and soil type parameters to generate simulation scenarios under different seasons and weather conditions; (2.2) Traditional data enhancement: Apply rotation, scaling, and flipping operations to increase data diversity; Use color dithering to simulate image changes under different lighting conditions; (2.3) Simulation data fusion: Combine the simulated data with the real remote sensing image and use the hybrid strategy to generate mixed samples, see formula (3): (3); In the formula is the mixing ratio, which controls the weight of real and simulated data; (3) CNN basic feature extractor construction and ecological pre-training: (3.1) CNN model design: Build ResNet or EfficientNet deep convolutional neural networks to adapt to remote sensing image feature extraction; Design specific convolutional layers to capture the unique texture and shape characteristics of the ecosystem; (3.2) Large-scale dataset pre-training: Use a large-scale ecosystem remote sensing image dataset for pre-training; Introducing ecosystem-specific labels, such as vegetation type and soil moisture, and using the cross-entropy loss function for training; (3.3) Feature representation learning: Through pre-training, the model learns the initial feature representation of the ecosystem : in To input remote sensing images; (4) GAN ecological image generation and stability optimization: (4.1) GAN model construction: Design a generator-discriminator structure to generate remote sensing images with ecosystem characteristics; Introducing ecological constraints, such as the ecological balance principle, as a regularization term for the generator; (4.2) Stability optimization: Wasserstein GAN (WGAN) is used to reduce the risk of mode collapse, and a gradient penalty term is introduced to stabilize the training process, as shown in formula (4): (4); In the formula, is the penalty coefficient; (4.3) Ecological indicator assessment: Define ecological indicators, such as vegetation index (NDVI), species diversity index, and evaluate the quality of generated images: ,in are the near-infrared and red band reflectances, respectively; (5) Ecological feature screening and GAN image verification: (5.1) Feature screening: Use the random forest ecological feature screening algorithm to retain key ecological features: ,in Features of images generated by GAN; (5.2) Expert knowledge verification: Organize ecological experts to visually interpret and verify the screened images; Adjust screening thresholds based on expert feedback to ensure image quality; (6) CNN feature refinement guided by attention mechanism: (6.1) Attention mechanism design: Introducing channel attention (SE Block) or spatial attention (CBAM) ecological attention mechanism: ,in Features extracted for CNN; (6.2) Key Areas of Focus: During training, the model is made to focus on key areas relevant to ecosystem assessment; Use attention maps to visualize key areas, such as areas with high vegetation cover; (7) Ecological transfer learning and model adaptation: (7.1) Transfer learning framework: Based on the pre-trained CNN model, fine-tuning is performed using historical data of a specific ecosystem, as shown in formula (5): (5); In the formula, For mission loss, is the regularization term, is the weight; (7.2) Model adaptation: Incorporate ecological knowledge to adjust model parameters and improve adaptability to new ecosystems; Use cross-validation to evaluate model performance and select the optimal parameter combination; (8) Multi-source feature fusion and ecological indicator construction: (8.1) Feature fusion: Use concatenation, weighted sum or deep fusion network feature fusion algorithms to fuse CNN, GAN and other data source features: : (8.2) Construction of ecological indicators: Construct comprehensive ecological indicators based on ecosystem assessment needs; (9) Ecosystem remote sensing comprehensive evaluation and feedback iteration: (9.1) Comprehensive evaluation: Use the fused feature representations and ecological indicators to classify ecosystems and assess their quality; Use classifiers, such as SVM or random forest, for classification; use regression models, such as linear regression, for quality assessment; (9.2) Performance evaluation: Compare with real ecological data to evaluate model accuracy, recall rate, and F1 score indicators: ,in They are true positive, true negative, false positive, and false negative; (9.3) Feedback iteration: According to the evaluation results, adjust the model parameters, feature selection, and ecological indicators; Iterate training until the model performance is stable and meets the needs of ecosystem remote sensing evaluation.

[0025] Through the above steps, the present invention constructs the correlation between multi-source data and ecosystem remote sensing evaluation through steps S1 to S9, and provides a comprehensive, accurate and stable ecosystem remote sensing evaluation method through customized preprocessing, ecological scene simulation, ecological pre-training, ecological feature screening, attention mechanism guidance, ecological transfer learning, multi-source feature fusion and ecological indicator construction.

[0026] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. Ecosystem remote sensing assessment method based on deep image features, characterized in that: The following steps are involved: S1. Ecosystem-specific data collection and customized pre-processing; S2. Conduct ecological scenario simulation and data enhancement; S3. Start building the CNN basic feature extractor and perform ecological pre-training; S4. Generate ecological images through GAN and optimize stability; S5. Screen ecological features and then perform image verification through GAN; S6. Introduce attention mechanism to refine CNN features; S7. Conduct ecological transfer learning and model adaptation; S8. Fusion of multi-source features and construction of ecological indicators; S9. Conduct comprehensive evaluation of ecosystem remote sensing data and implement feedback iteration mechanisms.

2. The ecosystem remote sensing evaluation method based on deep image features according to claim 1 is characterized in that: The implementation process of step S1 is as follows: S1.

1. Collect multispectral and hyperspectral remote sensing images from different satellites, drone aerial images, and ground-based data, including vegetation plot surveys and soil moisture measurements, to provide high-resolution details as validation data; S1.

2. Use masking techniques to remove cloud cover and shadow interference factors and unify data from different sources to the same coordinate system, resolution and format; S1.

3. Apply adaptive filtering methods to remove peculiar noise, use empirical linear correction methods to adjust the spectral response, and then use a deep super-resolution network to enhance image details.

3. The ecosystem remote sensing evaluation method based on deep image features according to claim 2 is characterized in that: The implementation process of step S2 is as follows: S2.

1. Simulate the dynamic changes of ecosystems based on the DNDC ecological process model, input climate and soil type parameters, and generate simulation scenarios under different seasons and weather conditions; S2.

2. Apply rotation, scaling, and flipping operations to increase data diversity, and use color jitter to simulate image changes under different lighting conditions; S2.

3. Combine the simulated data with the real remote sensing images, use the hybrid strategy to generate mixed samples, and control the weights of the real and simulated data.

4. The ecosystem remote sensing evaluation method based on deep image features according to claim 3 is characterized in that: The implementation process of step S3 is as follows: S3.

1. Construct a deep convolutional neural network to adapt to remote sensing image feature extraction, and capture the unique texture and shape characteristics of the ecosystem through the convolution layer; S3.

2. Use ecosystem remote sensing image dataset for pre-training, introduce ecosystem vegetation type and soil moisture labels, and use cross entropy loss function for training; S3.

3. Learning preliminary ecosystem characteristics through pre-trained models.

5. The ecosystem remote sensing evaluation method based on deep image features according to claim 4 is characterized in that: The implementation process of step S4 is as follows: S4.

1. Design a generator-discriminator structure to generate remote sensing images with ecosystem characteristics, and introduce the ecological balance principle as an ecological constraint as a regularization term for the generator; S4.

2. Use WGAN to reduce the risk of mode collapse, introduce gradient penalty terms, and stabilize the training process; S4.

3. Define ecological indicators such as vegetation index or species diversity index to evaluate the quality of the generated image.

6. The ecosystem remote sensing evaluation method based on deep image features according to claim 5 is characterized in that: The implementation process of step S5 is as follows: S5.

1. Use ecological trait screening methods to retain key ecological traits; S5.

2. Visually interpret and verify the screened images and adjust the screening threshold based on expert or historical data feedback to ensure image quality.

7. The ecosystem remote sensing assessment method based on deep image features according to claim 6 is characterized in that: The implementation process of step S6 is as follows: S6.

1. Introduce channel attention or spatial attention ecological attention mechanism; S6.

2. During training, the model is trained to focus on key areas related to ecosystem assessment and key areas are visualized using attention maps.

8. The ecosystem remote sensing assessment method based on deep image features according to claim 7 is characterized in that: The implementation process of step S7 is as follows: S7.

1. Based on the pre-trained CNN model, fine-tune it using the historical data of the ecosystem; S7.

2. Adjust model parameters based on ecological knowledge to improve adaptability to new ecosystems, use cross-validation to evaluate model performance, and select the optimal parameter combination.

9. The ecosystem remote sensing assessment method based on deep image features according to claim 8 is characterized in that: The implementation process of step S8 is as follows: S8.

1. Use network feature fusion methods such as concatenation, weighting, or deep fusion to fuse features from CNN, GAN, and other data sources; S8.

2. Construct comprehensive ecological indicators based on ecosystem assessment needs.

10. The ecosystem remote sensing assessment method based on deep image features according to claim 9, characterized in that: The implementation process of step S9 is as follows: S9.

1. Use the fused feature representation and ecological indicators to classify and assess the quality of ecosystems, using SVM or random forest classifiers for classification and linear regression models for quality assessment; S9.

2. Compare with real ecological data to evaluate model accuracy, recall rate, and F1 score indicators; S9.

3. Based on the evaluation results, adjust model parameters, feature selection, and ecological indicators, and iterate training until the model performance is stable and meets the needs of ecosystem remote sensing evaluation.

Images