Remote sensing ecological index assessment method based on deep learning

Abstract

The invention relates to the technical field of remote sensing, and discloses a remote sensing ecological index assessment method based on deep learning, which comprises the following steps: acquiring Landsat remote sensing image data and land utilization data of a research area, and extracting remote sensing ecological index factors of the land utilization data and the preprocessed Landsat remote sensing image data; normalizing the remote sensing ecological index factors and performing dimension reduction analysis to obtain dimension-reduced remote sensing ecological index factors, inputting the dimension-reduced remote sensing ecological index factors into a time sequence convolutional network model, and generating a comprehensive feature vector through causal convolution, extended convolution and global average pooling operations; performing linear combination on the comprehensive feature vector and the principal component analysis feature value, and constructing a time sequence convolutional neural network remote sensing ecological index model for evaluating the ecological environment quality condition of the research area; according to the method, the accuracy, the reliability and the practicability of remote sensing ecological index evaluation are improved by adding the remote sensing ecological landmark factors reflecting human disturbance and water environment change and constructing the time sequence convolutional neural network remote sensing ecological index model.

Description

Technical Field

[0001] The present invention relates to the field of remote sensing technology, and in particular to a remote sensing ecological index evaluation method based on deep learning. Background Art

[0002] In the context of rapid urbanization, the deterioration of the ecological environment has become increasingly prominent. Problems such as soil erosion, reduced vegetation cover, and changes in hydrological conditions caused by human activities have had a significant impact on the ecosystem. In order to effectively monitor and evaluate changes in the regional ecological environment, remote sensing technology has been widely used as a non-contact monitoring method. The Remote Sensing Ecological Index (RSEI), as an ecological environment evaluation indicator based on remote sensing data, can comprehensively reflect multiple ecological factors such as vegetation, humidity, surface temperature and land use, providing an important tool for ecological environment monitoring and evaluation.

[0003] However, the traditional RSEI evaluation calculation method has some limitations. First, the calculation process is complex and computationally intensive, requiring a large amount of data processing and computing resources. Second, the calculation of the RSEI evaluation relies on a large amount of prior knowledge and empirical parameters, which limits the generalization and robustness of the model. In addition, the RSEI evaluation has high requirements on the quality and timeliness of remote sensing data, is sensitive to the geographical location of the study area and the imaging time of remote sensing satellite data, and has poor reproducibility and interpretability of the algorithm. Summary of the invention

[0004] In view of the above-mentioned deficiencies in the prior art, the present invention provides a remote sensing ecological index evaluation method based on deep learning, proposes to improve the time series characteristics of the remote sensing ecological index in combination with a time series convolutional network model, optimizes the parallel processing capability of the remote sensing ecological index evaluation model for time series data, and linearly combines the comprehensive feature vector of the remote sensing ecological index factor predicted by the time series convolutional network model with the eigenvalue of the principal component analysis to obtain the time series convolutional neural network remote sensing ecological index value, so as to solve the limitations of the traditional remote sensing ecological index evaluation calculation.

[0005] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is: A remote sensing ecological index evaluation method based on deep learning includes the following steps: S1. Obtain Landsat remote sensing image data and land use data of the study area, preprocess the Landsat remote sensing image data, and extract remote sensing ecological index factors using the preprocessed Landsat remote sensing image data and land use data; S2, normalizing the remote sensing ecological index factor to obtain a normalized remote sensing ecological index factor; S3. Use principal component analysis to perform dimensionality reduction analysis on the normalized remote sensing ecological index factors to obtain the reduced dimensionality remote sensing ecological index factors and principal component analysis eigenvalues; S4, inputting the reduced dimension remote sensing ecological index factor into the time series convolutional network model, and generating a comprehensive feature vector of the remote sensing ecological index factor through causal convolution, dilated convolution, and global average pooling operations; S5. Linearly combine the comprehensive eigenvector of the remote sensing ecological index factor with the eigenvalue of the principal component analysis to construct a time series convolutional neural network remote sensing ecological index model, which is used to evaluate the ecological environment quality status of the study area.

[0006] The present invention has the following beneficial effects: 1. The remote sensing ecological index evaluation method based on deep learning proposed in this invention effectively improves the accuracy of ecological remote sensing index evaluation by making full use of the time series parallel processing capability of the temporal convolutional network model, thereby more accurately capturing the dynamic trend of ecological environment changes; 2. By combining the time series convolutional network model to evaluate ecological remote sensing indices, it is compatible with data of different resolutions and reduces the evaluation error caused by resolution differences through flexible estimation; 3. Taking advantage of the simple structure and intuitive decision-making process of the temporal convolutional network model, the interpretability of the model is significantly improved, making the evaluation results more transparent and credible; 4. Using the time series convolutional neural network remote sensing ecological index model to assess the current status of ecological quality can better adapt to new data and changing ecological environments, especially when the ecosystem is affected by engineering disturbances or other external factors, and can better assess the recovery or degradation trend of the ecosystem; 5. The use of convolutional neural network models to conduct parallel analysis of the temporal characteristics of ecological remote sensing indices can more accurately assess the quality of the ecological environment and more effectively carry out environmental monitoring and management, reducing environmental pollution and ecological damage; at the same time, it can also conduct in-depth environmental research under a limited budget, significantly reduce the cost of obtaining accurate environmental data, and greatly improve decision-making efficiency and market competitiveness. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] Figure 1 This is a flow chart of a remote sensing ecological index assessment method based on deep learning proposed in the present invention; Figure 2 It is the spatiotemporal distribution diagram of the green factor index of the Baihetan Hydropower Project area in the embodiment; Figure 3 It is the spatiotemporal distribution diagram of humidity factor index in Baihetan Hydropower Project area in the embodiment; Figure 4 It is the spatiotemporal distribution diagram of the dryness factor index in the Baihetan Hydropower Project area in the embodiment; Figure 5 It is the temporal and spatial distribution diagram of the heat factor index of the Baihetan Hydropower Project area in the embodiment; Figure 6 It is the spatiotemporal distribution map of the land use change factor index in the Baihetan Hydropower Project area in the embodiment; Figure 7 This is the spatiotemporal distribution map of the normalized water index factor indicators in the Baihetan Hydropower Project area in the example; Figure 8 It is the spatiotemporal distribution map of the Time Series Convolutional Neural Network Remote Sensing Ecological Index (TCN-RESI) of the Baihetan Hydropower Project Area in the embodiment. DETAILED DESCRIPTION

[0008] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0009] like Figure 1 As shown, a remote sensing ecological index evaluation method based on deep learning includes the following steps S1-S5: S1. Obtain Landsat remote sensing image data and land use data of the study area, preprocess the Landsat remote sensing image data, and use the preprocessed Landsat remote sensing image data and land use data to extract remote sensing ecological index factors.

[0010] In this embodiment, a 6-year data set with a time scale of 2014-2019, namely Landsat remote sensing image data and land use data, is obtained for calculating the remote sensing ecological index factor, so as to more scientifically reflect the ecological and environmental difference characteristics of the study area. In addition, the boundary of the study area of ​​the present invention is delineated based on terrain data, but the boundary of the study area can be comprehensively delineated based on factors such as terrain, administrative boundaries, and watershed scope in actual application; among them, the terrain data is directly downloaded using the ALOS PALSAR 12.5m elevation data product with a data resolution of 12.5m.

[0011] Specifically, step S1 specifically includes S11-S18: S11. Obtain Landsat remote sensing image data and land use data of the study area. Landsat remote sensing image data include Landsat-8 remote sensing data and Landsat5 TM data. Land use data include farmland, woodland, shrubs, grassland, water area, ice and snow, bare land, building land and wetland data.

[0012] In this embodiment, the land use data is the China Annual Land Cover Product (CLCD) derived from Landsat by Professor Huang Xin of Wuhan University, which is directly downloaded and used based on Google Earth Engine (GEE). The data has a total of 9 classification systems, including farmland, woodland, shrubs, grassland, water area, ice and snow, bare land, impervious surface and wetland. The time scale is from 2014 to 2019, and the data resolution is 30m; among them, the impervious surface is represented by various types of urbanized buildings (including houses, roads, bridges, parking lots, etc.), and the present invention changes the impervious surface data to building land to indicate the scope of human activities in the project area.

[0013] S12, performing data preprocessing including radiation correction, atmospheric correction, geometric correction and noise removal on the Landsat-8 remote sensing data and the Landsat5 TM data respectively, to obtain preprocessed Landsat-8 remote sensing data and preprocessed Landsat5 TM data.

[0014] S13. Calculate the greenness factor based on the preprocessed Landsat-8 remote sensing data, namely:

[0015] in, represents the greenness factor, , They respectively represent the near-infrared band reflectance data and the red band reflectance data of the preprocessed Landsat-8 remote sensing data.

[0016] In this embodiment, the purpose of calculating the greenness factor is to reflect the vegetation coverage status of the study area so as to evaluate the ecological environment quality status of the study area.

[0017] S14, based on the pre-processed Landsat-8 remote sensing data, the humidity factor is calculated by extracting the humidity component of each pixel, that is:

[0018] in, represents the humidity factor, , , , , , They represent the constant coefficients of the surface reflectance data corresponding to different bands. , , , , , They respectively represent the surface reflectance data of the blue band, green band, near-red band, mid-infrared band 1, and mid-infrared band 2 of the preprocessed Landsat-8 remote sensing data.

[0019] In this embodiment, the humidity factor is the humidity index (WET). Since the humidity factor can effectively reflect the moisture status in the environment, the change of moisture has a direct impact on ecological elements such as vegetation, soil, and water bodies. Especially after engineering disturbance, engineering activities often change the surface characteristics and destroy natural vegetation, thereby affecting soil moisture and overall water balance. Therefore, monitoring the change of humidity factor helps to judge the impact of engineering disturbance on ecology, as well as the progress and effect of regional ecosystem recovery after disturbance.

[0020] S15. Based on the preprocessed Landsat-8 remote sensing data, the dryness factor is constructed by calculating the bare soil index and the building index, that is:

[0021] in, represents the dryness factor, represents the building index, Represents the bare soil index.

[0022] In this embodiment, the dryness factor, namely the dryness index (Normalized Difference Built-up and Soil Index, referred to as NDBSI), is the degree of dryness in the environment and its impact on the ecosystem. It plays an important role in ecosystem health assessment, vegetation growth, water resource management, etc. And in arid or semi-arid areas, monitoring of the dryness factor is a key step in understanding the health of the ecosystem. After an engineering disturbance, especially projects involving large-scale mining, construction or land use changes, changes in the dryness factor can reflect whether the water balance of the region is affected. Therefore, in such an area, an increase in the dryness factor means that the ecosystem may be in a state of water scarcity and vegetation recovery is difficult. Therefore, the dryness factor is the key to evaluating the ecological restoration effect of the disturbed area.

[0023] S16. Based on the pre-processed Landsat5 TM data, the heat factor is calculated by extracting the surface brightness temperature, that is:

[0024] in, It represents the radiation value of the sixth band of the preprocessed Landsat5 TM data at the sensor. , Respectively represent the gain and offset values ​​of the band, Represents the gray value of the pixel, Represents the temperature value at the sensor, , Both represent the calibration parameters of the sensor. represents the heat factor, represents the logarithmic function, is the surface emissivity, represents the central wavelength of the thermal infrared band of band 6 of the preprocessed Landsat5 TM data, Represents a constant.

[0025] In this embodiment, the thermal factor (Land Surface Temperature, referred to as LST) is an important indicator for measuring environmental temperature and its impact on the ecosystem. It mainly reflects the effects and feedback of temperature changes on the ecosystem. Since temperature is one of the key environmental variables in the ecosystem, the thermal factor plays an important role in affecting plant growth, animal behavior, soil processes and the overall health of the ecosystem. Especially in projects involving large-scale mining or land transformation after engineering disturbances, the thermal factor can reflect changes in regional microclimate. The thermal balance of the region is destroyed after the disturbance, and the increase in the thermal factor may aggravate ecological pressure and affect plant recovery and animal habitat. Therefore, by monitoring the thermal factor, the progress of ecological restoration can be evaluated and guidance can be provided for restoration strategies.

[0026] S17. Based on the land use data, obtain the land use change factor.

[0027] In this embodiment, land use refers to the way humans develop, manage and use land resources according to their own needs; and there are many forms of land use, including agriculture, industry, commerce, residence, transportation, forestry, grassland and water, etc.; therefore, land use plays an important role in the ecosystem and environment. Therefore, land use data plays a vital role in reflecting the impact of engineering disturbances on the ecological environment, where engineering disturbances usually include infrastructure construction, mining, agricultural expansion or urbanization, etc. These activities will significantly change the way land is used and land cover. Therefore, through the monitoring and analysis of land use data, the impact of engineering disturbances on the ecosystem can be revealed, the environmental changes after the disturbance can be evaluated, and a scientific basis can be provided for ecological restoration planning. In this embodiment, land use data, i.e., land use change factors, include farmland, woodland, shrubs, grassland, water, ice and snow, bare land, building land and wetland data.

[0028] S18. Based on the preprocessed Landsat-8 remote sensing data, the normalized water index factor is calculated, namely:

[0029] in, represents the normalized water index factor, Represents the reflectance of the green band of the preprocessed Landsat-8 remote sensing data.

[0030] In this embodiment, the Modified Normalized Difference Water Index (MNDWI) is an ecological factor used to identify and monitor water distribution and its changes. It plays an important role in monitoring water and wetland environments, assessing ecosystem health, and studying the impact of climate change. Since water distribution and water resources often change significantly after engineering disturbances (such as infrastructure construction, mineral development, agricultural irrigation, etc.), the normalized water index can identify water reduction, drying up, or pollution caused by these engineering activities, helping ecological managers take timely restoration measures.

[0031] In addition, the above-mentioned greenness factor, humidity factor, dryness factor, heat factor, land use change factor and normalized water index factor are all remote sensing ecological index factors, which together constitute the remote sensing ecological index factor combination.

[0032] S2. Normalize the remote sensing ecological index factor to obtain a normalized remote sensing ecological index factor.

[0033] In this embodiment, the purpose of normalization is to scale the raw data of different ranges to a fixed range (usually between [0, 1]) to eliminate the dimension effect, so that different features have the same scale, accelerate the convergence speed of model training, improve the stability and accuracy of the model, and reduce data deviation.

[0034] Specifically, the calculation formula of the normalized remote sensing ecological index factor in step S2 is:

[0035] in, Represents the normalized value of each factor in the normalized remote sensing ecological index factor, It represents the value of each factor in the remote sensing ecological index factor at the pixel. It represents the minimum value of each factor in the remote sensing ecological index factor at the pixel. It represents the maximum value of each factor in the remote sensing ecological index factor at the pixel.

[0036] S3. Use principal component analysis to perform dimensionality reduction analysis on the normalized remote sensing ecological index factors to obtain the reduced dimensionality remote sensing ecological index factors and principal component analysis eigenvalues.

[0037] In this embodiment, the normalized remote sensing ecological index factor is reduced in dimension by principal component analysis (PCA), which facilitates model training in subsequent steps, thereby improving the efficiency of model training.

[0038] Specifically, step S3 specifically includes S31-S34: S31. According to the normalized remote sensing ecological index factor, a normalized remote sensing ecological index factor matrix is ​​constructed, and the covariance matrix is ​​calculated, that is:

[0039] in, represents the covariance matrix, represents the normalized remote sensing ecological index factor matrix, represents the number of rows of the normalized remote sensing ecological index factor matrix, Indicates transpose.

[0040] S32. Perform eigenvalue decomposition on the covariance matrix to obtain a number of principal component analysis eigenvalues ​​and corresponding eigenvectors.

[0041] S33, based on several principal component analysis eigenvalues, select the front of the covariance matrix The eigenvectors corresponding to the principal component analysis eigenvalues ​​form the transformation matrix .

[0042] in, The value of The cumulative variance contribution rate of the principal component analysis eigenvalues ​​exceeds the set threshold.

[0043] In this embodiment, the threshold is set to 85%, that is, the front The principal component analysis eigenvalues ​​were selected so that their cumulative variance contribution rate exceeded 85% and were regarded as valid information.

[0044] S34, multiplying the transformation matrix by the normalized remote sensing ecological index factor matrix to obtain the reduced dimension remote sensing ecological index factor, that is:

[0045] in, Represents the remote sensing ecological index factor of dimensionality reduction.

[0046] In this embodiment, the dimension-reduced remote sensing ecological index factor Each row represents a time step and each column represents a PCA eigenvalue.

[0047] S4. Input the reduced-dimensional remote sensing ecological index factor into the time series convolutional network model, and generate a comprehensive feature vector of the remote sensing ecological index factor through causal convolution, dilated convolution, and global average pooling operations.

[0048] Specifically, the number of convolutional layers of the temporal convolutional network model in step S4 is layer, the convolution kernel size is , the expansion factor is ; Among them, the first convolution layer is causal convolution and the subsequent convolution layers are dilated convolution.

[0049] In this embodiment, a time series convolutional network model (TCN) is used to capture the temporal characteristics of the data so that a remote sensing ecological index factor time series convolutional neural network evaluation model (TCN-RSEI) can be constructed in the subsequent steps to monitor the changes in ecological quality in the effective study area and predict the future ecological environment change trends in the study area.

[0050] In addition, since the size of the convolution kernel determines the window size of each operation on the time series, the convolution kernel size is generally Set to 2 or 3. In this embodiment, the specific value is 3. At the same time, since increasing the number of convolutional layers can expand the depth of the temporal convolutional network model, enabling it to capture more complex temporal dependencies, it is generally selected according to the requirements of data complexity and model depth. The expansion factor is 5-10 layers, and the specific value is 5 in this embodiment; secondly, because the expansion factor allows the temporal convolutional network model to expand the perception field without increasing the size of the convolution kernel, the model can capture a wider range of time dependencies with fewer layers and smaller convolution kernels, making the convolution operation sparse in time. The expansion factor is generally selected as At the same time, the first convolution layer of the temporal convolutional network model in this embodiment adopts causal convolution, the purpose of which is: causal convolution can ensure that Output It only depends on the current and previous inputs, not on future input data. The calculation formula is: ; Dilated convolution and non-linear activation are used in the second and subsequent convolutional layers. The purpose is: in the second and subsequent layers, dilated convolution is applied to enable the model to capture dependencies over a longer time range. The non-linear activation selects the ReLU activation function, which can be used to increase the expressive power of the model.

[0051] Specifically, step S4 includes S41-S44: S41. Input the dimension-reduced remote sensing ecological index factor into the first convolution layer of the temporal convolutional network model for causal convolution operation to obtain the output features of the first convolution layer, namely:

[0052] in, Indicates that the first convolutional layer is at time step The output features of represents the first convolutional layer convolution kernel weight matrix, represents the number of convolution kernels, Indicates that at time step Remote sensing ecological index factor for dimensionality reduction.

[0053] S42. Input the output features of the first convolutional layer into the second convolutional layer for extended convolution operation, and introduce the ReLU activation function for nonlinear activation to obtain the output features of the second convolutional layer.

[0054] S43, the output features of the second convolutional layer are input into the third convolutional layer for extended convolution operation and nonlinear activation, and the process is repeated until the last convolutional layer is input for extended convolution operation and nonlinear activation, and the predicted time series features are obtained, namely:

[0055] in, Indicates The convolutional layer is constructed at time step The output features of Indicates The convolutional layer convolution kernel weight matrix, Indicates The convolutional layer is constructed at time step The output features of .

[0056] In this example, the expansion factor determines the interval of the input data, and each time step of the predicted time series features contains Features.

[0057] S44. Perform global average pooling operation on the predicted time series features to generate a comprehensive feature vector of the remote sensing ecological index factor, namely:

[0058] in, Represents the comprehensive characteristic vector of remote sensing ecological index factors, Represents the total time step.

[0059] In this embodiment, Global Average Pooling (GAP) is a pooling operation used in a convolutional neural network (CNN). Its function is to take the average value of all elements of each feature map as the output of the feature map, thereby reducing the feature map to a scalar. Therefore, global average pooling is introduced in the temporal convolutional network model (TCN) to reduce the number of model parameters, avoid overfitting, and improve the generalization ability of the model.

[0060] S5. Linearly combine the comprehensive eigenvector of the remote sensing ecological index factor with the eigenvalue of the principal component analysis to construct a time series convolutional neural network remote sensing ecological index model, which is used to evaluate the ecological environment quality status of the study area.

[0061] Specifically, the temporal convolutional neural network remote sensing ecological index model in step S5 is:

[0062] in, represents the time series convolutional neural network remote sensing ecological index model, represents the number of remote sensing ecological index factors, Indicates Principal component analysis eigenvalues ​​of remote sensing ecological index factors.

[0063] In this embodiment, is the output result of the temporal convolutional network model, representing the output feature of the temporal convolutional network at the last layer. It is the final result after temporal convolution of the input data at different time steps. Each output contains the combined information of multiple temporal features. By linearly combining with the principal component analysis eigenvalues, we can get the time series convolutional neural network remote sensing ecological index value, namely the TCN-RSEI value. Among them, TCN-RSEI between 0-0.2 indicates poor ecological environment quality, 0.2-0.4 indicates poor ecological environment quality, 0.4-0.6 indicates general ecological quality, 0.6-0.8 indicates good ecological quality, and 0.8-1.0 indicates excellent ecological quality.

[0064] In order to verify the effectiveness of the remote sensing ecological index evaluation method based on deep learning proposed in this invention and reflect the impact of human engineering disturbance on the ecological environment, the Baihetan Hydropower Project on the main stream of Jinsha River was selected as an example for the study area. The preparation of the hydropower project began in 2010 and construction started in 2017. The scope of the study area was extracted based on the terrain data and the project impact range was delineated in combination with the literature data. Among them, the spatiotemporal distribution of the green factor index in the Baihetan Hydropower Project area is shown in the figure. Figure 2 As shown, Figure 2(a), (b), (c), (d), (e), and (f) are the spatiotemporal distribution maps of green factor indicators in 2014, 2015, 2016, 2017, 2018, and 2019 respectively; the spatiotemporal distribution of humidity factor indicators in the Baihetan Hydropower Project area is shown in Figure 3 As shown, Figure 3 (a), (b), (c), (d), (e), and (f) are the spatiotemporal distribution diagrams of humidity factor indicators in 2014, 2015, 2016, 2017, 2018, and 2019 respectively; the spatiotemporal distribution of dryness factor indicators in the Baihetan Hydropower Project area is shown in Figure 4 As shown, Figure 4 (a), (b), (c), (d), (e), and (f) are the spatiotemporal distribution diagrams of the dryness factor index in 2014, 2015, 2016, 2017, 2018, and 2019 respectively; the spatiotemporal distribution of the heat factor index in the Baihetan Hydropower Project area is shown in Figure 5 As shown, Figure 5 (a), (b), (c), (d), (e), and (f) are the temporal and spatial distribution maps of heat factor indicators in 2014, 2015, 2016, 2017, 2018, and 2019 respectively; the temporal and spatial distribution of land use change factor indicators in the Baihetan Hydropower Project area is shown in Figure 6 As shown, Figure 6 (a), (b), (c), (d), (e), and (f) are the spatiotemporal distribution maps of land use change factor indicators in 2014, 2015, 2016, 2017, 2018, and 2019, respectively. Since there is no ice and snow, bare land, or wetland in the land types of the Baihetan Hydropower Project, only six types of data, namely farmland, forest land, shrubs, grassland, water area, and construction land, are retained; The spatiotemporal distribution of the normalized water index factor indicators in the Baihetan Hydropower Project area is shown in Figure 2. Figure 7 As shown, Figure 7 (a), (b), (c), (d), (e), and (f) are the spatiotemporal distribution maps of the normalized water index factor indicators in 2014, 2015, 2016, 2017, 2018, and 2019 respectively; the spatiotemporal distribution of the Time Series Convolutional Neural Network Remote Sensing Ecological Index (TCN-RESI) in the Baihetan Hydropower Project area is shown in Figure 8 As shown in the figure, (a) and (b) are the spatiotemporal distribution maps of TCN-RESI from 2014 to 2016 and from 2017 to 2019, respectively; therefore, the Baihetan Hydropower Project was selected as an example for the experiment, which can reflect the impact of human engineering disturbances on the ecological environment and capture the dynamic trend of ecological environment changes.

[0065] In summary, the remote sensing ecological index evaluation method based on deep learning proposed in the present invention has the following advantages: (1) Improving evaluation accuracy: Through the time series parallel processing capability of the time series convolutional network model, the accuracy of the ecological remote sensing index (RSEI) evaluation is effectively improved, and the dynamic trend of ecological environment changes can be captured more accurately; (2) Optimizing spatial resolution adaptability: By combining the time series convolutional network model to evaluate the ecological remote sensing index (RSEI), it can be compatible with data of different resolutions. This flexibility is expected to reduce the evaluation error caused by resolution differences; (3) Improving model interpretability: The time series convolutional network model has a simple structure and an intuitive decision-making process, which will significantly improve the interpretability of the model and make the evaluation more accurate. The results are more transparent and credible; (4) Enhanced generalization ability: The TCN-RSEI model can better adapt to new data and changing environments, especially when the ecosystem is affected by engineering disturbances or other external factors, and can accurately predict the recovery or degradation trend of the ecosystem; (5) The convolutional neural network model is used to predict the time series characteristics of the ecological remote sensing index, and global average pooling operations and linear combinations are performed, which can more accurately and timely evaluate the ecological environment quality, help to more effectively carry out environmental monitoring and management, and reduce environmental pollution and ecological damage; at the same time, it can conduct in-depth environmental research under a limited budget, significantly reduce the cost of obtaining accurate environmental data, and greatly improve decision-making efficiency and market competitiveness.

[0066] The present invention uses specific embodiments to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present invention.

[0067] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.

Claims

1. A remote sensing ecological index evaluation method based on deep learning, characterized in that: The following steps are involved: S1. Obtain Landsat remote sensing image data and land use data of the study area, preprocess the Landsat remote sensing image data, and extract remote sensing ecological index factors using the preprocessed Landsat remote sensing image data and land use data; S2, normalizing the remote sensing ecological index factor to obtain a normalized remote sensing ecological index factor; S3. Use principal component analysis to perform dimensionality reduction analysis on the normalized remote sensing ecological index factors to obtain the reduced dimensionality remote sensing ecological index factors and principal component analysis eigenvalues; S4, inputting the reduced dimension remote sensing ecological index factor into the time series convolutional network model, and generating a comprehensive feature vector of the remote sensing ecological index factor through causal convolution, dilated convolution, and global average pooling operations; S5. Linearly combine the comprehensive eigenvector of the remote sensing ecological index factor with the eigenvalue of the principal component analysis to construct a time series convolutional neural network remote sensing ecological index model, which is used to evaluate the ecological environment quality status of the study area.

2. The remote sensing ecological index evaluation method based on deep learning according to claim 1 is characterized in that: Step S1 specifically includes: S11. Obtain Landsat remote sensing image data and land use data of the study area, where Landsat remote sensing image data include Landsat-8 remote sensing data and Landsat5 TM data, and land use data include farmland, woodland, shrubs, grassland, water area, ice and snow, bare land, building land and wetland data; S12, performing data preprocessing including radiation correction, atmospheric correction, geometric correction and noise removal on the Landsat-8 remote sensing data and the Landsat5 TM data respectively, to obtain preprocessed Landsat-8 remote sensing data and preprocessed Landsat5 TM data; S13. Calculate the greenness factor based on the preprocessed Landsat-8 remote sensing data, namely: in, represents the greenness factor, , They represent the near-infrared band reflectance data and the red band reflectance data of the preprocessed Landsat-8 remote sensing data respectively; S14, based on the pre-processed Landsat-8 remote sensing data, the humidity factor is calculated by extracting the humidity component of each pixel, that is: in, represents the humidity factor, , , , , , They represent the constant coefficients of the surface reflectance data corresponding to different bands. , , , , , They represent the surface reflectance data of the blue band, green band, near-red band, mid-infrared band 1, and mid-infrared band 2 of the preprocessed Landsat-8 remote sensing data respectively; S15. Based on the preprocessed Landsat-8 remote sensing data, the dryness factor is constructed by calculating the bare soil index and the building index, that is: in, represents the dryness factor, represents the building index, represents the bare soil index; S16. Based on the pre-processed Landsat5 TM data, the heat factor is calculated by extracting the surface brightness temperature, that is: in, It represents the radiation value of the sixth band of the preprocessed Landsat5 TM data at the sensor. , Respectively represent the gain and offset values ​​of the band, Represents the gray value of the pixel, Represents the temperature value at the sensor, , Both represent the calibration parameters of the sensor. represents the heat factor, represents the logarithmic function, is the surface emissivity, represents the central wavelength of the thermal infrared band of band 6 of the preprocessed Landsat5 TM data, represents a constant; S17. Based on the land use data, a land use change factor is obtained; S18. Based on the preprocessed Landsat-8 remote sensing data, the normalized water index factor is calculated, namely: in, represents the normalized water index factor, Represents the reflectance of the green band of the preprocessed Landsat-8 remote sensing data.

3. The remote sensing ecological index evaluation method based on deep learning according to claim 2 is characterized in that: The calculation formula of the normalized remote sensing ecological index factor in step S2 is: in, Represents the normalized value of each factor in the normalized remote sensing ecological index factor, It represents the value of each factor in the remote sensing ecological index factor at the pixel. It represents the minimum value of each factor in the remote sensing ecological index factor at the pixel. It represents the maximum value of each factor in the remote sensing ecological index factor at the pixel.

4. The remote sensing ecological index evaluation method based on deep learning according to claim 3 is characterized in that: Step S3 specifically includes: S31. According to the normalized remote sensing ecological index factor, a normalized remote sensing ecological index factor matrix is ​​constructed, and the covariance matrix is ​​calculated, that is: in, represents the covariance matrix, represents the normalized remote sensing ecological index factor matrix, represents the number of rows of the normalized remote sensing ecological index factor matrix, represents transpose; S32, performing eigenvalue decomposition on the covariance matrix to obtain a number of principal component analysis eigenvalues ​​and corresponding eigenvectors; S33, based on several principal component analysis eigenvalues, select the front of the covariance matrix The eigenvectors corresponding to the principal component analysis eigenvalues ​​form the transformation matrix ; in, The value of The cumulative variance contribution rate of the principal component analysis eigenvalues ​​exceeds the set threshold for determination; S34, multiplying the transformation matrix by the normalized remote sensing ecological index factor matrix to obtain the reduced dimension remote sensing ecological index factor, that is: in, Represents the remote sensing ecological index factor of dimensionality reduction.

5. The remote sensing ecological index evaluation method based on deep learning according to claim 4 is characterized in that: The number of convolutional layers of the temporal convolutional network model in step S4 is layer, the convolution kernel size is , the expansion factor is ; Among them, the first convolution layer is causal convolution and the subsequent convolution layers are dilated convolution.

6. The remote sensing ecological index evaluation method based on deep learning according to claim 5 is characterized in that: Step S4 specifically includes: S41, inputting the dimension-reduced remote sensing ecological index factor into the first convolution layer of the temporal convolutional network model to perform a causal convolution operation, and obtaining the output features of the first convolution layer; S42, inputting the output features of the first convolutional layer into the second convolutional layer for extended convolution operation, and introducing the ReLU activation function for nonlinear activation to obtain the output features of the second convolutional layer; S43, inputting the output features of the second convolution layer into the third convolution layer for dilated convolution operation and nonlinear activation, iterating in sequence until the output features of the last convolution layer are input for dilated convolution operation and nonlinear activation, thereby obtaining predicted time series features; S44. Perform a global average pooling operation on the predicted time series features to generate a comprehensive feature vector of the remote sensing ecological index factor.

7. The remote sensing ecological index evaluation method based on deep learning according to claim 6 is characterized in that: The calculation formula of the output feature of the first convolutional layer in step S41 is: in, Indicates that the first convolutional layer is at time step The output features of represents the first convolutional layer convolution kernel weight matrix, represents the number of convolution kernels, Indicates that at time step Remote sensing ecological index factor for dimensionality reduction.

8. The remote sensing ecological index assessment method based on deep learning according to claim 7 is characterized in that: The calculation formula of the predicted time series feature in step S43 is: in, Indicates The convolutional layer is constructed at time step The output features of Indicates The convolutional layer convolution kernel weight matrix, Indicates The convolutional layer is constructed at time step The output features of .

9. The remote sensing ecological index assessment method based on deep learning according to claim 8, characterized in that: The calculation formula of the comprehensive characteristic vector of the remote sensing ecological index factor in step S44 is: in, Represents the comprehensive characteristic vector of remote sensing ecological index factors, Represents the total time step.

10. The remote sensing ecological index assessment method based on deep learning according to claim 9, characterized in that: The temporal convolutional neural network remote sensing ecological index model in step S5 is: in, represents the time series convolutional neural network remote sensing ecological index model, represents the number of remote sensing ecological index factors, Indicates Principal component analysis eigenvalues ​​of remote sensing ecological index factors.

Images