Artificial intelligence-based method for detecting changes in ecological protection red line areas

Abstract

The present invention discloses an artificial intelligence-based change detection method for ecological protection redline areas, which relates to the field of artificial intelligence technology and includes the following steps: Step 1: Converting remote sensing data at each time point into directly comparable surface reflectance; Step 2: Using a pixel decomposition method, splitting the mixed reflectance spectrum in the remote sensing data into vegetation reflectance spectrum data and soil reflectance spectrum data; Step 3: Establishing a deep learning model that uses the difference between the vegetation reflectance spectrum data and the soil reflectance spectrum data of the dual-temporal data as input to automatically extract the change characteristics of the ecological protection area and output the change probability; Step 4: Setting an adaptive change detection threshold based on the sensitivity of the ecological protection redline area; and determining the detection result based on the change probability and the adaptive change detection threshold. The present invention effectively overcomes the problems of large mixed pixel errors and poor threshold fixity in traditional change detection.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an artificial intelligence-based method for detecting changes in ecological protection red line areas. Background Art

[0002] Ecological redline areas often possess extremely high ecological value, providing core ecological services such as water conservation, biodiversity protection, wind and sand control, and maintaining regional climate balance. Therefore, timely and accurate monitoring of their dynamic changes is crucial for ensuring ecological security. Remote sensing technology, with its large-scale, periodic, non-contact, and high spatial resolution, has been widely used for dynamic monitoring of ecological redline areas. Numerous studies and technical approaches rely on remote sensing imagery to identify ecological changes, such as the Normalized Difference Vegetation Index (NDVI), the Enhanced Vegetation Index (EVI), Change Vector Analysis (CVA), and image difference analysis. These traditional remote sensing analysis methods primarily identify surface changes based on spectral differences between remote sensing images at two time points. However, these methods still present significant challenges in monitoring ecological changes in redline areas.

[0003] First, existing technologies fail to adequately handle mixed pixels, leading to large errors in change identification. Redline areas are often located in complex terrain and landforms such as mountains, hills, and wetlands. Pixels in remote sensing images typically represent a mixed response of various landforms, including vegetation, bare soil, and water bodies. A single vegetation index or spectral difference cannot accurately distinguish subtle ecological disturbances from changes in the natural background, and can easily produce "false changes" or "missed detections." While some studies have introduced spectral mixing analysis (such as Linear Spectral Unmixing) to deconstruct pixels, these models are often limited to experimental verification under ideal conditions and lack the ability to model dynamic changes in vegetation cover in real time, making them difficult to adapt to the highly heterogeneous nature of ecosystems.

[0004] Secondly, fixed thresholds or manual experience-based judgment criteria limit the accuracy and robustness of change identification. Traditional change detection methods usually set a unified change threshold for spectral changes or vegetation index changes. Once crossed, a change is considered to have occurred. This method ignores the natural volatility, seasonal periodicity, and local outlier disturbances within the ecosystem. For example, in the forest edge zone of the red line area, non-structural changes such as leaf color change and light obstruction may occur with the change of seasons. These changes may show significant fluctuations in NDVI and can easily be mistakenly identified as ecological degradation; while mild disturbances in the early stages of land reclamation may not be captured due to the small amplitude of cover changes. Current studies have attempted to introduce dynamic thresholds or fuzzy logic rules to improve adaptability, but there are still limitations in spatial scale, temporal consistency, and automated model deployment.

[0005] Third, the lack of modeling of the spatial continuity of ecosystem change results in change identification results that are not ecologically interpretable. Ecosystem changes often have obvious spatial expansion characteristics. For example, forest degradation, water body shrinkage, and grassland fragmentation are not isolated events, but rather manifest as regional patch shifts or boundary movements. However, existing remote sensing change detection methods generally use pixel-level binary judgments, which do not fully incorporate spatial neighborhood information. They determine the change state based solely on single-point spectral changes, ignoring the overall consistency of the neighborhood structure. This can easily lead to spotty detection maps or discontinuous boundary identification, which in turn affects policymakers' identification and management responses to ecological risk areas. Summary of the Invention

[0006] To address these technical issues, we propose an AI-based change detection method for ecological redline areas, enabling high-precision intelligent identification of small yet critical ecological changes within these areas. This method effectively overcomes the challenges of traditional change detection, such as large mixed pixel errors, poor threshold stability, missing spatial information, and a lack of ecological explanatory power. This significantly improves the sensitivity, stability, and ecological interpretability of change detection.

[0007] In order to achieve the above objects, the technical solution adopted by the present invention is:

[0008] The artificial intelligence-based method for detecting changes in ecological protection red line areas includes:

[0009] Step 1: Perform atmospheric correction and radiation correction on the remote sensing data at each time to convert the remote sensing data at each time into directly comparable surface reflectance;

[0010] Step 2: Split the mixed reflectance spectrum in the remote sensing data into vegetation reflectance spectrum data and soil reflectance spectrum data by pixel decomposition method; the vegetation reflectance spectrum data and soil reflectance spectrum data are both bi-temporal data;

[0011] Step 3: Build a deep learning model that uses the difference between vegetation reflectance spectral data and soil reflectance spectral data from dual-temporal data as input, extracts the change characteristics of the ecological protection area, and outputs the change probability;

[0012] Step 4: Set an adaptive change detection threshold based on the sensitivity of the ecological protection redline area; determine the detection result based on the change probability and the adaptive change detection threshold.

[0013] Furthermore, in step 1, at time The surface reflectivity at :

[0014]

[0015] in: Indicates time Lower pixel Surface reflectivity; Indicates time The pixels received The radiance of Indicates the distance between the Earth and the Sun; Indicates time Solar irradiance under Indicates time The solar zenith angle below is the horizontal coordinate of the pixel; is the vertical coordinate of the pixel.

[0016] Furthermore, the radiation brightness is obtained through the following process: a radiation detection instrument is installed on the remote sensing sensor; the radiation detection instrument includes a multispectral sensor and an imaging spectrometer; when the remote sensing sensor passes over the ground, the sunlight reflected from the surface propagates through the atmosphere to reach the photosensitive element of the remote sensing sensor, and the remote sensing sensor records the electromagnetic radiation energy intensity received in different bands and stores it in the form of digital grayscale values; the digital grayscale values recorded by the remote sensing sensor are converted into radiation brightness with physical meaning through radiation calibration processing.

[0017] Furthermore, the process of converting the digital grayscale values recorded by the remote sensing sensor into radiant brightness with physical meaning through radiometric calibration includes: obtaining digital grayscale calibration parameters, including gain and bias, and converting through the radiometric brightness formula to obtain the radiant brightness value of each pixel in the remote sensing data in a specific band. The radiant brightness formula is: radiant brightness = gain * digital grayscale value + bias, and the unit of radiant brightness is W·m⁻²·sr⁻¹·μm⁻¹.

[0018] Furthermore, in step 2, the mixed reflectance spectrum in the remote sensing data is split into vegetation reflectance spectrum data and soil reflectance spectrum data by pixel decomposition method as follows:

[0019]

[0020] in: For time Lower pixel The proportion of vegetation cover area; For time Lower pixel Vegetation reflectance spectral data; For time Lower pixel Soil reflectance spectral data.

[0021] Furthermore, in step 3, the output change probability is:

[0022]

[0023] in: Indicates time Lower pixel The probability of change; Represents the sigmoid activation function; Indicates the number of layers of the deep learning network, the value is greater than or equal to 6; Indicates the The weight of the layer; Indicates the Non-linear activation function of the layer; and Indicates two different times; Indicates time Lower pixel Surface reflectivity; Indicates time Lower pixel Surface reflectivity; For time Lower pixel The proportion of vegetation cover area; For time Lower pixel The proportion of vegetation cover area; For time Lower pixel Vegetation reflectance spectral data; For time Lower pixel Soil reflectance spectral data; For time Lower pixel Soil reflectance spectral data.

[0024] Further, ;when When greater than or equal to 1 and less than 3, is a linear rectification function; when When greater than or equal to 3 and less than 5, is an exponential linear function; when Greater than or equal to 5 and less than hour, is the hyperbolic tangent function.

[0025] Furthermore, in step 4, the adaptive change detection threshold is set for:

[0026]

[0027] in: is the detection reference value, and its value range is a real number greater than 0 and less than 1. The larger it is, the more sensitive the ecological protection red line area is.

[0028] Furthermore, in step 4, based on the change probability and the adaptive change detection threshold, the detection result is determined as:

[0029]

[0030] in: Expressed as time Lower pixel The test result is 1, which means change and 0 means no change. Represents pixels The spatial neighborhood of Represents an indicator function, which is 1 when the condition is met and 0 otherwise; Indicates the number of pixels in the neighborhood; Indicates the neighborhood consistency threshold, ranging from 0 to 1; and Represents pixels The horizontal and vertical coordinates of other pixels in the spatial neighborhood of ; Indicates time Lower pixel The probability of change; Indicates time Lower pixel Adaptive change detection threshold.

[0031] Compared with the existing technology, the beneficial effect of the present invention is that it realizes high-precision identification and dynamic monitoring of ecological disturbances in ecological protection red line areas. Compared with the existing technology that only relies on vegetation index changes or fixed thresholds, the present invention has significant beneficial effects: first, through physical radiation brightness inversion and standardized reflectance calculation, the consistency and comparability of multi-phase remote sensing data are ensured, laying a solid foundation for subsequent modeling; second, by introducing the vegetation-soil decomposition model, the ecological information in the mixed pixels is effectively stripped away, and the sensitivity of the model to weak changes is enhanced; third, the deep neural network constructed by combining multi-layer activation functions and band-sensitive weights can automatically extract change features and generate change probability outputs with physical interpretations; finally, through the exponential adaptive threshold and spatial neighborhood consistency judgment mechanism, the stability and ecological interpretability of the detection results are significantly improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 This is a schematic diagram of the method flow of the artificial intelligence-based ecological protection red line area change detection method proposed in the present invention. DETAILED DESCRIPTION

[0033] The following description is intended to disclose the present invention so that those skilled in the art can implement the present invention. The preferred embodiments described below are merely examples, and those skilled in the art may conceive of other obvious variations.

[0034] Reference Figure 1 As shown, the artificial intelligence-based ecological protection red line area change detection method in the embodiment of the present invention includes:

[0035] Step 1: Perform atmospheric correction and radiation correction on the remote sensing data at each time to convert the remote sensing data at each time into directly comparable surface reflectance;

[0036] Step 1 is the foundation of the entire methodology, with the core goal of converting remote sensing observation data acquired at different time phases into surface reflectance that can be compared across time, thereby providing highly consistent and physically realistic input data for subsequent change feature extraction and deep learning model analysis. During the original acquisition of remote sensing images, due to complex changes in observation conditions, solar irradiance, atmospheric composition, and sensor response, the reflectance information of the same object at different times and under different climatic conditions exhibits significant differences. Therefore, precise atmospheric correction and radiation calibration must be performed to remove the influence of these non-object attributes and restore their true surface radiation characteristics, so that the artificial intelligence model can scientifically and reliably learn and judge changes in ecological protection red line areas.

[0037] In this method, surface reflectance is acquired from raw images recorded by remote sensing sensors. These images are presented as digital grayscale values, with each pixel's grayscale value corresponding to the radiant energy received by that pixel within a specific wavelength band. However, because grayscale values themselves lack direct physical meaning, they must first be converted into radiant brightness, which has practical energy significance, through a radiometric calibration process. This process relies on the calibration parameters of the remote sensing sensor, including gain and bias, and achieves the conversion between grayscale values and radiant brightness through linear mapping. This physical modeling process ensures the physical authenticity of the data and provides a reliable foundation for subsequent reflectance calculations.

[0038] Furthermore, considering that the surface radiation signal is affected by multiple scattering and absorption by atmospheric gases, aerosols, and water vapor during transmission, the present invention introduces an atmospheric correction mechanism, using key parameters such as the solar zenith angle, the Earth-Sun distance, and the solar irradiance in the observation band for comprehensive correction. Through the derived reflectivity expression, a quantitative relationship is established between the original radiation brightness and the observation geometry and solar energy distribution parameters, thereby obtaining a standardized surface reflectivity. This reflectivity has been stripped of the interference caused by the observation time and climatic conditions, so that remote sensing data from different phases can be regarded by the artificial intelligence model as physical data under the same evaluation benchmark, making change detection possible.

[0039] It is worth emphasizing that the surface reflectivity in the present invention is not just an intermediate product of data preprocessing, but plays a key role as an information carrier in the entire ecological change detection. During the long-term monitoring of the ecological protection red line area, due to the interweaving of multiple factors such as vegetation growth cycle, climate evolution and human intervention, the surface reflectivity characteristics in the area will undergo subtle and complex changes. Without precise atmospheric correction and radiation standardization steps, these changes will be masked by observation errors, and it will be difficult for artificial intelligence models to learn effective change patterns from them. The reflectivity conversion method proposed in the present invention not only retains the true spectral characteristics of the ground objects, but also effectively reduces the interference of temporal noise on model training, thereby improving the sensitivity and robustness of change detection.

[0040] Step 2: Split the mixed reflectance spectrum in the remote sensing data into vegetation reflectance spectrum data and soil reflectance spectrum data by pixel decomposition method; the vegetation reflectance spectrum data and soil reflectance spectrum data are both bi-temporal data;

[0041] The core of step 2 is to invert and decompose the mixed spectrum of each pixel in the remote sensing image, thereby converting the observed data into two components with ecological and physical significance: vegetation reflectance and soil reflectance. This decomposition is not only a key step in data preprocessing, but also a structural foundation that supports the subsequent deep learning model to achieve accurate change identification in ecological protection red line areas. Due to the limitations of the spatial resolution of remote sensing imaging, the ground object corresponding to a pixel is often a mixture of multiple ground object types, such as bare soil areas partially covered with vegetation, desert areas scattered with shrubs, or buffer zones where weeds and buildings coexist. In this case, the original surface reflectance value cannot reflect the true changes of a single ecological factor. Therefore, it is necessary to use mixed pixel decomposition technology to reconstruct the remote sensing observation value into a structural expression composed of a weighted combination of basic components.

[0042] On this basis, the present invention adopts a linear mixed model with physical interpretability to decompose the overall reflectance of each pixel at any time into a linear superposition result composed of a certain proportion of vegetation reflectance and soil reflectance. This model assumes that at medium-scale resolution, the spectral response of vegetation and soil as the dominant ecological elements can represent the two basic components of green cover and non-green background in the pixel respectively. By introducing the vegetation cover ratio as a weighting factor, each reflectance value not only carries spectral information, but also integrates the structural information of the distribution of land objects. The introduction of this method not only improves the discrimination of internal changes in ecological elements, but also establishes a mapping relationship between ecological structure and spectral response at the data level, so that subsequent models can rely on more structurally discriminating feature data when learning and judging changes in ecological red line areas.

[0043] The essence of this step is the decoupling process of refining macro remote sensing signals into micro ecological indicators, which not only solves the inherent problems of “fuzziness” and “mixing” of remote sensing pixels, but also provides a more sparse and clear input channel for deep learning networks. Especially in red line areas, such areas are often the most sensitive and variable spatial belts of ecosystem changes. Their vegetation coverage and soil exposure fluctuate rapidly with seasonality, climate disturbances or human development. Therefore, if there is no accurate pixel decomposition process, it is very easy to produce “false changes” or “false alarms” in the model, thereby weakening the intelligent system’s ability to respond to real changes. Through the pixel spectral deconstruction mechanism adopted by the present invention, the model can focus on two clear change indicators during the learning process - the dual-phase difference of vegetation reflectance and soil reflectance - thereby effectively avoiding the interference of complex ground background confusion on the classification boundary and improving the credibility of dynamic monitoring of ecological red line areas.

[0044] In addition, the decomposition results output by this step also have a high degree of physical consistency, which enables the subsequent neural network to converge more easily during the training phase and better fit the nonlinear evolution process of the ecosystem. Since the deconstruction model adopted by the present invention takes into account the coverage ratio of the spatial structure and the intensity characteristics of the spectral response in the modeling, its output not only serves the change detection itself, but also provides a data basis for the temporal evolution analysis of the ecological state, vegetation degradation assessment and the derivation of the soil drought index. This feature reflects the versatility and extensibility of this method in ecological remote sensing intelligent analysis. In summary, step 2 is not only a technical processing process, but also a key bridge for converting raw data into structural ecological variables and building a more expressive feature space for artificial intelligence models.

[0045] Step 3: Build a deep learning model that uses the difference between vegetation reflectance spectral data and soil reflectance spectral data from dual-temporal data as input, extracts the change characteristics of the ecological protection area, and outputs the change probability;

[0046] Step 3 is the core of the entire technical system. Its fundamental task is to automatically extract the change characteristics of the ecological protection red line area based on the differences between the dual-phase vegetation and soil reflectance spectral data obtained in the first two steps by building a deep learning model, and output the change probability at the pixel level. Compared with traditional change detection methods that rely solely on static threshold judgments based on single-band differences or exponential combinations, this invention fully incorporates the temporal perception and nonlinear feature learning capabilities of artificial intelligence, allowing the system to identify weak but critical ecological dynamics in complex surface change scenarios. In particular, in the highly sensitive area of the ecological protection red line, its ability to identify small-amplitude interference and high-frequency disturbances is significantly better than traditional methods, reflecting the breakthrough value of intelligent monitoring.

[0047] The design of the deep learning model in this invention is closely built around the physical nature of ecological change. The input end collects the spectral differences of vegetation and soil after pixel decomposition between the two time phases. These differences not only contain information about the changes in the reflectance characteristics of the ground objects, but also imply the evolution trend of the surface cover structure. Through multiple layers of nonlinear transformation, the model maps these original spectral differences into a feature space with higher discrimination, so that the ecological disturbance presents a more stable and easy-to-identify characteristic form in this space. Especially in the deep structure, the model introduces a combination of multiple types of activation functions, such as linear rectification, exponential linear function and hyperbolic tangent function, to simulate the response of ecosystems to changes of different scales and intensities, thereby improving the ability to identify weak changes and complex changes. Corresponding to this is the imbalance of the weight design of each layer. Through the hierarchical attenuation mechanism, the dominant weight of high-level abstract features in the final judgment is gradually enhanced, so that the model pays more attention to the deep structural differences hidden in the temporal evolution, rather than just the drastic fluctuations of the surface spectrum.

[0048] The output of the change probability in this invention is not simply a linear function of a significant difference, but rather a nonlinear output derived from the comprehensive judgment of the input information by the entire deep neural network. This approach makes the model tolerant to noise perturbations and enhances sensitivity to local dynamics. For example, while some areas may experience minimal changes in surface reflectivity, their vegetation cover may fluctuate slightly, and soil moisture may undergo subtle adjustments. This ecological state evolution is a typical example of redline sensitivity, easily overlooked by traditional methods. However, the deep model constructed in this invention explicitly expresses this as a high-probability output through multi-layer feature combination and weighting, reflecting the intelligent system's high sensitivity to changes in ecological safety boundaries.

[0049] Furthermore, the model not only captures local features but also incorporates spatial similarity modeling capabilities through a convolution kernel weight and activation function adjustment mechanism. This means that similarities or differences in reflectance structure between pixels also influence the probability of change in the current pixel. This mechanism effectively mitigates the "island effect" and "blurred boundaries" in remote sensing data, improving the coherence and spatial consistency of detection results. Furthermore, the change probability output by the model is not a binary result, but a continuous value range index of change trend, providing a mathematical foundation and semantic interpretation for subsequent adaptive threshold judgment and spatial neighborhood constraints.

[0050] Step 4: Set an adaptive change detection threshold based on the sensitivity of the ecological protection redline area; determine the detection result based on the change probability and the adaptive change detection threshold.

[0051] Step 4 bears the key responsibility from the model output results to the final change identification decision. Its core lies in designing a set of adaptive change detection threshold mechanisms that are highly compatible with ecological sensitivity, and combining them with the spatial neighborhood consistency strategy to accurately determine the change probability of each pixel. This process is not only an effective explanation and physical constraint for the output of the deep learning model, but also an important guarantee for achieving high-accuracy and high-stability identification in the ecological red line protection system. Since the ecological protection red line areas often have strong spatial heterogeneity and time sensitivity, the fixed threshold method is often difficult to adapt to the complex evolutionary rhythm of different ecosystems. The present invention introduces an exponential threshold function that is dynamically coupled with the spectral characteristics of vegetation to achieve intelligent adjustment capabilities that are adapted to local conditions in spatial distribution and to time changes, so that the system has stronger ecological adaptability and intelligent judgment capabilities.

[0052] In the specific implementation process, the present invention compares the probability of change of the model output with the adaptive threshold to determine whether an ecological change has occurred in a certain pixel. This threshold is not a fixed value set by humans, but is dynamically generated according to the amplitude of the change in vegetation reflectance between the two time phases of the pixel. When the vegetation reflectance in a certain area changes dramatically, it may mean a real ecological disturbance. The system will automatically lower the judgment threshold and enhance sensitivity. Conversely, when the change is weak, the threshold will be increased accordingly to avoid the interference of accidental noise on the change judgment. This exponential adjustment mechanism enables the system to have an amplification mechanism for weak changes at the red line boundary, and exhibits fault tolerance characteristics for slight fluctuations in non-critical areas, effectively avoiding the dual risks of false alarms and missed judgments. Especially in some areas with complex landforms and frequent ecological fluctuations but not caused by human intervention, this mechanism can fully release the change perception of the artificial intelligence model while maintaining the physical rationality and ecological interpretability of the overall judgment of the system.

[0053] More importantly, the present invention does not limit the judgment of changes to the single pixel dimension, but further introduces a spatial neighborhood consistency judgment strategy, and uses the change probability of pixels in the neighborhood to compare with the threshold value to perform integrated judgment. This strategy stems from the spatial continuity characteristics of ecological changes, that is, real ecological disturbances are often not limited to a certain isolated pixel, but expand on the regional scale in the form of groups or patches. By constructing a local neighborhood system centered on the current pixel, and calculating and judging the proportion of changed pixels in the neighborhood, the present invention effectively eliminates the "noise change" phenomenon caused by isolated pixel errors, and also improves the spatial coherence and ecological structure explanatory power of the change results. The introduction of the neighborhood consistency threshold not only enhances the stability of the judgment, but also gives the system a certain degree of "group judgment ability", which is difficult to match with the traditional change detection method based on single-pixel logical judgment.

[0054] In addition, the change judgment logic itself adopts a strict combination of logical conditions in mathematical form, which makes the final output result replicable and highly controllable. Change judgment not only requires that the probability of change of the pixel is higher than the adaptive threshold, but also requires that the proportion of pixels in the neighborhood that meet the same conditions exceeds a certain set threshold. This "probability + structure" judgment mechanism reflects the comprehensive mobilization of the multi-scale structural perception ability of the ecosystem driven by artificial intelligence in the present invention, ensuring that the change judgment is not only accurate, but also has ecological coherence. This spatial collaborative judgment method is particularly suitable for special ecological units such as ecological protection red line areas with clear boundaries, clear functions, and significant consequences of changes, so that the system can still maintain stability and discrimination capabilities in dynamic changes.

[0055] Furthermore, in step 1, at time The surface reflectivity at :

[0056]

[0057] in: Indicates time Lower pixel Surface reflectivity; Indicates time The pixels received The radiance of Indicates the distance between the Earth and the Sun; Indicates time Solar irradiance under Indicates time The solar zenith angle below is the horizontal coordinate of the pixel; is the vertical coordinate of the pixel.

[0058] The reflectivity calculation model used is a physical derivation model with radiance as the core input variable. Its goal is to convert the amount of electromagnetic radiation received by the remote sensor in the sky (i.e. radiance) into ) is converted into the actual surface reflectivity of the corresponding object at the time of observation . This conversion process is based on the accurate modeling of solar irradiance, solar zenith angle, and the distance between the earth and the sun. Specifically, the radiation energy reflected by the ground object after being illuminated by the sun first propagates through the atmosphere, reaches the remote sensor lens and is recorded by the detector. This process is affected by the coupling of multiple geometric and radiation parameters. In order to eliminate these external interference factors and restore the reflection ability of the ground object itself in a specific band, it is necessary to introduce a set of inversion mechanisms to normalize the received brightness value with the solar radiation intensity and illumination angle.

[0059] in, This is a factor introduced by the Lambertian reflectance model commonly used in remote sensing systems to average the angular distribution of radiation energy, ensuring energy conservation during the calculation process. The square term is used to correct the energy density change caused by the length of the radiation path between the sun and the earth. Considering that the earth's orbit around the sun is elliptical and its distance varies periodically, if it is not corrected, it will cause errors in the solar energy input in winter and summer, thereby affecting the authenticity of the seasonal variation characteristics of vegetation. This is particularly critical in ecological red line areas, because the small periodic characteristics of vegetation changes are often early signals of ecological disturbances.

[0060] and , that is, the sun at time The irradiance under is the solar energy input per unit area defined in the band. It, together with the reflective characteristics of the ground object, determines the final remote sensing received brightness. By normalizing the radiance to unit incident energy, the inconsistency of solar input energy is eliminated and the comparability between time series is enhanced. The term is used to correct the projection effect caused by the solar zenith angle. The larger the solar zenith angle, the closer the sun is to the horizon, the longer its illumination path in the atmosphere, and the smaller the energy per unit area actually projected onto the ground; when the zenith angle is close to zero, that is, the sun is close to being directly overhead, the illumination intensity is the greatest. Dividing the brightness by this cosine value is actually restoring the amount of reflection from the obliquely illuminated surface to the standard reflectivity under direct sunlight conditions. This design allows data acquired at different times and different observation angles to be physically unified, ensuring that the artificial intelligence system can process highly consistent data during the training and prediction stages.

[0061] In conjunction with the goal of the present invention, which is to automatically identify changes in ecological protection red line areas with high precision, the use of the above-mentioned reflectivity calculation model has extremely high practical adaptability and engineering value. On the one hand, it provides physical attribute values of land objects that can be directly compared across time, making dual-phase difference analysis possible, which is a prerequisite for subsequent pixel decomposition and change feature extraction. On the other hand, the calculation process fully suppresses non-land object interference factors in ecological changes, and enhances the model's sensitivity to real ecological disturbances. For example, on the same plot, if the remote sensing image has significant differences in solar zenith angle due to different imaging times, if geometric normalization is not performed, the reflectivity differences in the same vegetation area may be entirely caused by the observation conditions, causing the model to misidentify it as an ecological change; and by introducing and By using correction factors such as ΔE and ΔS, this error can be eliminated, thereby improving the authenticity of change detection and the ecological interpretability of the judgment results.

[0062] Furthermore, the radiation brightness is obtained through the following process: a radiation detection instrument is installed on the remote sensing sensor; the radiation detection instrument includes a multispectral sensor and an imaging spectrometer; when the remote sensing sensor passes over the ground, the sunlight reflected from the surface propagates through the atmosphere to reach the photosensitive element of the remote sensing sensor, and the remote sensing sensor records the electromagnetic radiation energy intensity received in different bands and stores it in the form of digital grayscale values; the digital grayscale values recorded by the remote sensing sensor are converted into radiation brightness with physical meaning through radiation calibration processing.

[0063] In practice, the present invention utilizes a high-precision radiation detection device, including a multispectral sensor and an imaging spectrometer, mounted on a remote sensing platform to enable band-specific observation of surface reflected light in multiple specific wavelengths. These detectors separate the reflected solar radiation signal from the surface into electromagnetic energy flows in different wavelength bands, transmitting them as electrical signals to photosensitive elements for recording. As the remote sensing platform passes over the surface, sunlight reflected from ground objects does not travel directly to the sensor, but instead must pass through the atmosphere, a complex radiation transmission medium. Aerosols, water vapor, ozone, and various gaseous components in the atmosphere absorb, scatter, and reflect radiation in different wavelength bands, causing the electromagnetic signal that ultimately reaches the sensor to vary in both intensity and spectrum. Therefore, the signal received by the remote sensing photosensitive element is not entirely identical to the original reflective properties of the ground object, but rather a composite signal modulated by atmospheric path modulation. For artificial intelligence models to learn the true characteristics of ecological state changes, it is necessary to link the raw remote sensing data with the actual reflective behavior of the surface through the physical quantity of radiance.

[0064] Remote sensing images recorded by sensors are presented as pixel grayscale values. These grayscale values are digital representations of electronic signals after analog-to-digital conversion, and their numerical range does not directly correspond to actual radiant energy. Therefore, the present invention introduces a radiometric calibration mechanism to convert the grayscale values of each band into physically meaningful radiant brightness values. This conversion process relies on pre-acquired calibration parameters, typically including the gain and offset for each band, which together define the linear mapping between grayscale values and true brightness. Specifically, the system uses calibration experiments to determine the response curve of each band under different grayscale inputs. This is then used to construct a grayscale-to-brightness conversion function, resulting in radiant brightness in units of W·m⁻²·sr⁻¹·μm⁻¹, which has strict physical dimensions and is therefore comparable. Through this conversion process, the present invention ensures the true comparability of pixel brightness values in remote sensing data, providing a stable input for the precise identification of spectral features of ground objects and the spatial consistency analysis of ecological indicators.

[0065] More importantly, the acquisition process of this radiance not only serves the single link of reflectance calculation, but also serves as the basic data structure throughout the entire change detection system. Its consistency and stability at different time nodes directly determine the credibility of the difference between the two-phase data. In the dynamic monitoring of ecological protection red line areas, it is often necessary to accurately perceive extremely subtle phenomena such as vegetation degradation, increased soil exposure, and the spread of local disturbances. These phenomena often do not form obvious changes in the macroscopic structure of the image, but can only be captured through small fluctuations in pixel brightness values. Therefore, only by accurately converting the original grayscale value into radiance and correcting it through a subsequent reflectance model can the deep learning model be provided with input data that can be used to identify such "hidden changes", enabling the model to have the ability to distinguish subtle but critical disturbances in the ecosystem.

[0066] Furthermore, the process of converting the digital grayscale values recorded by the remote sensing sensor into radiant brightness with physical meaning through radiometric calibration includes: obtaining digital grayscale calibration parameters, including gain and bias, and converting through the radiometric brightness formula to obtain the radiant brightness value of each pixel in the remote sensing data in a specific band. The radiant brightness formula is: radiant brightness = gain * digital grayscale value + bias, and the unit of radiant brightness is W·m⁻²·sr⁻¹·μm⁻¹.

[0067] In the present invention's "Artificial Intelligence-Based Change Detection Method for Ecological Protection Red Line Areas," a key step in ensuring the physical accuracy and ecological interpretability of remote sensing data in identifying ecological changes is to convert the digital grayscale values recorded in the remote sensing image into physically meaningful radiation brightness values. This conversion process is not only a data formatting process, but also a scientific modeling process that maps the electrical signals obtained from the sensing device to the radiation properties of the ground object. Its core principle is derived from the remote sensing radiation calibration theory and is a crucial link in the remote sensing physical modeling system. Especially for environmentally sensitive areas such as ecological protection red line areas, any slight disturbance of vegetation or change in surface exposure may affect the stability of the entire ecosystem. Only a change detection mechanism based on real radiation energy values can achieve high-precision identification and high-confidence judgment in the artificial intelligence system.

[0068] During Earth observation, each pixel recorded by a remote sensing sensor actually corresponds to the intensity of electromagnetic radiation received by the sensor within a specific wavelength band. However, the sensor itself does not directly record the energy value. Instead, it converts the electromagnetic signal into a voltage, which is then converted into a digital grayscale value by an analog-to-digital converter for storage. Although these grayscale values appear as integers ranging from 0 to 255 (or higher, such as 10-bit, 12-bit, or even 16-bit grayscale), they have no direct physical meaning and cannot reflect the true level of energy reflected by the ground object. Therefore, before conducting any remote sensing analysis with ecological and physical significance, it is necessary to convert the grayscale values into actual radiance. This invention addresses this core requirement and constructs a systematic and precise radiometric calibration processing module.

[0069] The core of this conversion process is to obtain the response characteristics of the remote sensing equipment for each band, that is, to obtain the two calibration parameters of gain and offset through experiments or data provided by the manufacturer. Gain represents the amount of energy increase corresponding to a unit grayscale value and is a proportional coefficient; offset represents the system background output at zero grayscale and is mainly used to compensate for the nonlinear response of the sensor and the influence of electronic noise. Therefore, for the grayscale value of each pixel , in a specific band Its radiant brightness The conversion can be performed through the following linear relationship:

[0070] ;

[0071] in, Indicates band The gain, The physical significance of this formula is that it converts remote sensing signals from "relative expression" to "absolute energy," giving pixel brightness values a strict energy measurement unit—W·m⁻²·sr⁻¹·μm⁻¹—thus providing solid data support for subsequent surface reflectance calculations.

[0072] Furthermore, in step 2, the mixed reflectance spectrum in the remote sensing data is split into vegetation reflectance spectrum data and soil reflectance spectrum data by pixel decomposition method as follows:

[0073] in: For time Lower pixel The proportion of vegetation cover area; For time Lower pixel Vegetation reflectance spectral data; For time Lower pixel Soil reflectance spectral data.

[0074] In its implementation, this decomposition method is essentially based on the assumption that the observed spectral response can be viewed as a weighted superposition of two main ground object types, vegetation and soil, at the pixel scale. The representation of different ground objects in spectral space has relatively stable characteristic peaks and reflectance ranges. Therefore, as long as an appropriate coverage factor can be found, the mixed spectrum of a single observation can be decomposed into a linear combination of several components. The key concept here is the vegetation cover area ratio, which indicates the distribution of green plants within a pixel at the time of observation, thereby quantitatively distinguishing the amount of spectral mixing in a physical sense. A higher ratio means that the vegetation reflectance of the pixel has a greater impact on the overall spectrum; conversely, a lower ratio means that the soil reflectance dominates. Through this explicit weight design, the present invention provides an ecological interpretation of the observed spectrum for each pixel: it can not only identify the spatiotemporal evolution of vegetation conditions, but also reveal the various variations in soil exposure or non-green cover, thereby making observations in ecological protection redline areas more targeted and flexible.

[0075] Compared with the traditional single threshold or fixed index, this mixed pixel decomposition method has higher adaptability. Because the ecological protection red line area is not necessarily all dense and pure vegetation coverage, it also contains some shrubs, grasslands, desert edges and even human disturbance areas. These areas are likely to have a small amount of vegetation and bare soil or sandy surface within the same pixel range. If we only rely on simple vegetation indices (such as the common normalized difference vegetation index) to make judgments, we will often overestimate or underestimate the true coverage of the ecosystem and make it difficult to reflect local subtle differences. With the help of this linear decomposition, the present invention decomposes the surface reflectance at each moment into two main components: vegetation and soil, and simultaneously estimates the corresponding coverage ratio, thereby accurately capturing the diversity and differences of surface reflectance characteristics and ecological structure. After the decomposition is completed, vegetation reflectance becomes an important indicator for measuring changes in green cover, while soil reflectance is used to evaluate the impact of changes in bare surface or background interference. This "two-factor decomposition" has significant advantages in monitoring ecological protection red line areas.

[0076] What is more noteworthy is that the results output by this decomposition method are not only meaningful at the pixel scale, but can also form a continuous distribution of ecological information at a larger spatial scale. Because when processing the entire remote sensing image, each pixel will obtain the corresponding vegetation cover ratio and vegetation and soil reflectance values. Combined with high-resolution geographic information, it can form a quantitative description of the vegetation-soil structure within the ecological red line area. This can not only assist artificial intelligence models in identifying changing trends, but also provide decision makers or ecologists with a visual basis for in-depth understanding of the state of the ecosystem. For example, in which areas vegetation cover is gradually declining, in which areas increased soil reflectance means an increased risk of drought or desertification, whether there is an expansion of bare land caused by external construction, etc., early warnings and precise interventions can be obtained through analysis of decomposed data.

[0077] In addition, this spectral inversion for mixed pixels is closely connected with the subsequent deep learning process in the present invention, and the two promote each other. The decomposed vegetation and soil reflectance components will be input into the deep network in the next step to form a sparser and more ecologically structural feature space, so that the neural network is no longer disturbed by a large amount of mixed information or noise when discerning ecological changes. Correspondingly, the network can also more accurately learn the temporal laws and ecological driving factors behind the changes in surface cover. This "decomposition first, then learning" logical chain greatly reduces the convergence difficulties that may arise when the model faces non-pure pixels, and significantly improves the detection sensitivity of weak changes and early micro-disturbance. For some gradually occurring ecological degradation phenomena, even if the overall degree of spectral change is not significant, the vegetation coverage is slowly declining. Such subtle trends can often be captured in the decomposition results of the present invention and high-confidence intelligent judgments can be made based on this.

[0078] From a more macro perspective, this pixel decomposition method based on linear combinations also provides a general approach for multi-source remote sensing data, cross-temporal comparisons, and cross-regional ecological monitoring. In the dynamic management of ecological protection red lines, it is often necessary to integrate data sets from different platforms, different bands, and different temporal resolutions. The spatial resolution and spectral resolution of these remote sensing data may vary, but as long as the appropriate coverage factor and basic reflectance template can be found, the decomposition process can be applied to unify the data caliber and generate comparable vegetation and soil reflectance information. In this way, the artificial intelligence model can avoid the deviations caused by inconsistent bands or different degrees of pixel mixing during subsequent training and application, and better realize the integrated analysis of cross-scale and cross-platform ecological data.

[0079] Furthermore, in step 3, the output change probability is:

[0080]

[0081] in: Indicates time Lower pixel The probability of change; Represents the sigmoid activation function; Indicates the number of layers of the deep learning network, the value is greater than or equal to 6; Indicates the The weight of the layer; Indicates the Non-linear activation function of the layer; and Indicates two different times; Indicates time Lower pixel Surface reflectivity; Indicates time Lower pixel Surface reflectivity; For time Lower pixel The proportion of vegetation cover area; For time Lower pixel The proportion of vegetation cover area; For time Lower pixel Vegetation reflectance spectral data; For time Lower pixel Soil reflectance spectral data; For time Lower pixel Soil reflectance spectral data.

[0082] In the formula, the probability of change It is a pixel In time The confidence level of the change under the condition is between 0 and 1, which reflects a probabilistic estimate of the intensity of ecological disturbance. This output is completed through a composite structure, and the outermost layer uses the sigmoid activation function. , this function is often used to compress continuous values to between 0 and 1, which is suitable for expressing the probability judgment of "whether to change". It also makes the model have good gradient propagation characteristics, which is convenient for network convergence and generalization. The input received by the sigmoid function is not a simple differential term, but a weighted combination result after the action of multiple layers of nonlinear functions. This structure is defined.

[0083] Nonlinear mapping function of each layer Represents the deep learning network Layer-by-layer nonlinear transformation, its design follows the modeling principle of neural network response to multi-scale ecological changes. As the number of layers increases, these transformation functions gradually evolve from extracting local boundaries and difference features in the early stage to encoders of structural hierarchical features. For example, the linear rectification (ReLU) function may be used in the initial layers of the model to capture local significant change areas; the exponential linear function (ELU) is used in the middle layer to improve the response ability to weak gradients; and the hyperbolic tangent function (tanh) is used in the high layer to enhance the representation ability of temporal nonlinear evolution characteristics. This layer-by-layer combination of activation functions and weight coefficients The attenuation factor is combined with ), which enables the model to not only preserve the sensitivity of the shallow layer to the original change intensity, but also emphasize the dominance of the deep structure in the judgment of ecological patterns, thereby taking into account both local differences and overall structural evolution in change detection.

[0084] In terms of input feature construction, the present invention combines three types of key ecological change indicators into the network: the absolute difference of the original spectral reflectance , the coupling term of vegetation cover change and vegetation reflectivity change , and soil reflectivity change . These three features are combined with weight coefficients of 0.2, 0.5 and 0.3 respectively, reflecting the hierarchical design of the model's sensitivity to different types of changes. The original reflectivity difference mainly reflects the direct change of the pixel in the overall energy response, and is an important feature for capturing macro-disturbance; the change in vegetation cover combined with the reflectivity difference focuses more on the changes in the health status and distribution structure of the green ecosystem, and is the most direct ecological indicator in the red line area; the soil reflectivity term takes into account the dynamic changes of the background bare land, which is of great significance for detecting non-vegetation changes such as human disturbance, bare soil expansion, and soil drought. The combination of these three forms a multi-dimensional modeling of the ecosystem change mechanism, and also reflects the precise matching made by the present invention between feature selection and ecological response.

[0085] In addition, in order to further improve the model's ability to understand band differences, the present invention also introduces band-sensitive weights into the network structure, including 、 、 Used to adjust the Layer This design allows the deep learning model to automatically learn the response characteristics of ground objects at different wavelengths, thereby dynamically adjusting the importance of feature channels during the information fusion process. For example, in the near-infrared band, the reflectance difference of vegetation is most significant, so The weight will automatically increase, and in the mid-infrared or short-wave infrared bands, soil moisture and exposure status are more sensitive. This mechanism essentially builds a "band attention mechanism," a deep fusion strategy for remote sensing data and deep networks at the ecological understanding level.

[0086] Therefore, this change probability output model is not only a mathematical mapping function, but also a composite ecological change judgment system that integrates remote sensing spectral knowledge, ecological structure changes and artificial intelligence feature extraction mechanisms. By systematically modeling spectral differences, structural evolution and spatial responses, and combining the strong expression capabilities of multi-layer neural networks, the present invention constructs a probability estimation mechanism with ecological interpretability and intelligent discrimination capabilities. It can not only capture significant changes in ecological protection red line areas, but also stably output change probabilities in complex scenarios such as the coexistence of multiple weak disturbances, coordinated changes in vegetation and soil, and spectral contrast blur, providing a solid data foundation and model guarantee for intelligent supervision of ecological red lines.

[0087] Further, ;when When greater than or equal to 1 and less than 3, is a linear rectification function; when When greater than or equal to 3 and less than 5, is an exponential linear function; when Greater than or equal to 5 and less than hour, is the hyperbolic tangent function.

[0088] Specifically, for the The weight distribution of each layer is based on an adaptive decreasing design method. ,in Represents the total number of network layers. This design, while retaining the contribution of deep structures to change judgment, appropriately suppresses the noise interference that may be introduced by shallow layers due to excessive response. Since ecological changes often have implicit structures and slow evolution trends in space and time scales, shallow models are prone to misjudgment when faced with apparent changes (such as sudden spectral perturbations and image noise), while deep networks can extract long-term stable evolutionary patterns through more complex nonlinear transformations. Therefore, the present invention allows the weight to increase linearly with the level, and overall, because the denominator is Maintaining overall stability makes deep output more impactful but not overly enhanced, thus achieving a deep balance in the perception of change.

[0089] More importantly, in the activation function In terms of design, the present invention does not adopt a fixed structure, but sets a segmented nonlinear transformation mechanism according to the functional positioning of different levels: when At the initial layer of the network, the rectified linear unit (ReLU) activation function is used. Its core advantage lies in its ability to quickly respond to strong mutations in input features, making it suitable for capturing directly significant perturbations such as changes in the original spectrum and sharp fluctuations in coverage ratio. The model response at this stage tends to "capture," sensitively screening potential change factors and delivering rich local information to subsequent layers.

[0090] When the number of layers enters the middle section, When , the model activation function switches to the exponential linear unit (ELU), a functional structure with strong nonlinear response capabilities. It is particularly suitable for processing ecological data with certain slowly varying characteristics but the change signal may be negative. The ELU has a smooth decreasing property when the output is negative, which can respond delicately to the "micro-perturbations" changes common in ecological red lines, such as slight degradation of vegetation cover and slow changes in soil moisture. This prevents the model from discarding these weak ecological signals as noise, while maintaining a balanced input distribution, thereby enhancing the model's ability to identify transitional areas.

[0091] In the deep stage of the network, This paper chooses to use the hyperbolic tangent function (tanh) as the activation mechanism. This is because the tanh function has good bidirectional symmetry and smoothness in the interval [-1,1], which is particularly suitable for expressing the long-term trends of "positive and negative evolution" in ecosystems. For example, the gradual growth and degradation of vegetation cover and the slow contraction and expansion of wetlands can be characterized by the positive and negative output directions and amplitudes of tanh. In addition, the saturation characteristics of tanh help the model suppress extreme input values, preventing non-ecological jumps in the probability of change when affected by a single strong interference factor, making the model more robust in reflecting ecological evolution trends at the regional scale.

[0092] This segmented activation strategy combines the responsiveness of different functions in feature space, enabling the model to form a progressive expression framework of "change reception - change understanding - change judgment" throughout the network hierarchy. Shallow ReLU layers quickly capture potential changes, mid-layer ELU layers smooth transitions and extract details, and deep tanh layers perform comprehensive judgments and model spatial structure. This design not only improves the model's adaptability to both change intensity and structure, but also makes the final output change probability mathematically interpretable, providing practical ecological value.

[0093] Furthermore, in step 4, the adaptive change detection threshold is set for:

[0094]

[0095] in: is the detection reference value, and its value range is a real number greater than 0 and less than 1. The larger it is, the more sensitive the ecological protection red line area is.

[0096] The design logic of this structure is very clear: when the vegetation reflectance of a pixel changes greatly between two time points, it means that significant ecological evolution may have occurred in the area. At this time, the fractional term in the exponential function is large, which leads to the increase of the entire threshold. There is a significant attenuation, that is, the system becomes more "tolerant" to the detection results. As long as the probability of change in the model output is slightly higher than this lower threshold, it can be judged as a change, ensuring that major disturbances are not missed. When the vegetation reflectance changes very little between two time points, the system believes that the area is relatively stable and there is no significant disturbance. At this time, the threshold attenuation is very small and basically remains at Near the peak, detection standards become more stringent to avoid misjudging normal fluctuations. This mechanism is particularly important in ecological redline areas, as their goal is not only to identify severe damage but also to provide early warning of minor ecological trends and avoid false alarms. Therefore, the judgment criteria must be dynamically adaptable.

[0097] The parameters in the formula Defined as the "detection benchmark value", it is a quantitative expression of the sensitivity of the area to ecological changes. Its value ranges from (0, 1). The larger the value, the more sensitive the system is to changes, the lower the tolerance, and the easier it is to detect changes. Conversely, the smaller the value, the more significant the change signal must be before the system responds. The setting of this parameter can be combined with the management level of different ecological functional zones within the ecological protection red line area. For example, in the core protection area, A value of 0.9 or even higher can be used to improve the model's response to minor changes. Within buffer zones or restoration zones, this value can be appropriately lowered to prevent over-response to seasonal fluctuations in natural ecology. This flexible configuration of baseline values also gives the present invention high scalability and management integration capabilities, enabling it to be used not only for automatic identification but also for linkage with ecological policy standards, achieving a deep integration of technology and institutions.

[0098] Furthermore, the introduction of the exponential function itself demonstrates a high degree of depiction of the nonlinear characteristics of the change response mechanism. During ecological evolution, surface reflectivity, especially vegetation reflectivity, does not grow or decay linearly, but often evolves in a phased, critical manner. Small fluctuations do not necessarily represent a trend, but once a certain threshold of change is exceeded, a chain reaction may occur. Therefore, by achieving rapid attenuation of the threshold through an exponential form, the model actually implies an "ecological critical response mechanism," that is, once the amplitude of change exceeds the threshold of the ecosystem's self-stabilization capacity, the detection mechanism will quickly relax the judgment threshold to ensure the efficient identification of potential ecological turning points.

[0099] More importantly, this adaptive detection threshold mechanism is not only used for single-point judgment at the pixel level, but also provides basic support for subsequent spatial neighborhood consistency constraints. In the red line area where the spatial distribution is uneven and local disturbances are frequent, whether a change occurs depends not only on the comparison between the probability and threshold of the pixel itself, but also on the common evolution trend of the surrounding pixels. If the threshold of each pixel is automatically adjusted according to the change of its own ecological state, then the entire spatial judgment system will have a highly individualized but collaborative response judgment logic, realizing "multi-threshold collaborative judgment" in a spatial sense. This mechanism overcomes the difficult problem of traditional methods in balancing spatial consistency and change sensitivity, and is a key breakthrough in the intelligent direction of change detection of the present invention.

[0100] Furthermore, in step 4, based on the change probability and the adaptive change detection threshold, the detection result is determined as:

[0101]

[0102] in: Expressed as time Lower pixel The test result is 1, which means change and 0 means no change. Represents pixels The spatial neighborhood of Represents an indicator function, which is 1 when the condition is met and 0 otherwise; Indicates the number of pixels in the neighborhood; Indicates the neighborhood consistency threshold, ranging from 0 to 1; and Represents pixels The horizontal and vertical coordinates of other pixels in the spatial neighborhood of ; Indicates time Lower pixel The probability of change; Indicates time Lower pixel Adaptive change detection threshold.

[0103] The core structure of the judgment formula is a piecewise function in the form of a conditional expression, and its output is Represents pixels In time The final detection state under , a value of 1 indicates that it is judged as "changed", and a value of 0 indicates "unchanged". First, the system will determine the change probability of the current pixel Is it higher than its corresponding adaptive change detection threshold? This judgment reflects the system's confidence in the change in the ecological state of the current pixel between the two time phases. Only when this value significantly exceeds the threshold does it indicate that there is a real structural change trend at that point. However, to avoid "pseudo-changes" caused by factors such as local noise, data loss, or single-point reflectivity anomalies, the present invention further introduces spatial neighborhood consistency judgment as an auxiliary criterion.

[0104] Specifically, for each pixel, its spatial neighborhood It is a set of multiple adjacent pixels, usually in the form of a 3×3, 5×5 sliding window in a two-dimensional image. The system will count the number of pixels with the probability of change within this neighborhood. Also higher than its corresponding adaptive threshold This judgment is made through the indicator function When a pixel in the neighborhood meets the condition of "probability of change is higher than the threshold", its corresponding value is 1, otherwise it is 0. Then, the indicator function values of all the neighborhood pixels are summed to get the number of pixels in the current neighborhood that are judged to have changed. This number will be compared with a threshold For comparison, is a pre-set neighborhood consistency threshold, is the total number of pixels in the neighborhood. If the number of pixels that meet the threshold value exceeds or is equal to this ratio, it means that the change is not an isolated phenomenon, but presents spatial continuity. Finally, the system will convert the current pixel into a new pixel. If the value is determined to be changed (output is 1), otherwise, it is determined to be unchanged (output is 0).

[0105] The design of this dual-judgment mechanism essentially embeds the concept of spatial consistency into change judgment, thereby enhancing the system's ability to discern true disturbance signals within complex ecological contexts. Within ecological redline areas, true ecological change is often not confined to a single isolated pixel but rather exhibits continuous, clustered, or boundary-extending evolutionary trends. For example, phenomena such as vegetation degradation, land reclamation, and water body reduction exhibit distinct spatial extension characteristics in remote sensing imagery, while "false changes" such as natural noise, localized shadows, and imaging errors often exhibit random, isolated pixel response characteristics. Therefore, only when the majority of pixels within a local area meet the change conditions is it ecologically justified.

[0106] At the same time, this judgment structure also makes the system have strong adjustment ability and adaptability. The size of can control the tolerance of the system to local changes. A value of 0.8 or 0.9 requires that the vast majority of pixels in the neighborhood show consistent changes, which is suitable for highly sensitive ecological areas and emphasizes false alarm control; lower values A value such as 0.3 or 0.5 allows for change determination to be valid even in the presence of local inconsistencies, making it suitable for transition zones within change boundaries or areas prone to human disturbance. This flexible setting not only improves the accuracy of change determination overall, but also adapts to the different management zones, functional areas, or natural landscape features within ecological redline areas, achieving a spatially differentiated intelligent identification strategy.

[0107] Furthermore, this spatial neighborhood consistency mechanism helps improve the spatial coherence and structural stability of the model output, avoiding the appearance of "spotted" change patterns or broken edges in the final change image. This makes the change detection results more geographically interpretable and more suitable for subsequent ecological statistical analysis, trend simulation, and conservation intervention. In conjunction with the aforementioned adaptive change detection threshold mechanism, the present invention constructs a change detection judgment system consisting of three layers of logic: probability output, threshold adjustment, and spatial constraint. From sensing the magnitude of ecological change, to judging its credibility, and then to verifying its spatial structural consistency, this system forms an intelligent recognition mechanism for the entire chain of ecological disturbance events, at multiple scales and dimensions.

[0108] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions merely illustrate the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed by the present invention is defined by the appended claims and their equivalents.

Claims

1. An artificial intelligence-based ecological protection red line area change detection method, characterized by: include: Step 1: Perform atmospheric correction and radiation correction on the remote sensing data at each time to convert the remote sensing data at each time into directly comparable surface reflectance; Step 2: Split the mixed reflectance spectrum in the remote sensing data into vegetation reflectance spectrum data and soil reflectance spectrum data by pixel decomposition method; the vegetation reflectance spectrum data and soil reflectance spectrum data are both bi-temporal data; Step 3: Build a deep learning model that uses the difference between vegetation reflectance spectral data and soil reflectance spectral data from dual-temporal data as input, extracts the change characteristics of the ecological protection area, and outputs the change probability; Step 4: Set an adaptive change detection threshold based on the sensitivity of the ecological protection redline area; determine the detection result based on the change probability and the adaptive change detection threshold; In step 3, the output change probability is: in: Indicates time Lower pixel The probability of change; Represents the sigmoid activation function; Indicates the number of layers of the deep learning network, the value is greater than or equal to 6; Indicates the The weight of the layer; Indicates the Non-linear activation function of the layer; and Indicates two different times; Indicates time Lower pixel Surface reflectivity; Indicates time Lower pixel Surface reflectivity; For time Lower pixel The proportion of vegetation cover area; For time Lower pixel The proportion of vegetation cover area; For time Lower pixel Vegetation reflectance spectral data; For time Lower pixel Soil reflectance spectral data; For time Lower pixel Soil reflectance spectral data.

2. The artificial intelligence-based ecological protection red line area change detection method according to claim 1, characterized in that: In step 1, at time The surface reflectivity at : in: Indicates time Lower pixel Surface reflectivity; Indicates time The pixels received The radiance of Indicates the distance between the Earth and the Sun; Indicates time Solar irradiance under Indicates time The solar zenith angle below is the horizontal coordinate of the pixel; is the vertical coordinate of the pixel.

3. The method for detecting changes in ecological protection red line areas based on artificial intelligence according to claim 2, characterized in that: The radiation brightness is obtained through the following process: a radiation detection instrument is installed on the remote sensing sensor; the radiation detection instrument includes a multispectral sensor and an imaging spectrometer; when the remote sensing sensor passes over the ground, the sunlight reflected from the surface propagates through the atmosphere to reach the photosensitive element of the remote sensing sensor, and the remote sensing sensor records the electromagnetic radiation energy intensity received in different bands and stores it in the form of digital grayscale values; the digital grayscale values recorded by the remote sensing sensor are converted into radiation brightness with physical meaning through radiation calibration processing.

4. The artificial intelligence-based ecological protection red line area change detection method according to claim 3, characterized in that: The process of converting the digital grayscale values recorded by remote sensing sensors into radiance with physical meaning through radiometric calibration includes: obtaining digital grayscale calibration parameters, including gain and bias, and converting them through the radiometric brightness formula to obtain the radiance value of each pixel in the remote sensing data in a specific band. The radiance formula is: radiance = gain * digital grayscale + bias. The unit of radiance is W·m⁻ 2 ·sr⁻ 1 μm⁻ 1 .

5. The method for detecting changes in ecological protection red line areas based on artificial intelligence according to claim 4, characterized in that: In step 2, the mixed reflectance spectrum in the remote sensing data is split into vegetation reflectance spectrum data and soil reflectance spectrum data by pixel decomposition method: in: For time Lower pixel The proportion of vegetation cover area; For time Lower pixel Vegetation reflectance spectral data; For time Lower pixel Soil reflectance spectral data.

6. The method for detecting changes in ecological protection red line areas based on artificial intelligence according to claim 5, characterized in that: ;when When greater than or equal to 1 and less than 3, is a linear rectification function; when When greater than or equal to 3 and less than 5, is an exponential linear function; when Greater than or equal to 5 and less than hour, is the hyperbolic tangent function.

7. The artificial intelligence-based ecological protection red line area change detection method according to claim 6, characterized in that: In step 4, set the adaptive change detection threshold for: in: is the detection reference value, and its value range is a real number greater than 0 and less than 1. The larger it is, the more sensitive the ecological protection red line area is.

8. The method for detecting changes in ecological protection red line areas based on artificial intelligence according to claim 7, characterized in that: In step 4, based on the change probability and the adaptive change detection threshold, the detection result is determined as: in: Expressed as time Lower pixel The test result is 1, which means change and 0 means no change. Represents pixels The spatial neighborhood of Represents an indicator function, which is 1 when the condition is met and 0 otherwise; Indicates the number of pixels in the neighborhood; Indicates the neighborhood consistency threshold, ranging from 0 to 1; and Represents pixels The horizontal and vertical coordinates of other pixels in the spatial neighborhood of ; Indicates time Lower pixel The probability of change; Indicates time Lower pixel Adaptive change detection threshold.

Images