Photovoltaic power station remote sensing identification method and system based on time constraint

Abstract

The invention relates to the technical field of remote sensing image processing, and particularly discloses a photovoltaic power station remote sensing identification method and system based on time constraint, and the method comprises the steps: reading image data, and obtaining projection parameters and geographical transformation parameters; processing the image data by a minimum-maximum normalization method; selecting a reference year, acquiring reference year photovoltaic power station vector data processed by a normalization method, converting image data and vector data processed by the reference year normalization method into the same coordinate system according to projection parameters and / or geographical transformation parameters, rasterizing the vector data, and generating a binary mask consistent with the resolution of the image data; and by taking the binary mask as a label, designing a deep learning network structure to train image data, and outputting a predicted binary mask. According to the method, the uncertainty of single-phase image classification is remarkably reduced, and the stability and reliability of a multi-phase classification result are ensured.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing image processing, and in particular to a photovoltaic power station remote sensing identification method and system based on time constraints. Background Art

[0002] Current methods for remote sensing identification of photovoltaic power plants suffer from the following major limitations. First, existing methods overly rely on the spectral and texture features of a single temporal image for classification and identification, failing to fully utilize the information correlations between multi-temporal images and ignoring the temporal continuity of photovoltaic power plant construction and development. Second, the independent processing of each period of imagery results in a lack of temporal consistency in classification results. The same plot of land may experience repeated classification results in different years, lacking effective constraints on the photovoltaic power plant construction process. Furthermore, traditional methods are susceptible to interference from atmospheric conditions such as clouds and aerosols. Variations in light intensity and specular reflections can also significantly alter the spectral characteristics of the objects, severely impacting the stability of classification results. Furthermore, the lack of a unified standard for classification results across different periods makes quantitative comparative analysis across multiple years difficult, and thus fails to accurately reflect the changing trends in photovoltaic power plant area. Summary of the Invention

[0003] The purpose of the present invention is to provide a photovoltaic power station remote sensing identification method based on time constraints, which is suitable for scenarios such as energy infrastructure monitoring and land use change analysis. It ensures the logical continuity of the development of photovoltaic power stations, significantly reduces the uncertainty of single-period image classification, and can effectively reduce the impact of environmental factors on classification results, thereby ensuring the stability and reliability of multi-period classification results.

[0004] The purpose of the present invention can be achieved through the following technical solutions:

[0005] A photovoltaic power station remote sensing identification method based on time constraints, comprising:

[0006] Read image data and obtain projection parameters and geographic transformation parameters;

[0007] The minimum-maximum normalization method is used to process image data;

[0008] Selecting a base year and obtaining the base year photovoltaic power station vector data processed by the normalization method, converting the base year normalization image data and the vector data to the same coordinate system according to the projection parameters and / or geographic transformation parameters, rasterizing the vector data, and generating a binary mask consistent with the resolution of the image data;

[0009] Using the binary mask as a label, designing a deep learning network structure to train the image data to output a predicted binary mask;

[0010] Designing a loss function that balances classification accuracy and temporal continuity consistency to process the predicted binary mask;

[0011] performing dynamic morphological optimization on the predicted binary mask processed by the loss function;

[0012] The optimized binary mask is corrected through area change analysis and time series consistency test to obtain the photovoltaic power station remote sensing recognition result.

[0013] In a further embodiment, the method of reading image data and obtaining projection parameters and geographic transformation parameters includes:

[0014] Read image pixel values through GDAL;

[0015] Extract projection parameters to ensure that the image has a clear spatial coordinate system;

[0016] Extract geographic transformation parameters and define the conversion relationship between pixel coordinates and geographic coordinates.

[0017] In a further embodiment, the method of processing image data using the minimum-maximum normalization method includes:

[0018] Through formula 1: X norm =(X max -X min ) / (XX min ) to obtain the normalized image data X norm , where X max is the maximum value of the image data read, X min is the minimum value of the read image data, and X is the unprocessed value of the read image data.

[0019] In a further embodiment, the image data includes image pixel values.

[0020] In a further embodiment, the photovoltaic region in the binary mask is assigned a value of 1, and the non-photovoltaic region is assigned a value of 0.

[0021] In a further embodiment, the method of using the binary mask as a label and designing a deep learning network structure to train the image data output to predict the binary mask includes:

[0022] Based on the deep learning model, the binary mask of the benchmark year is used as the label. The feature extractor of the multi-layer convolutional network structure of the deep learning model is combined with instance normalization and nonlinear activation function to extract the temporal features before and after the benchmark year. The temporal features before and after are adaptively fused through the temporal attention module to output the predicted binary mask.

[0023] In a further embodiment, the method for processing the predicted binary mask by designing a loss function that balances classification accuracy and temporal continuity consistency includes:

[0024] By adopting weighted binary cross entropy loss, a higher weight is given to the photovoltaic class to alleviate the class imbalance and ensure the correct pixel-level classification of photovoltaic / non-photovoltaic areas in a single phase;

[0025] The weighting coefficient is set according to the distance between the time and the base year, and the decreasing speed of the constraint strength of the weighting coefficient is controlled by the attenuation exponential function.

[0026] In a further embodiment, the performing of a dynamic morphological optimization method on the predicted binary mask obtained by processing the loss function comprises:

[0027] Performing an opening operation on the predicted binary mask processed by the loss function before the base year to remove noise;

[0028] A closing operation is performed on the predicted binary mask processed by the loss function after the base year to fill the holes.

[0029] In a further embodiment, the method of correcting the optimized binary mask through area change analysis and time series consistency test to obtain the photovoltaic power station remote sensing identification result includes:

[0030] Count the total number of photovoltaic pixels in each year, draw an area change curve, and verify whether the curve trend meets the quantitative indicators;

[0031] Verify whether the recognition results are contradictory in the time dimension.

[0032] The present invention also proposes a photovoltaic power station remote sensing identification system based on time constraints, comprising:

[0033] The acquisition module is used to read image data and obtain projection parameters and geographic transformation parameters;

[0034] A first processing module is used to process image data using a minimum-maximum normalization method;

[0035] a second processing module, configured to select a base year and obtain the base year photovoltaic power station vector data processed by the normalization method, convert the image data and the vector data processed by the base year normalization method into the same coordinate system according to the projection parameters and / or geographic transformation parameters, rasterize the vector data, and generate a binary mask consistent with the resolution of the image data;

[0036] A third processing module is configured to use the binary mask as a label to design a deep learning network structure to train the image data to output a predicted binary mask;

[0037] a fourth processing module, configured to design a loss function that balances classification accuracy and temporal continuity consistency to process the predicted binary mask;

[0038] a fifth processing module, configured to perform dynamic morphological processing on the prediction mask obtained through loss function optimization;

[0039] The feedback module corrects the optimized binary mask through area change analysis and time series consistency test to obtain the remote sensing identification results of the photovoltaic power station.

[0040] Beneficial effects of the present invention:

[0041] This method establishes a bidirectional temporal constraint mechanism based on high-precision labeled data from a standard year. It extracts stable feature expressions through a deep learning model and effectively integrates multi-period image information using a temporal attention mechanism. This complete temporal constraint framework ensures the logical continuity of photovoltaic power station development and significantly reduces the uncertainty of single-period image classification. By organically combining deep learning with temporal constraints, this method effectively reduces the impact of environmental factors on classification results, ensuring the stability and reliability of multi-period classification results. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following is a brief introduction to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0043] Figure 1 This is a flowchart of a photovoltaic power station remote sensing identification method based on time constraints in an embodiment. DETAILED DESCRIPTION

[0044] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present invention.

[0045] like Figure 1 As shown, a photovoltaic power station remote sensing identification method based on time constraints includes:

[0046] S1. Read image data and obtain projection parameters and geographic transformation parameters;

[0047] S2, minimum-maximum normalization method to process image data;

[0048] S3. Selecting a base year and obtaining normalized base year photovoltaic power station vector data, converting the base year normalized image data and vector data to the same coordinate system based on the projection parameters and / or geographic transformation parameters, rasterizing the vector data, and generating a binary mask consistent with the resolution of the image data.

[0049] S4, using the binary mask as a label, designing a deep learning network structure to train the image data to output a predicted binary mask;

[0050] S5. Design a loss function that balances classification accuracy and temporal continuity consistency to process the predicted binary mask; this can be achieved by proposing a temporal attention module, fusing multi-phase features through convolution and sigmoid dynamic weighting, and designing a two-stage spatiotemporal constraint loss function to jointly optimize classification accuracy and temporal logic consistency;

[0051] S6. performing dynamic morphological optimization on the prediction mask obtained by optimizing the loss function; this allows the development of a dynamic morphological optimization strategy that adaptively adjusts the size of the structural element according to the age of the data;

[0052] S7. The optimized binary mask is modified through area change analysis and time series consistency test to obtain the photovoltaic power station remote sensing identification result.

[0053] This method improves classification accuracy: the temporal attention module increases the Intersection over Union (IoU) by over 15%. Temporal consistency is ensured: the constraint loss reduces the logical error rate to <2%. Computational efficiency: dynamic morphological optimization reduces post-processing time by 30% compared to fixed parameter methods.

[0054] In some embodiments, step S1 may include:

[0055] Read image pixel values through GDAL; obtain numerical matrices of remote sensing images (such as multi-band reflectance, temperature).

[0056] Extract projection parameters to ensure that the image has a clear spatial coordinate system (such as WGS84, UTM);

[0057] Extract geographic transformation parameters and define the conversion relationship between pixel coordinates and geographic coordinates. For example, geographic transformation parameters such as the upper left corner latitude and longitude and pixel resolution are used to input the upper left corner coordinates and resolution through the transformation relationship to generate a georeferenced image. The transformation relationship can be an existing affine transformation.

[0058] Parameters generally include the upper left corner X coordinate (longitude / easting coordinate), pixel width (X-direction resolution), rotation parameter (usually 0), upper left corner Y coordinate (latitude / northing coordinate), rotation parameter (usually 0), and pixel height (Y-direction resolution, usually a negative value because the image coordinate system increases from top to bottom).

[0059] This allows for spatial alignment: subsequent processing (such as cropping and multi-temporal overlay) relies on a unified geographic reference to avoid pixel misalignment.

[0060] Area calculation: The number of pixels can be converted into actual area through geographic transformation parameters (e.g. at 30m resolution, 1 pixel = 900m 2 ).

[0061] This ensures the spatial accuracy of subsequent analysis (such as overlay with vector data and area statistics).

[0062] In some embodiments, step S2 may include:

[0063] Through formula 1: X norm =(X max -X min ) / (XX min ) to obtain the normalized image data X norm , where X max is the maximum value of the image data read, X min is the minimum value of the read image data, and X is the unprocessed value of the read image data.

[0064] where X max and X min Can be specified manually, such as the Landsat reflectivity theoretical range [0,1], so that X can be specified max is 1, X min It is set to 0, or truncated according to the actual pixel distribution of the image (such as taking the 2% to 98% quantile to avoid the influence of outliers).

[0065] Normalization can eliminate dimensionality differences: scaling pixel values from images of different time phases and sensors to the same range (e.g., [0, 1]) prevents differences in numerical scales from impacting model training. It can also suppress outliers: by truncating extreme values (e.g., cloud cover and noise), it improves model robustness. It can also standardize model inputs: neural networks are sensitive to the numerical distribution of input data, and normalization can accelerate gradient descent convergence. It can also facilitate cross-temporal comparisons: images from different times must be at the same numerical scale to be effectively compared (e.g., for NDVI calculation and change detection). It can also improve model generalization, making it independent of the numerical range of the original data.

[0066] For example, input data: Landsat8 image (band 5: near infrared, original DN value range 0 to 65535).

[0067] Normalization: Reflectance is unified to [0, 1], allowing direct joint model training with data from other sensors such as Sentinel-2. Without normalization, the model may overemphasize Landsat features because Landsat's DN value (0-65535) is much larger than Sentinel-2's reflectance (0-1).

[0068] Through this preprocessing process, the image data is standardized in both spatial and numerical dimensions, providing a high-quality input foundation for subsequent deep learning models (such as U-Net and ResNet).

[0069] In some embodiments, the image data includes image pixel values.

[0070] In some embodiments, the operation steps of step S3 can be: select a certain year, i.e., the base year, give priority to years with comprehensive data and high quality, ensure that key indicators such as GDP, carbon emissions, land use, etc. are not missing, manually mark the photovoltaic power stations in the processed image data, and extract the boundary data of the photovoltaic power stations, usually in vector format (such as Shapefile, GeoJSON), containing polygon (Polygon) geometric information, accurately mark the actual range of the photovoltaic power station, vector data and remote sensing image data may use different coordinate reference systems, so conversion is required to ensure that the projection reference is the same. Convert the vector polygon to a binary mask with the same resolution as the remote sensing image, in which the photovoltaic area is assigned a value of 1 and the non-photovoltaic area is assigned a value of 0. The pixel size after rasterization must be consistent with the remote sensing image (such as 30m / pixel).

[0071] High-precision vector data (accurate to the boundaries of individual photovoltaic panels) is used to avoid the loss of small polygons during rasterization, preserve small photovoltaic facilities, and serve as a benchmark for spatiotemporal analysis, ensuring that the classification results of all years are consistent with the logic of the base year.

[0072] In some embodiments, step S4 may include:

[0073] Based on the deep learning model, the binary mask of the benchmark year is used as the label. The feature extractor of the multi-layer convolutional network structure of the deep learning model is combined with instance normalization and nonlinear activation function to extract the temporal features before and after the benchmark year. The temporal features before and after are adaptively fused through the temporal attention module to output the predicted binary mask.

[0074] The deep learning model can be an improved U-Net architecture, and the feature extractor contains instance normalization (InstanceNorm) and ReLU activation function.

[0075] This solves the problem of feature fusion in multi-temporal image classification, allowing the model to automatically learn the weights of images from different times, eliminating the need for manually set fixed rules. For example, in images from 2020 and 2022, the model should focus on newly added photovoltaic areas rather than the unchanging background.

[0076] Input to the deep learning model: feature maps of the previous and next phases (such as the features of the images in 2020 and 2022 after passing through the encoder). Through convolution operations, such as using 3×3 convolution to analyze the spatial correlation of temporal features. Through Sigmoid activation, the convolution output is compressed to the interval [0,1] to generate an attention weight map, which represents the importance of features at different positions. Then, adaptive fusion is performed, and the weight map is multiplied element by element with the original features to weightedly fuse the temporal features. This can more accurately identify newly added photovoltaic areas, that is, high-weight focused changes, and is more robust to temporary noise such as clouds and shadows, that is, low-weight suppression of unchanged areas. Through this deep model design, the model can not only extract spatial features, but also adaptively utilize temporal dimension information, significantly improving the accuracy of multi-phase classification (such as an increase of more than 15% in IoU).

[0077] In some embodiments, step S5 may include:

[0078] Because the number of photovoltaic regions (positive samples) in remote sensing images is typically much smaller than the background (negative samples), directly predicting a binary mask from the model output or using a common loss function for balancing results in a biased model toward the majority class (background). Using a binary cross-entropy loss ensures correct pixel-level classification of photovoltaic / non-photovoltaic regions within a single time phase. Weighting can increase the loss weight for photovoltaic regions, forcing the model to focus on rare classes. Specifically, a weighted binary cross-entropy loss function can be used to obtain appropriate weight coefficients, providing a flexible balance between precision and recall to suit different task requirements.

[0079] The weighting coefficient is set according to the distance from the base year, and the decaying exponential function is used to control the rate of decrease of the constraint strength of the weighting coefficient. The farther the time is from the base year, the higher the data uncertainty (such as image quality degradation and accumulated annotation errors), and the constraint strength should be gradually reduced. The weight of the year can be calculated according to the formula: It is calculated as follows: λ0 is the initial weight of the base year (e.g., 1.0), α is the decay rate (e.g., 0.1), which controls the speed of decline of the constraint strength, t0 is the base year, and t is the calculation year.

[0080] Initial classification results (such as the PV / non-PV binary image output by a deep learning model) often contain two types of noise: small misclassified points (such as buildings or vegetation mistakenly identified as PV) and holes within the PV area (area discontinuities caused by cloud cover or classification errors).

[0081] In some embodiments, step S6 may include:

[0082] Performing an opening operation on the predicted binary mask processed by the loss function before the base year to remove noise;

[0083] Operation order: Erosion followed by Dilation. Purpose: Eliminates isolated misclassified small points (e.g., noise whose area is smaller than the structural element). Smoothes the edges of the photovoltaic region.

[0084] Close the predicted binary mask processed by the loss function after the base year to fill any holes. The order of operations is: dilation followed by erosion. Purpose: Fill small holes within the photovoltaic region (e.g., missing classifications due to cloud cover). Connect adjacent fragmented regions.

[0085] During the opening operation, the structuring element size is dynamically adjusted, and strict denoising is performed on images older than the baseline year. For example, adjusting the structuring element size to a larger size (e.g., 7×7 pixels) increases the risk of misclassification due to lower quality historical images (e.g., poor resolution, high noise). A conservative strategy is used to avoid false detections, even at the expense of some true small photovoltaic areas. This approach completely removes small noise, but may also miss true small photovoltaic areas.

[0086] During the closing operation, details are preserved in imagery after the baseline year, and the structuring element size is adjusted to a smaller size (e.g., 3×3 pixels). Modern imagery has high resolution, good quality, and rich details. Small targets such as distributed rooftop photovoltaics need to be preserved. Only minor noise is removed to preserve more of the true photovoltaic area.

[0087] Through dynamic morphological optimization, this method adapts to the high-detail requirements of modern imagery while maintaining strict quality control for historical data. It achieves denoising (eliminating isolated misclassified points), infilling (repairing discontinuities in photovoltaic regions), and self-adaptation (adjusting processing intensity based on data age, balancing accuracy and detail). This strategy is particularly well-suited for long-term time series analysis (such as monitoring photovoltaic power plant expansion and tracking deforestation), where data quality and target characteristics vary significantly across time periods.

[0088] In some embodiments, step S7 may include:

[0089] Count the total number of photovoltaic pixels in each year, draw the area change curve, and verify whether the curve trend conforms to the quantitative indicators; this can verify whether the change of the photovoltaic power station area over time conforms to the actual law (such as continuous growth, phased expansion), such as multiplying by the area of a single pixel (such as Landsat 30m resolution, 1 pixel = 900m2), to get the annual total area, check whether the curve trend is reasonable, reasonable cases: monotonically increasing (new photovoltaic), step-by-step growth (phased construction). Abnormal cases: sudden decrease in area (possible misclassification), violent fluctuations (noise interference). Annual growth rate: compared with policy documents or energy reports (such as a province's annual average photovoltaic growth of 20%). Goodness of fit (R2): R of linear fit 2 >0.9 indicates a stable trend.

[0090] Verify whether the identification results are inconsistent across time. This ensures that the classification results are consistent across time (e.g., a PV plant being misclassified before completion). Specific verification rules can include: Forward Check: If an area is classified as PV in the base year (e.g., 2020), it must remain PV in subsequent years (2021–2023) (additions are allowed, but deletions are not). Backward Check: If an area is not PV in the base year, PV must not appear in the historical years (2015–2019) (to avoid "time travel" errors). Addition rationality: New PV areas must be spatiotemporally aligned with construction records (e.g., satellite construction imagery, approval documents). Anomaly detection example: Error case: An area was classified as PV in 2019 but changed to non-PV in 2020. Possible cause: Cloud cover caused the misclassification in 2020, requiring manual review. Correction: Force the area to be reclassified as PV in 2020 (following the forward constraint).

[0091] The identification results can be used for policy evaluation: verifying whether PV installed capacity meets planned targets. Illegal construction monitoring: identifying PV projects that have been built without approval. Carbon accounting: Accurately calculating PV emissions reductions requires consistent classification results across time and space. This provides reliable data support for energy decision-making and environmental monitoring.

[0092] In one specific embodiment, image data is first read using the GDAL open-source library, simultaneously acquiring projection and geographic transformation parameters to ensure geographic reference integrity for subsequent processing. The raw data is then normalized using a minimum-maximum normalization method to effectively eliminate outliers and invalid values and unify the numerical ranges of images from different time phases to the same scale, laying the foundation for subsequent deep learning classification.

[0093] Using vector data from a standard year of photovoltaic power plants as a benchmark, a coordinate system conversion was performed to ensure that the vector data and the remote sensing imagery had the same projection basis. Rasterization was then used to convert the vector data into a binary mask with the same spatial resolution as the remote sensing imagery. Photovoltaic areas were assigned a value of 1, and non-photovoltaic areas were assigned a value of 0. This ensured that even small photovoltaic facilities were not lost during the conversion process.

[0094] A deep learning model based on the U-Net architecture was constructed. Its feature extractor utilizes a multi-layer convolutional network structure, and improves feature extraction capabilities through instance normalization and ReLU activation functions. The model's core innovation lies in the introduction of a temporal attention module, which uses convolution operations and sigmoid activation functions to adaptively integrate features from previous and subsequent temporal phases, effectively improving the model's ability to utilize temporal features.

[0095] The basic classification loss uses binary cross entropy and addresses class imbalance through weight adjustment. The temporal constraint loss implements constraints on the baseline data and continuity constraints on adjacent time phases by setting different weight coefficients. The decay exponent controls how the constraint strength changes over time.

[0096] Model training uses benchmark year data as a benchmark and employs various data augmentation strategies to improve model generalization. Spatial transformations such as random rotation and flipping, as well as spectral enhancement methods such as brightness and contrast adjustments, are used to increase the diversity of training samples. The Adam optimizer is used during training, and an adaptive learning rate strategy is employed to ensure model convergence to the optimal solution.

[0097] After completing training for the base year, the model classifies and identifies the time phases before and after the base year. For years after the base year, forward constraints are used to ensure that identified PV regions remain stable in subsequent years, while allowing for the emergence of new PV regions. For years before the base year, backward constraints are used to ensure that regions identified as non-PV in the base year remain non-PV in historical years, effectively avoiding temporal logic errors.

[0098] Spatial optimization of classification results is primarily achieved through morphological operations. Opening operations are first used to remove small noise points, followed by closing operations to fill small holes within the photovoltaic region. Structuring elements of different sizes are used for different years. Larger structuring elements are used before the base year to strictly control misclassification, while smaller structuring elements are used after the base year to preserve more detailed information.

[0099] By calculating the total area of PV power plants in each year, we analyze area changes and assess the rationality of these trends. We focus on the logical relationships between adjacent years and ensure that the classification results conform to the objective laws of PV power plant development, thus ensuring the reliability of the results.

[0100] According to the above embodiment, it is obvious that a photovoltaic power station remote sensing identification system based on time constraints can be obtained, including:

[0101] The acquisition module is used to read image data and obtain projection parameters and geographic transformation parameters;

[0102] A first processing module is used to process image data using a minimum-maximum normalization method;

[0103] a second processing module, configured to select a base year and obtain the base year photovoltaic power station vector data processed by the normalization method, convert the image data and the vector data processed by the base year normalization method into the same coordinate system according to the projection parameters and / or geographic transformation parameters, rasterize the vector data, and generate a binary mask consistent with the resolution of the image data;

[0104] A third processing module is configured to use the binary mask as a label to design a deep learning network structure to train the image data to output a predicted binary mask;

[0105] a fourth processing module, configured to design a loss function that balances classification accuracy and temporal continuity consistency to process the predicted binary mask;

[0106] a fifth processing module, configured to perform dynamic morphological processing on the prediction mask obtained through loss function optimization;

[0107] The feedback module corrects the optimized binary mask through area change analysis and time series consistency test to obtain the remote sensing identification results of the photovoltaic power station.

[0108] The U-Net architecture deep learning model used in the present invention can use other deep learning models, such as Deeplab, SwinTransformer, etc.

[0109] The binary cross entropy loss function used in the present invention may use other loss functions, such as FocalLoss, weighted cross entropy, etc.

[0110] The Adam optimizer used in the present invention may use other optimization algorithms, such as the stochastic gradient method, AdaBelief, etc. It is conceivable that the subject performing the above-mentioned actions may be a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the above-mentioned photovoltaic power station remote sensing identification method based on time constraints are implemented.

[0111] It should be noted that the terms "first," "second," etc., in this application are used to distinguish similar objects, and are not necessarily used to describe a specific order or precedence. It should be understood that the terms used in this manner are interchangeable where appropriate, so as to facilitate the embodiments of the present application described herein.

[0112] Throughout this specification, references to terms such as "one embodiment," "example," or "specific example" indicate that the specific features, structures, materials, or characteristics described in conjunction with that embodiment or example are included in at least one embodiment or example of the present invention. In this specification, schematic representations of these terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0113] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are merely illustrative of 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, and such changes and modifications fall within the scope of the invention as claimed.

Claims

1. A photovoltaic power station remote sensing identification method based on time constraints, characterized in that: include: Read image data and obtain projection parameters and geographic transformation parameters; The minimum-maximum normalization method is used to process image data; Selecting a base year and obtaining the base year photovoltaic power station vector data processed by the normalization method, converting the base year normalization image data and the vector data to the same coordinate system according to the projection parameters and / or geographic transformation parameters, rasterizing the vector data, and generating a binary mask consistent with the resolution of the image data; Using the binary mask as a label, designing a deep learning network structure to train the image data to output a predicted binary mask; Designing a loss function that balances classification accuracy and temporal continuity consistency to process the predicted binary mask; performing dynamic morphological optimization on the predicted binary mask processed by the loss function; The optimized binary mask is corrected through area change analysis and time series consistency test to obtain the photovoltaic power station remote sensing recognition result.

2. The photovoltaic power station remote sensing identification method based on time constraints according to claim 1 is characterized in that: The method of reading image data and obtaining projection parameters and geographic transformation parameters includes: Read image pixel values through GDAL; Extract projection parameters to ensure that the image has a clear spatial coordinate system; Extract geographic transformation parameters and define the conversion relationship between pixel coordinates and geographic coordinates.

3. The photovoltaic power station remote sensing identification method based on time constraints according to claim 1, characterized in that: The method for processing image data using the minimum-maximum normalization method includes: Through formula 1: X norm =(X max -X min ) / (XX min ) to obtain the normalized image data X norm , where X max is the maximum value of the image data read, X min is the minimum value of the read image data, and X is the unprocessed value of the read image data.

4. The photovoltaic power station remote sensing identification method based on time constraints according to claim 3 is characterized in that: The image data includes image pixel values.

5. The photovoltaic power station remote sensing identification method based on time constraints according to claim 1 is characterized in that: In the binary mask, the photovoltaic area is assigned a value of 1, and the non-photovoltaic area is assigned a value of 0.

6. The photovoltaic power station remote sensing identification method based on time constraints according to claim 1, characterized in that: The method of using the binary mask as a label and designing a deep learning network structure to train the image data output to predict the binary mask includes: Based on the deep learning model, the binary mask of the benchmark year is used as the label. The feature extractor of the multi-layer convolutional network structure of the deep learning model is combined with instance normalization and nonlinear activation function to extract the temporal features before and after the benchmark year. The temporal features before and after are adaptively fused through the temporal attention module to output the predicted binary mask.

7. The photovoltaic power station remote sensing identification method based on time constraints according to claim 1, characterized in that: The method for designing a loss function that balances classification accuracy and temporal continuity consistency to process the predicted binary mask includes: By adopting weighted binary cross entropy loss, a higher weight is given to the photovoltaic class to alleviate the class imbalance and ensure the correct pixel-level classification of photovoltaic / non-photovoltaic areas in a single phase; The weighting coefficient is set according to the distance between the time and the base year, and the decreasing speed of the constraint strength of the weighting coefficient is controlled by the attenuation exponential function.

8. The photovoltaic power station remote sensing identification method based on time constraints according to claim 1, characterized in that: The method of performing dynamic morphological optimization on the predicted binary mask obtained through loss function processing includes: Performing an opening operation on the predicted binary mask processed by the loss function before the base year to remove noise; A closing operation is performed on the predicted binary mask processed by the loss function after the base year to fill the holes.

9. The photovoltaic power station remote sensing identification method based on time constraints according to claim 1, characterized in that: The method for obtaining a photovoltaic power station remote sensing recognition result by correcting the optimized binary mask through area change analysis and time sequence consistency test includes: Count the total number of photovoltaic pixels in each year, draw an area change curve, and verify whether the curve trend meets the quantitative indicators; Verify whether the recognition results are contradictory in the time dimension.

10. A photovoltaic power station remote sensing identification system based on time constraints, characterized in that: include: The acquisition module is used to read image data and obtain projection parameters and geographic transformation parameters; A first processing module is used to process image data using a minimum-maximum normalization method; a second processing module, configured to select a base year and obtain the base year photovoltaic power station vector data processed by the normalization method, convert the image data and the vector data processed by the base year normalization method into the same coordinate system according to the projection parameters and / or geographic transformation parameters, rasterize the vector data, and generate a binary mask consistent with the resolution of the image data; A third processing module is configured to use the binary mask as a label to design a deep learning network structure to train the image data to output a predicted binary mask; a fourth processing module, configured to design a loss function that balances classification accuracy and temporal continuity consistency to process the predicted binary mask; a fifth processing module, configured to perform dynamic morphological processing on the prediction mask obtained through loss function optimization; The feedback module corrects the optimized binary mask through area change analysis and time series consistency test to obtain the remote sensing identification results of the photovoltaic power station.

Images