Large-area photovoltaic land classification method and system based on combined decision tree and deep learning

Abstract

The present invention discloses a large-area photovoltaic land classification method that combines decision trees and deep learning, belonging to the field of machine learning and remote sensing intelligent mapping. In view of the problem that single traditional machine learning and deep learning are difficult to balance the accuracy and speed of large-area remote sensing image classification, the present invention produces a global representative regional sample set based on remote sensing images, constructs a photovoltaic land index, and adopts a two-step classification method: the first step is to construct a decision tree classifier based on features such as the photovoltaic land index, and realize the rapid classification of large-area photovoltaic land on the GEE remote sensing cloud platform. After the classification results are processed by post-processing, raster-vector conversion, and buffer, a coarse classification area of ​​photovoltaic land is generated; the second step is to add a secondary branch and an index space attention module to the baseline ResUNet++ model to construct a photovoltaic land classification model that integrates multi-scale photovoltaic land features, and accurately identify photovoltaic land from the coarse classification area based on the model, so as to realize rapid and accurate mapping of large-area photovoltaic land.

Description

Technical Field

[0001] The present invention belongs to the field of machine learning and remote sensing intelligent mapping, and in particular relates to a large-area photovoltaic land classification method and system combining decision trees and deep learning. Background Art

[0002] Photovoltaic land (PVL) refers to a new land use pattern involving the installation of photovoltaic panels on underlying land uses / covers, including bare land (such as deserts, bare soil, and saline-alkali land), vegetation (such as grassland, woodland, and farmland), water surfaces, and rooftops. Large-scale PV land monitoring can provide data support for PV project site selection, PV power generation forecasting, the ecological and environmental impacts of PV project construction, and the formulation of green energy policies.

[0003] Existing PV data is primarily published by energy organizations and related think tanks, using megawatts (GW) as the unit of measurement and an annual timescale. This data has the advantage of reflecting national or regional PV installed capacity and the location of large PV facilities. However, it lacks information on area and the location of smaller PV facilities. Therefore, this data has limited value for dynamic monitoring of PV land use.

[0004] Satellite remote sensing is the most effective means of dynamically monitoring large-scale photovoltaic land use. While there have been numerous successful cases of identifying large-scale photovoltaic land use from satellite or airborne remote sensing imagery using traditional machine learning or deep learning methods, current research lacks a global, refined sample set that considers underlying land use / cover background type, photovoltaic array installation inclination, and geographic location. Furthermore, the lack of incorporation of domain knowledge makes it difficult to achieve both rapid and accurate identification of large-scale PVL monitoring. Therefore, based on the construction of a globally representative regional sample set, there is an urgent need to develop a method for rapid and precise identification of large-scale photovoltaic land use using open-source satellite imagery with the ability to migrate over time and space. Summary of the Invention

[0005] The purpose of the present invention is to solve the problem in the prior art that it is difficult to achieve both rapidity and accuracy in remote sensing classification of large-area photovoltaic land, and to provide a large-area photovoltaic land classification method and system that combines decision trees and deep learning.

[0006] In order to achieve the above-mentioned object of the invention, the present invention specifically adopts the following technical solutions:

[0007] In a first aspect, the present invention provides a large-area photovoltaic land classification method combining decision trees and deep learning, which comprises the following steps:

[0008] S1. Select data of representative photovoltaic land areas from the global photovoltaic land location dataset and classification dataset, and construct a photovoltaic land sample dataset containing remote sensing images, geographic coordinates, imaging time, underlying land cover type, and photovoltaic array arrangement;

[0009] S2. Input the photovoltaic land sample dataset into the constructed decision tree classifier for coarse classification, and obtain the coarse segmentation result in the form of a binary raster map;

[0010] S3. Post-processing the coarse segmentation results, setting a buffer zone of a preset size around the post-processed coarse segmentation results, defining an outer envelope rectangle outside the buffer zone, and forming a coarse classification area image of the photovoltaic land from pixels within the outer envelope rectangle in the remote sensing image;

[0011] S4. Generate a true photovoltaic land label for each photovoltaic land coarse classification area image through human-computer interaction, and perform a slicing operation on the photovoltaic land coarse classification area image and the photovoltaic land true label. The sliced ​​photovoltaic land coarse classification area image and the photovoltaic land true label constitute a photovoltaic land classification dataset;

[0012] S5. Train a photovoltaic land classification model that integrates multi-scale photovoltaic land features on the photovoltaic land classification dataset;

[0013] S6. Input the coarse classification area image of the photovoltaic land to be classified into the trained photovoltaic land classification model to obtain the segmentation result of the coarse classification area image of the photovoltaic land, thereby realizing fine identification of the photovoltaic land.

[0014] Based on the above solution, each step can be implemented in the following preferred specific manner.

[0015] As a preferred embodiment of the above-mentioned first aspect, in step S2, the classification limitation conditions of the decision tree classifier are divided into photovoltaic land index conditions and other conditions. The photovoltaic land index conditions are the first photovoltaic land index, the second photovoltaic land index and the blue-red band ratio. The other conditions are slope, altitude, normalized building index and reflectivity of the first short-wave infrared band.

[0016] As a preferred embodiment of the first aspect, the first photovoltaic land index PVLI1, the second photovoltaic land index PVLI2, the blue-red band ratio BDRI, and the normalized building index NDBI are specifically calculated as follows:

[0017]

[0018]

[0019] Among them, ρ Blue ,ρ Red ,ρ NIR ,ρSWIR1 ,ρ SWIR2 They represent the reflectivity of the blue band, red band, near-infrared band, first short-wave infrared band, and second short-wave infrared band respectively.

[0020] As a preference for the first aspect above, in step S2, the basis for coarse classification of the photovoltaic land sample data set is set as: simultaneously satisfying the first photovoltaic land index PVLI1∈[0.75,3.65], the second photovoltaic land index PVLI1∈[0.30,3.60], the blue-red band ratio BDRI∈[0.56,1.21], the slope SLOPE∈[0,20°], the altitude ALTITUDE∈[0,3500m], the normalized building index NDBI∈[-0.03,0.46] and the reflectivity ρ of the first short-wave infrared band SWIR1 ∈[0.07,0.58].

[0021] As a preferred embodiment of the above-mentioned first aspect, in step S3, the specific process of post-processing the coarse segmentation result is: first, the coarse segmentation result is subjected to spatial mean filtering, then the spatial mean filtering result is subjected to reclassification and denoising, and then the reclassification and denoising result in the form of a binary raster image is vectorized to obtain a vectorized binary image, and finally, all elements in the vectorized binary image whose area is larger than a preset area threshold are removed to obtain a post-processed coarse segmentation result.

[0022] As a preferred embodiment of the above-mentioned first aspect, the photovoltaic land classification model is a main-sub dual-branch structure, the main branch includes an input module, an encoder, a void space pyramid pooling module, a decoder and an output module, and the sub-branch includes an input module, an encoder and five exponential space attention modules, wherein the encoder of the main branch, the encoder of the sub-branch and the decoder of the main branch each contain four residual blocks, the main branch is a U-shaped symmetrical structure, each residual block of the main branch decoder is jump-connected with each residual block at the same level of the main branch encoder, and the input and output of each residual block of the sub-branch encoder are densely connected. The exponential feature space weights generated by each exponential space attention module in the sub-branch are densely connected with each residual block at the same level of the main branch encoder. The exponential feature space weights generated by each exponential space attention module in the sub-branch are densely connected with the exponential feature space weights generated by the subsequent exponential space attention module, and then densely connected with each residual block at the same level of the main branch decoder.

[0023] As a preference of the above-mentioned first aspect, in the photovoltaic land classification model of step S5, the specific processing flow is: based on the image of the coarse classification area of ​​the photovoltaic land after slicing, a multispectral image and an index image are generated, the multispectral image is input into the input module of the main branch to obtain a first spectral feature map, the index image is input into the input module of the sub-branch to obtain a first index feature map, the first index feature map is input into the first index space attention module of the sub-branch to obtain a first index feature space weight matrix, the first index feature space weight matrix and the first spectral feature map are multiplied to obtain a first fusion feature map; the first fusion feature map is input into the first residual block of the main branch encoder to obtain a second spectral feature map, and the first index is multiplied into the first spectral feature map. The feature map is input into the first residual block of the sub-branch encoder to obtain a second exponential feature map, the first exponential feature map and the second exponential feature map are densely connected to obtain a first densely connected feature map, the first densely connected feature map is input into the second exponential space attention module of the sub-branch to obtain a second exponential feature space weight matrix, the second exponential feature space weight matrix and the second spectral feature map are multiplied to obtain a second fused feature map; the second fused feature map is input into the second residual block of the main branch encoder to obtain a third spectral feature map, the first densely connected feature map is input into the second residual block of the sub-branch encoder to obtain a third exponential feature map, the third exponential feature map and the first densely connected feature map are multiplied to obtain a third spectral feature map. The rows are densely connected to obtain a second densely connected feature map, and the second densely connected feature map is input into the third exponential space attention module of the sub-branch to obtain a third exponential feature space weight matrix. The third exponential feature space weight matrix and the third spectral feature map are multiplied to obtain a third fusion feature map; the third fusion feature map is input into the third residual block of the main branch encoder to obtain a fourth spectral feature map, and the second densely connected feature map is input into the third residual block of the sub-branch encoder to obtain a fourth exponential feature map. The fourth exponential feature map and the second densely connected feature map are densely connected to obtain a third densely connected feature map. The third densely connected feature map is input into the fourth exponential space attention module of the sub-branch. Obtain a fourth exponential feature space weight matrix, multiply the fourth exponential feature space weight matrix and the fourth spectral feature map to obtain a fourth fused feature map; input the fourth fused feature map into the fourth residual block of the main branch encoder to obtain a fifth spectral feature map, input the third densely connected feature map into the fourth residual block of the sub-branch encoder to obtain a fifth exponential feature map, densely connect the fifth exponential feature map and the third densely connected feature map to obtain a fourth densely connected feature map, input the fourth densely connected feature map into the fifth exponential space attention module of the sub-branch to obtain a fifth exponential feature space weight matrix, and multiply the fifth exponential feature space weight matrix and the fifth spectral feature map to obtain a fifth fused feature map;The fifth fused feature map is input into the dilated space pyramid pooling module to obtain the dilated space pooling feature map, the dilated space pooling feature map is multiplied by the fifth exponential feature space weight matrix to obtain the sixth fused feature map, the sixth fused feature map is input into the fourth residual block of the main branch decoder to obtain the sixth spectral feature map, the fifth exponential feature space weight matrix and the fourth exponential feature space weight matrix are densely connected to obtain the first dense connection weight matrix, the sixth spectral feature map is multiplied by the first dense connection weight matrix to obtain the seventh fused feature map; the seventh fused feature map is input into the third residual block of the main branch decoder to obtain the seventh spectral feature map, the first dense connection weight matrix and the third exponential feature space weight matrix are densely connected to obtain the second dense connection weight matrix, the seventh spectral feature map and the second dense connection weight matrix are densely connected. The eighth fused feature map is obtained by multiplying the weight matrix; the eighth fused feature map is input into the second residual block of the main branch decoder to obtain an eighth spectral feature map, the second dense connection weight matrix and the second exponential feature space weight matrix are densely connected to obtain a third dense connection weight matrix, and the eighth spectral feature map and the third dense connection weight matrix are multiplied to obtain a ninth fused feature map; the ninth fused feature map is input into the first residual block of the main branch decoder to obtain a ninth spectral feature map, the third dense connection weight matrix and the first exponential feature space weight matrix are densely connected to obtain a fourth dense connection weight matrix, and the ninth spectral feature map and the fourth dense connection weight matrix are multiplied to obtain a tenth fused feature map; the tenth fused feature map is input into the output module of the main branch to obtain the segmentation result output by the photovoltaic land classification model.

[0024] As a preferred embodiment of the first aspect, the multispectral image is obtained in the following manner:

[0025] In the coarse classification area image of the photovoltaic land after slicing, the images of the blue band, green band, red band and the first short-wave infrared band are synthesized to form a multispectral image.

[0026] As a preferred embodiment of the first aspect, the index image is obtained in the following manner:

[0027] The red band, near-infrared band, first short-wave infrared band and second short-wave infrared band images are taken from the sliced ​​photovoltaic land coarse classification area image and calculated according to the calculation formula of the first photovoltaic land index to obtain the first photovoltaic land index image. The blue band, red band, near-infrared band, first short-wave infrared band and second short-wave infrared band images are taken from the sliced ​​photovoltaic land coarse classification area image and calculated according to the calculation formula of the second photovoltaic land index to obtain the second photovoltaic land index image. The first photovoltaic land index image and the second photovoltaic land index image are band-synthesized to form an index image.

[0028] In a second aspect, the present invention provides a large-area photovoltaic land classification system combining decision trees and deep learning, comprising:

[0029] The data acquisition module is used to select data of representative photovoltaic land areas from the global photovoltaic land location dataset and classification dataset, and construct a photovoltaic land sample dataset that includes remote sensing images, geographic coordinates, imaging time, underlying land cover type, and photovoltaic array arrangement;

[0030] The coarse classification module is used to input the photovoltaic land sample data set into the constructed decision tree classifier for coarse classification, and obtain the coarse segmentation result in the form of a binary raster map;

[0031] A post-processing module is used to post-process the coarse segmentation results, set a buffer zone of a preset size around the post-processed coarse segmentation results, frame an outer envelope rectangle outside the buffer zone, and form a coarse classification area image of the photovoltaic land from the pixels within the outer envelope rectangle in the remote sensing image;

[0032] The dataset construction module is used to generate a true photovoltaic land label for each photovoltaic land coarse classification area image through human-computer interaction, and to slice the photovoltaic land coarse classification area image and the photovoltaic land true label. The sliced ​​photovoltaic land coarse classification area image and the photovoltaic land true label constitute the photovoltaic land classification dataset;

[0033] Model training module, used to train a photovoltaic land classification model that integrates multi-scale photovoltaic land features on the photovoltaic land classification dataset;

[0034] The result acquisition module is used to input the coarse classification area image of the photovoltaic land to be classified into the trained photovoltaic land classification model, obtain the segmentation result of the coarse classification area image of the photovoltaic land, and realize the fine identification of photovoltaic land.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] To address the difficulty of balancing speed and accuracy in remote sensing classification of large-scale photovoltaic land, this paper proposes a large-scale photovoltaic land classification method that combines decision trees and deep learning, based on a globally representative dataset of photovoltaic land samples. Based on the coarse classification results of photovoltaic land obtained using a decision tree classifier, this method combines four post-processing techniques: spatial mean filtering, reclassification denoising, area threshold denoising, and grid-vector conversion to correct the coarse classification results. The coarse classification area of ​​the PVL is then delineated by creating a buffer zone and an outer envelope rectangle, significantly reducing the amount of data required for deep learning classification.

[0037] The present invention also constructs a photovoltaic land classification model PVLI-FSSM that integrates the photovoltaic land index. After training and hyperparameter optimization, the model achieves higher segmentation accuracy than the existing deep learning model. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 Schematic diagram of the steps of the method of the present invention;

[0039] Figure 2 Schematic diagram of the framework of the method of the present invention;

[0040] Figure 3 This is a schematic diagram of the structure of the photovoltaic land classification model of the present invention;

[0041] Figure 4 This is a system block diagram of the present invention. DETAILED DESCRIPTION

[0042] In order to make the above-mentioned objects, features and advantages of the present invention more clearly understood, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings. In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways than those described herein, and those skilled in the art can make similar improvements without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below. The technical features in the various embodiments of the present invention can be combined accordingly without conflicting with each other.

[0043] In the description of the present invention, it should be understood that the terms "first" and "second" are used solely for descriptive purposes and are not to be construed as indicating or implying relative importance or implicitly specifying the number of technical features being described. Therefore, features defined as "first" or "second" may explicitly or implicitly include at least one of such features.

[0044] like Figure 1 As shown, in a preferred implementation of the present invention, the above-mentioned large-area photovoltaic land classification method combining decision tree and deep learning includes the following steps S1 to S6. The specific implementation process is described in detail below.

[0045] S1. Data of representative photovoltaic land (PVL) areas are selected from the global photovoltaic land location dataset and classification dataset to construct a PV land sample dataset containing remote sensing images, geographic coordinates, imaging time, underlying land cover type, and photovoltaic array arrangement.

[0046] It should be noted that the global photovoltaic land location dataset mentioned in the present invention refers to the various points where photovoltaic land is actually measured. Such point information mainly comes from the installation statistics of the National Energy Administration of each country and the measured data published by various scientific research institutions. Each measured point will save four pieces of information when recording: the main underlying surface land cover type, the photovoltaic array arrangement, and the corresponding geographic coordinates and sampling time. According to the geographic coordinates and sampling time information, the corresponding remote sensing image can be collected, and the above information can be attached to the remote sensing image to construct it. In addition, the above-mentioned classification dataset of global photovoltaic land can adopt the existing global photovoltaic land dataset or the data collected by itself. For example, the underlying surface land cover type, photovoltaic array arrangement, and the corresponding geographic coordinates and sampling time of photovoltaic land in different regions around the world can directly adopt the Global Photovoltaic Land Vector Dataset (GPVI). GPVI is the core global photovoltaic land vector data product for geographic information science recognized by the Nature journal. It contains the identification results of global photovoltaic land from 2016 to 2019, and the positioning accuracy exceeds 85%. The underlying land cover type, photovoltaic array arrangement, and corresponding geographic coordinates and sampling time data of photovoltaic land in different regions around the world can be extracted from GPVI data.

[0047] It should also be noted that since the method of the present invention requires a large amount of global photovoltaic land use data to maximize the versatility of the photovoltaic land use classification model, and there is spatial heterogeneity in different regions of the world, the photovoltaic land use sample dataset should also cover the global scope as much as possible. In actual use, if only large-scale photovoltaic land use is to be quickly and accurately identified in a specific area, the user can also collect remote sensing images of the specified area for training the photovoltaic land use classification model, further reducing the amount of data and achieving a better fitting effect of the photovoltaic land use classification model in the specified area.

[0048] As a preferred implementation of an embodiment of the present invention, this embodiment uses Sentinel-2 remote sensing imagery as a data source and selects 236 representative photovoltaic land areas from six continents: Asia, Europe, Africa, North America, South America, and Oceania. These remote sensing images cover different underlying land cover types and photovoltaic array arrangements on each continent. These remote sensing images form a photovoltaic land sample dataset. The constructed photovoltaic land sample dataset is diverse and large enough to meet the training requirements of the photovoltaic land classification model and maximize the model's versatility.

[0049] S2. Input the photovoltaic land sample dataset into the constructed decision tree classifier for coarse classification, and obtain the coarse segmentation result in the form of a binary raster map.

[0050] In step S2 of the present invention, to enhance the photovoltaic land classification model's ability to fit photovoltaic land in various surface environments, as much training data as possible is required. However, the current photovoltaic land data is generally of low accuracy and insufficient global coverage density. Using only existing photovoltaic land data to train the photovoltaic land classification model will result in insufficient accuracy and versatility in its fitting results. Therefore, it is necessary to collect remote sensing images from different regions to construct a globally distributed fine sample. These images are not only massive in size, but in most images, the photovoltaic land portion only accounts for a small portion of the image. Therefore, by using a decision tree classifier to perform coarse classification on these remote sensing images and retaining the key portions containing photovoltaic land, the amount of data processing can be minimized while ensuring a sufficiently rich training sample.

[0051] It should be noted that in the present invention, the structure and classification method of the decision tree classifier belong to the prior art. It is a tree structure (which can be a binary tree or a non-binary tree). Each non-leaf node represents a test on a feature attribute, each branch represents the output of this feature attribute in a certain value range, and each leaf node stores a category. The process of making decisions using a decision tree classifier is to start from the root node, test the corresponding feature attributes in the item to be classified, and select the output branch according to its value until it reaches the leaf node, and use the category stored in the leaf node as the decision result.

[0052] The classification and limitation conditions of the decision tree classifier proposed in the present invention can be categorized into two types: photovoltaic land index conditions and other conditions. The photovoltaic land index conditions include the first photovoltaic land index, the second photovoltaic land index, and the blue-red band ratio. The blue-red band ratio refers to the ratio of the reflectivity of the blue band to the reflectivity of the red band in the remote sensing image. The photovoltaic land index conditions are calculated with the purpose of highlighting the commonality of the spectral characteristics of photovoltaic land in remote sensing images across the world in multiple bands. The calculation involves the reflectivity of five bands, namely the blue band, the red band, the near-infrared band, the first short-wave infrared band, and the second short-wave infrared band. The specific expression is:

[0053]

[0054]

[0055] Among them, ρ Blue ,ρ Red ,ρ NIR ,ρ SWIR1 ,ρ SWIR2 They represent the reflectivity of the blue band, red band, near-infrared band, first short-wave infrared band, and second short-wave infrared band respectively.

[0056] In this embodiment, the 458nm~523nm band is used as the blue band, and its center wavelength is 490nm; the 650nm~680nm band is used as the red band, and its center wavelength is 665nm; the 785nm~900nm band is used as the near-infrared band, and its center wavelength is 842nm; the 1565nm~1655nm band is used as the first short-wave infrared band, and its center wavelength is 1610nm; the 2100nm~2280nm band is used as the second short-wave infrared band, and its center wavelength is 2190nm.

[0057] Other conditions for the decision tree classifier include slope, altitude, Normalized Difference Building Index (NDBI), and reflectance in the first shortwave infrared band. Their primary purpose is to distinguish photovoltaic land from other features that share similarities in imagery. By setting an upper limit for slope, photovoltaic land can be distinguished from mountain shadows; by setting an upper limit for altitude, misclassification in some mountainous areas can be eliminated; by using the Normalized Difference Building Index (NDBI), photovoltaic land can be distinguished from densely built-up areas; and by using reflectance in the first shortwave infrared band, photovoltaic land can be distinguished from dark water bodies.

[0058] Among the other conditions mentioned above, the calculation of slope and altitude requires elevation data, which can be obtained from the global digital elevation model (DEM). DEM is a digital simulation of ground terrain using limited terrain elevation data. Currently, there are relatively mature global DEM data such as the Shuttle Radar Topography Mission (SRTM) and the Advanced Spaceborne Thermal Emission and Reflection Radiometer Global Digital Elevation Model (ASTER GDEM). The Normalized Difference Building Index (NDBI) and the reflectivity of the first shortwave infrared band can be directly obtained from remote sensing images. Among them, the calculation formula of the Normalized Building Index is:

[0059]

[0060] As a preferred implementation method of an embodiment of the present invention, the basis for the decision tree coarse classification of the photovoltaic land sample data set is set as follows: simultaneously satisfying the first photovoltaic land index PVLI1∈[0.75,3.65], the second photovoltaic land index PVLI1∈[0.30,3.60], the blue-red band ratio BDRI∈[0.56,1.21], the slope SLOPE∈[0,20°], the altitude ALTITUDE∈[0,3500m], the normalized building index NDBI∈[-0.03,0.46] and the reflectivity ρ of the first short-wave infrared band SWIR1 ∈[0.07,0.58].

[0061] S3. Post-process the coarse segmentation results, set a buffer zone of a preset size around the post-processed coarse segmentation results, frame an outer envelope rectangle outside the buffer zone, and form a coarse classification area image of the photovoltaic land by the pixels inside the outer envelope rectangle in the remote sensing image.

[0062] It should be noted that in step S3, the rough segmentation result is corrected through four post-processing techniques: spatial mean filtering, reclassification denoising, area threshold denoising, and grid-vector conversion. The specific steps are as follows:

[0063] S31. First, perform spatial mean filtering on the coarse segmentation result.

[0064] In this embodiment, based on a preset domain range n, a sliding window of size (2n+1)×(2n+1) and all values ​​1 is used to perform mean convolution on the pixels in the coarse segmentation result. The new value of the center pixel is output as 1 only when all other pixels in the designated neighborhood of the center pixel are 1; otherwise, the output value of the center pixel will fall within the interval (0,1). The spatial mean filtering result retains the values ​​of non-edge pixels and weakens the values ​​of pixels at broken edges or edge jagged protrusions.

[0065] S32. Then, the spatial mean filtering result is subjected to reclassification and denoising processing, that is, by retaining pixels with a value of 1 and assigning a value of 0 to pixels with a value divided by 1, so as to achieve the purpose of removing free small noise and smoothing the edges of the coarsely classified photovoltaic land.

[0066] S33. Secondly, the reclassification and denoising results in the form of a binary raster image are vectorized to obtain a vectorized binary image.

[0067] S34. Finally, all elements in the vectorized binary image whose area is larger than a preset area threshold are removed (i.e., misclassified noise with a large area that still exists after spatial mean filtering and classification denoising is removed), and a coarse segmentation result after post-processing is obtained.

[0068] The purpose of post-processing the coarse classification results is to smooth the edges of the coarse segmentation results obtained by the decision tree classifier and to repair any incompleteness caused by misclassification or omission. The final result is converted into vector data, which also makes the coarse classification results of photovoltaic land use more independent and easier to edit.

[0069] It should also be noted that, in the present invention, considering that the coarse classification method often cannot completely extract the entire photovoltaic land, or the recognition accuracy is insufficient, it is necessary to delineate a complete area with photovoltaic land around the coarse segmentation result, and input the corresponding remote sensing image into the constructed neural network model for training, so as to obtain a model with better photovoltaic land recognition effect. In this embodiment, after the coarse segmentation result is post-processed, through the overall investigation of the photovoltaic land sample data set, a size of 460m (slightly larger than ) buffer zone and framing an outer envelope rectangle outside this buffer zone can minimize missed points and ensure the integrity of the photovoltaic land in the input, which is consistent with the goal of the present invention to maximize the use of key parts of photovoltaic land in the image with minimum data processing.

[0070] As a preferred implementation of the embodiment of the present invention, the domain range of the spatial mean filter is set to 3 pixels, and the area threshold value of the area threshold denoising is set to 2000m 2 , and the buffer size is set to 460m. The final PV land coarse classification area image is generated in this way, which has the best effect for the overall PV land sample dataset from 236 sample areas.

[0071] S4. Generate high-precision photovoltaic land real labels for each photovoltaic land coarse classification area image through human-computer interaction, and slice the photovoltaic land coarse classification area image and the photovoltaic land real labels. The sliced ​​photovoltaic land coarse classification area images and the photovoltaic land real labels constitute the photovoltaic land classification dataset.

[0072] It should be noted that in step S4 of the present invention, the real labels of photovoltaic land can be generated with the help of tools such as Labelme and ArcGIS Pro. These tools currently provide open source access, and the generated real labels of photovoltaic land will automatically mark unique values ​​according to the type of land feature, which is convenient for training the photovoltaic land classification model. At the same time, in order to save workload, some photovoltaic land elements with better fitting effects in the photovoltaic land sample data set (GPVI) can be directly migrated to the photovoltaic land classification data set. After completing the production of the above-mentioned real labels, it is necessary to slice the coarse classification area image of photovoltaic land and the real label of photovoltaic land. The photovoltaic land classification model accepts square-shaped photovoltaic land coarse classification area image slices and photovoltaic land real label slices as input.

[0073] Furthermore, the photovoltaic land classification dataset can be divided into a training set and a test set in a ratio of 8:2 for training and testing the photovoltaic land classification model.

[0074] As a preferred implementation of the present invention, by statistically analyzing the characteristics of photovoltaic land area in 236 sample areas around the world, it is determined that the average area of ​​a single photovoltaic land patch is about 0.07 km 2 Therefore, the size of the image slice of the photovoltaic land coarse classification area is determined to be 64 pixels × 64 pixels (about 0.4 km 2 ), so that an average-sized photovoltaic land patch can occupy about 1 / 6 of the slice area, thereby ensuring a certain degree of balance in the proportion of photovoltaic land and non-photovoltaic land in the slice.

[0075] S5. Train a photovoltaic land classification model that integrates multi-scale photovoltaic land features on the photovoltaic land classification dataset.

[0076] In the present invention, the photovoltaic land classification model is a main-sub dual-branch structure, the main branch includes an input module, an encoder, a void space pyramid pooling module, a decoder and an output module, the sub-branch includes an input module, an encoder and five exponential space attention modules, wherein the encoder of the main branch, the encoder of the sub-branch and the decoder of the main branch each contain four residual blocks, the main branch is a U-shaped symmetrical structure, each residual block of the main branch decoder is jump-connected with each residual block at the same level of the main branch encoder, the input and output of each residual block of the sub-branch encoder are densely connected, the exponential feature space weights generated by each exponential space attention module in the sub-branch are densely connected with each residual block at the same level of the main branch encoder, the exponential feature space weights generated by each exponential space attention module in the sub-branch are densely connected with the exponential feature space weights generated by the subsequent exponential space attention module, and then densely connected with each residual block at the same level of the main branch decoder.

[0077] It should be noted that in the photovoltaic land classification model in step S5, Figure 3 As shown, the specific processing flow is as follows: a multispectral image and an index image are generated based on the coarse classification area image of the photovoltaic land after slicing, the multispectral image is input into the input module of the main branch to obtain a first spectral feature map, the index image is input into the input module of the secondary branch to obtain a first index feature map, the first index feature map is input into the first index space attention module of the secondary branch to obtain a first index feature space weight matrix, and the first index feature space weight matrix and the first spectral feature map are multiplied to obtain a first fusion feature map;

[0078] Input the first fused feature map into the first residual block of the main branch encoder to obtain a second spectral feature map, input the first exponential feature map into the first residual block of the sub-branch encoder to obtain a second exponential feature map, densely connect the first exponential feature map and the second exponential feature map to obtain a first densely connected feature map, input the first densely connected feature map into the second exponential space attention module of the sub-branch to obtain a second exponential feature space weight matrix, and multiply the second exponential feature space weight matrix and the second spectral feature map to obtain a second fused feature map;

[0079] Input the second fused feature map into the second residual block of the main branch encoder to obtain a third spectral feature map, input the first densely connected feature map into the second residual block of the sub-branch encoder to obtain a third exponential feature map, densely connect the third exponential feature map and the first densely connected feature map to obtain a second densely connected feature map, input the second densely connected feature map into the third exponential space attention module of the sub-branch to obtain a third exponential feature space weight matrix, and multiply the third exponential feature space weight matrix and the third spectral feature map to obtain a third fused feature map;

[0080] Input the third fused feature map into the third residual block of the main branch encoder to obtain a fourth spectral feature map, input the second densely connected feature map into the third residual block of the sub-branch encoder to obtain a fourth exponential feature map, densely connect the fourth exponential feature map and the second densely connected feature map to obtain a third densely connected feature map, input the third densely connected feature map into the fourth exponential space attention module of the sub-branch to obtain a fourth exponential feature space weight matrix, and multiply the fourth exponential feature space weight matrix and the fourth spectral feature map to obtain a fourth fused feature map;

[0081] Inputting the fourth fused feature map into the fourth residual block of the main branch encoder to obtain a fifth spectral feature map, inputting the third densely connected feature map into the fourth residual block of the sub-branch encoder to obtain a fifth exponential feature map, densely connecting the fifth exponential feature map and the third densely connected feature map to obtain a fourth densely connected feature map, inputting the fourth densely connected feature map into the fifth exponential space attention module of the sub-branch to obtain a fifth exponential feature space weight matrix, and multiplying the fifth exponential feature space weight matrix and the fifth spectral feature map to obtain a fifth fused feature map;

[0082] Input the fifth fused feature map into the dilute spatial pyramid pooling module to obtain a dilute spatial pooling feature map, multiply the dilute spatial pooling feature map and the fifth exponential feature space weight matrix to obtain a sixth fused feature map, input the sixth fused feature map into the fourth residual block of the main branch decoder to obtain a sixth spectral feature map, densely connect the fifth exponential feature space weight matrix and the fourth exponential feature space weight matrix to obtain a first dense connection weight matrix, and multiply the sixth spectral feature map and the first dense connection weight matrix to obtain a seventh fused feature map;

[0083] Input the seventh fused feature map into the third residual block of the main branch decoder to obtain a seventh spectral feature map, densely connect the first dense connection weight matrix and the third exponential feature space weight matrix to obtain a second dense connection weight matrix, and multiply the seventh spectral feature map and the second dense connection weight matrix to obtain an eighth fused feature map;

[0084] Input the eighth fused feature map into the second residual block of the main branch decoder to obtain an eighth spectral feature map, densely connect the second dense connection weight matrix and the second exponential feature space weight matrix to obtain a third dense connection weight matrix, and multiply the eighth spectral feature map and the third dense connection weight matrix to obtain a ninth fused feature map;

[0085] Inputting the ninth fused feature map into the first residual block of the main branch decoder to obtain a ninth spectral feature map, densely connecting the third dense connection weight matrix and the first exponential feature space weight matrix to obtain a fourth dense connection weight matrix, and multiplying the ninth spectral feature map and the fourth dense connection weight matrix to obtain a tenth fused feature map;

[0086] Finally, the tenth fusion feature map is input into the output module of the main branch to obtain the segmentation result output by the photovoltaic land classification model.

[0087] It should be noted that, for the purpose of accurately identifying photovoltaic land in remote sensing images, this paper considers the multi-band characteristics of remote sensing images, which are more complex than natural images, and the index features derived from spectral band characteristics based on domain knowledge (and such features have been shown to help improve segmentation accuracy). A photovoltaic land classification model (PVLI-FSSM) is constructed. The ResUNet++ model is used as the baseline model, and its structure is improved accordingly based on the rich spectral information and index features of remote sensing images. The specific performance is as follows:

[0088] 1) Dual-branch structure: The photovoltaic land classification model integrates the multi-scale photovoltaic land index features into the spectral information of the main branch through the sub-branch, the exponential spatial attention module, and various connection methods between the main and sub-branch modules. The main branch is a ResUNet++ structure, which is responsible for extracting the multi-scale and contextual features of the spectral channels of the input image, and at the same time integrates the exponential features in the sub-branch into the spectral information through various connection methods. The main branch includes an input module, an encoder, an atrous spatial pyramid pooling module (ASPP), a decoder, and an output module. The sub-branch includes an input module, an encoder, and five exponential spatial attention modules (IFSA). The encoder of the main branch, the encoder of the sub-branch, and the decoder of the main branch each contain four residual blocks (ResBlock). Each residual block of the main branch decoder is jump-connected to the residual blocks of the same level in the main branch encoder. The output module is responsible for converting the output of the decoder into a segmentation result.

[0089] 2) Special connection methods: In order to maximize the use of the index and achieve multi-scale integration with spectral characteristics, in addition to the common connection, four special connection methods are designed in the photovoltaic land classification model:

[0090] 2.1) Dense connections are made between the input and output of each residual block of the secondary branch encoder;

[0091] 2.2) The exponential feature space weights generated by each exponential spatial attention module in the secondary branch are densely connected with the residual blocks ResBlock at the same level of the main branch encoder;

[0092] 2.3) The exponential feature space weight matrix (IFSW) generated by each exponential spatial attention module in the secondary branch is densely connected with each residual block at the same level of the main branch decoder;

[0093] 2.4) Before the exponential spatial attention module in the secondary branch is densely connected to the main branch decoder, the exponential feature space weight matrix generated by each exponential spatial attention module in the secondary branch is densely connected with the exponential feature space weight matrix generated by the next exponential spatial attention module.

[0094] Among them, the third and fourth connections are designed to highlight the foreground, suppress the background, and refine the boundaries.

[0095] 3) Atrous Spatial Pyramid Pooling (ASPP): In the connection between the main branch encoder and decoder, the photovoltaic land classification model uses the atrous spatial pyramid pooling module to replace the connection layer of ResUNet.

[0096] 4) Exponential Spatial Attention Module (IFSA): The spatial attention mechanism aims to improve the recognition of specific targets in remote sensing imagery. It can highlight multiple target features through operations such as multiplication of multiple target features. In neural networks, spatial attention captures and highlights features. Through various statistical modules or convolutional combination modules, it reduces the dimensionality and calculates the spatial weights of certain features to analyze the spatial importance of these features. The output feature spatial weights are then combined with other features to enhance the specific target features. This process can also be understood as a fusion process.

[0097] The photovoltaic land classification model proposed in this invention introduces an exponential spatial attention module based on spatial attention in the secondary branch. The input of the first exponential spatial attention module is the index image of the C layer. After passing through the first exponential spatial attention module, the first exponential feature space weight matrix IFSW is output and input into the encoder and decoder of the main branch. It is then fused with the spectral feature maps of each scale output in the main branch to highlight the characteristics of the photovoltaic land.

[0098] In the exponential spatial attention module, its input passes through the ASPP module, the Max module (calculating the cross-channel maximum), and the Mean module (calculating the cross-channel mean) branches for further feature mining. These three branches reduce the dimension of the input exponential feature (IF) from N×C×H×W to N×1×H×W. The outputs of the three branches are then concatenated to form an N×3×H×W feature map. This map is then added to the input feature map using a residual connection. The map is then processed through a 3×3 convolutional layer and a sigmoid activation function to form a single layer with a value range of [0, 1], namely the exponential feature space weight matrix.

[0099] 5) Multi-channel input: The multispectral image channels input into the main branch of the photovoltaic land classification model are the blue band, green band, red band, and first short-wave infrared band. Here, the green band corresponds to the band range of 543nm to 578nm, with a central wavelength of 560nm. The index image channels input into the secondary branch are the first photovoltaic land index and the second photovoltaic land index. The bands involved in the index calculation are: blue band, red band, near-infrared band, first short-wave infrared band, and second short-wave infrared band. The calculation of all indices is completed before input into the secondary branch.

[0100] The above multispectral images and index images are obtained in the following manner:

[0101] A) In the coarsely classified area image of the photovoltaic land after slicing, the blue band, green band, red band, and first short-wave infrared band images are synthesized to form a multispectral image.

[0102] B) In the sliced ​​coarse classification area image of photovoltaic land, images of the red band, near-infrared band, first short-wave infrared band, and second short-wave infrared band are taken and calculated according to the calculation formula of the first photovoltaic land index to obtain a first photovoltaic land index image; in the sliced ​​coarse classification area image of photovoltaic land, images of the blue band, red band, near-infrared band, first short-wave infrared band, and second short-wave infrared band are taken and calculated according to the calculation formula of the second photovoltaic land index to obtain a second photovoltaic land index image; the first photovoltaic land index image and the second photovoltaic land index image are band-synthesized to form an index image.

[0103] To help those skilled in the art better understand the structure of the photovoltaic land classification model, the principle of the ResUNet++ model is now explained as follows.

[0104] The ResUNet++ model is a variant of the classic convolutional neural network (CNN)—UNet. As a deep learning model with a simple structure, clear principles, and high segmentation accuracy, UNet is commonly used in remote sensing imagery interpretation and recognition. UNet consists of three parts: an encoder, a connection layer, and a decoder. Both the encoder and decoder contain four convolutional modules, each consisting of two consecutive convolutional layers. However, the encoder and decoder modules use different post-convolutional sampling methods. Each convolutional module in the encoder is followed by a max pooling layer for downsampling, halving the dimensionality and extracting more abstract features layer by layer. Each convolutional module in the decoder is preceded by an upsampling layer, halving the number of features in each layer compared to the next (backward) layer, restoring the spatial dimension layer by layer. The advantage of UNet is that each encoder layer is replicated, pruned, and connected to the corresponding upsampling layer in the decoder, preserving shallow-level information and providing more detailed information for upsampling in the decoder. This avoids the problem of shallow-level features being reduced with increasing depth.

[0105] ResUNet++ replaces the convolutional module in the original UNet with the residual module ResBlock, and uses Skip Connection to achieve cross-layer connection in each residual block, so that the model has the advantages of both the residual connection of ResNet and the contextual information fusion of UNet.

[0106] Furthermore, the ResUNet++ model makes the following improvements to the UNet network architecture:

[0107] 1) Compression and Excitation (SE) module,The encoder part of the network contains four downsampling modules. Except for the first downsampling module, the remaining downsampling modules contain a residual module and a SE module (i.e., an SE module is added after each residual module).

[0108] 2) Attention module. The decoder part of the network contains four upsampling modules. Except for the last downsampling module, the remaining upsampling modules contain an Attention module, an upsampling layer and a residual module, that is, an Attention module is added before each upsampling layer.

[0109] It should be noted that in the present invention, the photovoltaic land classification model PVLI-FSSM requires the use of multiple preset hyperparameters during training, namely, training round E, batch size B, optimizer O, initial learning rate Lr, momentum coefficient M, and weight decay W.

[0110] As a preferred implementation method of an embodiment of the present invention, the hyperparameters when training the photovoltaic land classification model are set as follows: the training round E is set to 300 rounds, the batch size B is set to 4, the optimizer O is set to SGD, the initial learning rate Lr is set to 0.01, the momentum coefficient M is set to 0.9, and the weight decay W is set to 1e-4.

[0111] S6. Input the coarse classification area image of the photovoltaic land to be classified into the trained photovoltaic land classification model to obtain the segmentation result of the coarse classification area image of the photovoltaic land, thereby realizing fine identification of the photovoltaic land.

[0112] It should be noted that, although experiments have shown that images from other sensors can also successfully identify photovoltaic land after being classified by the photovoltaic land classification model, the final recognition effect is often better when the image types of the training and validation sets are consistent. Furthermore, the method of the present invention requires that the image have sufficient spectral information. Therefore, the various methods described above are based on optical remote sensing images. When selecting any image for photovoltaic land identification, it is necessary to ensure that its wavelength range is within the range of 380nm to 1648nm.

[0113] It should be noted that when processing the identification of photovoltaic land in a large area, the scale of image data input by the user into the photovoltaic land classification model may be large. At this time, the key parts containing photovoltaic land in the remote sensing image can be preliminarily extracted through the aforementioned S2-S3 coarse classification and post-processing process, and then the key parts are input into the photovoltaic land classification model to generate identification results, and then the identification results are imported into the full-area image. This can reduce the amount of data processing, improve the calculation speed, and obtain reliable and high-precision photovoltaic land identification results.

[0114] It should be noted that the segmentation results generated by the photovoltaic land classification model in step S6 are already highly accurate, but they can be slightly modified based on user needs. After removing a small number of misclassifications and omissions, the corrected segmentation results are converted from raster to vector and output as polygonal photovoltaic land product-level data. The purpose of converting the segmentation results from raster data to vector data is to facilitate numerical statistics and editing operations such as area divisions. Users can also select the appropriate data output format based on their actual needs.

[0115] In summary, if Figure 2 As shown, the present invention proposes a "two-step" joint decision tree and deep learning large-area photovoltaic land classification method, which specifically includes:

[0116] Step 1: Delineation of coarse PV land classification zones using a decision tree classifier. Using the GEE remote sensing cloud platform and Sentinel-2 remote sensing imagery, a decision tree classifier was constructed to obtain a coarse segmentation result in the form of a binary raster image. After spatial mean filtering, reclassification denoising, area threshold denoising, and raster-to-vector conversion, a coarse PV land classification zone image was generated.

[0117] Step 2: Accurately segment PV land within the coarse classification area using deep learning. Using the trained PV land classification model, PV land is accurately extracted from Sentinel-2 image slices within the coarse classification area. After manual verification, data correction, and raster-to-vector conversion, product-level PV land vector data is obtained.

[0118] The present invention will further illustrate the detailed implementation process of the large-area photovoltaic land classification method of the joint decision tree and deep learning shown in S1 to S6 above in a specific example below, so as to facilitate understanding of the essence of the present invention.

[0119] Example

[0120] In this embodiment, the Sentinel-2 remote sensing satellite image is used as the source data, and the specific implementation process of the large-area photovoltaic land classification method using a joint decision tree and deep learning is as described above and will not be repeated here.

[0121] This example first selects remote sensing images of representative regions based on the Global Photovoltaic Land Vector Dataset (GPVI). In this example, images of PV land construction areas from 236 representative regions across six continents, including Asia, Europe, Africa, North America, South America, and Oceania, are selected to form the Global Representative PV Land Remote Sensing Image Dataset. Because cloud obstruction in remote sensing data can affect the identification of image features, the Global Representative PV Land Remote Sensing Image Dataset must first be preprocessed. Images with cloud cover exceeding 20% ​​are removed during data acquisition and cloud removal is performed. Spectral indices and slope, among other factors, are then calculated for each image and input into a decision tree classifier. Based on pre-defined constraints, PV land and non-PV land are classified into binary raster images. The resulting image is then smoothed through four post-processing steps: spatial mean filtering, reclassification denoising, area threshold denoising, and vectorization. Finally, a buffer zone and outer envelope are generated around the resulting image. Based on the outer envelope, a small number of key areas containing PV land can be extracted from each image through a clipping operation to form a coarsely classified image block. In this embodiment, the above steps are all completed on the Google Earth Engine online cloud platform (code.earthengine.com). The decision tree defines the land as photovoltaic land as follows: PVLI1>1.142∩PVLI2>1.11∩BDRI>0.674. At the same time, the slope must be less than 20°, the altitude must be less than 3500m, the normalized building index (NDBI) must be greater than 0.075, and the reflectivity in the 1572nm~1648nm band must be greater than 0.194 (mainly used to distinguish water bodies). The elevation data required to calculate the slope and altitude are directly derived from the global digital elevation model (DEM) provided by the Shuttle Radar Topography Mission (SRTM) on the Google Earth Engine platform. In the post-processing technology, the domain range of the spatial mean filter is set to 3 pixels, and the area threshold set for area threshold denoising is 2000m. 2 , the surrounding buffer size is set to 460m.

[0122] After obtaining the image blocks to be classified in the key areas, ArcGIS Pro was used to interactively create features to generate high-precision PV land labels. The images and labels were then sliced ​​into 64×64 pixels according to the PVLI-FSSM model's requirements. Where some boundary image dimensions did not meet the slicing requirements, the surrounding areas were padded with zero values. This slicing process ultimately yielded 10,219 pairs of images and labels, which were divided into training and test sets in an 8:2 ratio. The training set, containing the images and corresponding labels, was fed into the PVLI-FSSM model for iterative training. After training, hyperparameters were adjusted based on the performance on the test set. The optimal preset hyperparameters were determined: 300 epochs of model training, a batch size of 4, an SGD optimizer, an initial learning rate of 0.01, a momentum coefficient of 0.9, and a weight decay of 1e-4. After training, the PVLI-FSSM model can accurately identify PV land in remote sensing imagery from around the world.

[0123] In this example, in order to verify the accuracy of the PVLI-FSSM model combined with spectral index features obtained by the above training, its performance was evaluated by the overall accuracy (OA) and mean intersection over union (mIoU) indicators. The results showed that OA = 97.54% and mIoU = 95.02%, indicating that the model has good fitting ability.

[0124] It can be seen from this that after training the photovoltaic land classification model, large-scale photovoltaic land identification can be performed for any area that needs to be identified. After raster-to-vector conversion operations, the final photovoltaic land feature class data with precise details and complete edges can be obtained.

[0125] In this example, large-scale PV land mapping was performed using Jiangsu Province and the Ningxia Hui Autonomous Region in China as representative examples. The results, shown in Tables 1 and 2, achieved an overall classification accuracy of 95.56%, reducing the deep learning data processing overhead to 4.5% of the original. This enabled rapid and accurate mapping of large-scale PV land. Based on provincial-level remote sensing imagery from different years, accurate PV land use product data was generated, and interannual data was used to understand the spatiotemporal variations in PV land use in the two provinces.

[0126] Table 1. Photovoltaic land identification accuracy in six typical inspection areas in Jiangsu

[0127]

[0128] Table 2. Photovoltaic land identification accuracy in six typical inspection areas in Ningxia

[0129]

[0130] To compare the impact of multi-scale fusion of spectral index features on the accuracy of the PVLI-FSSM model, an ablation experiment was conducted on a photovoltaic land dataset, comparing the PVLI-FSSM model with the existing baseline ResUNet++ model. After 100 rounds of training, the two models were trained and their segmentation performance was compared, with the results shown in Table 3. The results demonstrate that the PVLI-FSSM model, incorporating spectral index features, can better identify photovoltaic land in remote sensing imagery, and the introduction of spectral index features significantly improves accuracy, indicating that the multi-scale fusion of spectral index features can further distinguish photovoltaic land from the image background, thereby achieving accurate identification.

[0131] Table 3. Accuracy comparison between the baseline ResUNet++ model and the PVLI_FSSM model

[0132]

[0133] In summary, this invention, based on the rapid development of geographic remote sensing big data and deep learning algorithms, addresses the challenges of large-scale data and a lack of globally distributed training samples in large-scale photovoltaic land identification research. This method provides a large-scale photovoltaic land classification method that combines decision trees and deep learning, and constructs a representative global photovoltaic deep learning dataset. Simultaneously, leveraging deep learning methods' strong ability to characterize nonlinear relationships and discover deep features, this method constructs long-term, high-precision photovoltaic land product data, exemplified by Jiangsu and Ningxia. This demonstrates that this invention can improve understanding of the spatiotemporal differentiation and changing trends of global photovoltaic land use, providing a path for analyzing global energy infrastructure development.

[0134] It should also be noted that the large-area photovoltaic land classification method using the combined decision tree and deep learning in the above embodiment can essentially be executed by a computer program or module. Therefore, similarly, based on the same inventive concept, another preferred embodiment of the present invention also provides a large-area photovoltaic land classification system using the combined decision tree and deep learning corresponding to the large-area photovoltaic land classification method using the combined decision tree and deep learning provided in the above embodiment, such as Figure 4 As shown, it includes:

[0135] The data acquisition module is used to select data of representative photovoltaic land areas from the global photovoltaic land location dataset and classification dataset, and construct a photovoltaic land sample dataset that includes remote sensing images, geographic coordinates, imaging time, underlying land cover type, and photovoltaic array arrangement;

[0136] The coarse classification module is used to input the photovoltaic land sample data set into the constructed decision tree classifier for coarse classification, and obtain the coarse segmentation result in the form of a binary raster map;

[0137] A post-processing module is used to post-process the coarse segmentation results, set a buffer zone of a preset size around the post-processed coarse segmentation results, frame an outer envelope rectangle outside the buffer zone, and form a coarse classification area image of the photovoltaic land from the pixels within the outer envelope rectangle in the remote sensing image;

[0138] The dataset construction module is used to generate a true photovoltaic land label for each photovoltaic land coarse classification area image through human-computer interaction, and to slice the photovoltaic land coarse classification area image and the photovoltaic land true label. The sliced ​​photovoltaic land coarse classification area image and the photovoltaic land true label constitute the photovoltaic land classification dataset;

[0139] Model training module, used to train a photovoltaic land classification model that integrates multi-scale photovoltaic land features on the photovoltaic land classification dataset;

[0140] The result acquisition module is used to input the coarse classification area image of the photovoltaic land to be classified into the trained photovoltaic land classification model, obtain the segmentation result of the coarse classification area image of the photovoltaic land, and realize the fine identification of photovoltaic land.

[0141] It should also be noted that those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working process of the system described above can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here. In the various embodiments provided in this application, the division of steps or modules in the system and method is only a logical function division. In actual implementation, there may be other division methods, for example, multiple modules or steps can be combined or integrated together, and a module or step can also be split.

[0142] The embodiment described above is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Persons skilled in the art may make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, any technical solution obtained by equivalent substitution or equivalent transformation falls within the scope of protection of the present invention.

Claims

1. A large-area photovoltaic land classification method combining decision tree and deep learning, characterized in that: The following steps are involved: S1. Select data of representative photovoltaic land areas from the global photovoltaic land location dataset and classification dataset, and construct a photovoltaic land sample dataset containing remote sensing images, geographic coordinates, imaging time, underlying land cover type, and photovoltaic array arrangement; S2. Input the photovoltaic land sample dataset into the constructed decision tree classifier for coarse classification, and obtain the coarse segmentation result in the form of a binary raster map; S3. Post-processing the coarse segmentation results, setting a buffer zone of a preset size around the post-processed coarse segmentation results, defining an outer envelope rectangle outside the buffer zone, and forming a coarse classification area image of the photovoltaic land from pixels within the outer envelope rectangle in the remote sensing image; S4. Generate a true photovoltaic land label for each photovoltaic land coarse classification area image through human-computer interaction, and perform a slicing operation on the photovoltaic land coarse classification area image and the photovoltaic land true label. The sliced ​​photovoltaic land coarse classification area image and the photovoltaic land true label constitute a photovoltaic land classification dataset; S5. Train a photovoltaic land classification model that integrates multi-scale photovoltaic land features on the photovoltaic land classification dataset; S6. Inputting the coarse classification area image of the photovoltaic land to be classified into the trained photovoltaic land classification model to obtain the segmentation result of the coarse classification area image of the photovoltaic land, thereby realizing fine identification of the photovoltaic land; The photovoltaic land classification model has a main-sub dual-branch structure. The main branch includes an input module, an encoder, a dilated spatial pyramid pooling module, a decoder, and an output module. The sub-branch includes an input module, an encoder, and five exponential spatial attention modules. The encoder, encoder, and decoder of the main branch each contain four residual blocks. The main branch has a U-shaped symmetrical structure. Each residual block of the main branch decoder is jump-connected with each residual block of the same level of the main branch encoder. The input and output of each residual block of the sub-branch encoder are densely connected. The exponential feature space weight generated by each exponential spatial attention module in the sub-branch is densely connected with each residual block of the same level of the main branch encoder. The exponential feature space weight generated by each exponential spatial attention module in the sub-branch is densely connected with the exponential feature space weight generated by the subsequent exponential spatial attention module, and then densely connected with each residual block of the same level of the main branch decoder. The input module of the main branch is a multispectral image generated based on the coarse classification area image of the photovoltaic land after slicing. The input module of the sub-branch is an index image generated based on the coarse classification area image of the photovoltaic land after slicing.

2. The large-area photovoltaic land classification method combining decision tree and deep learning as claimed in claim 1, characterized in that: In step S2, the classification limitation conditions of the decision tree classifier are divided into photovoltaic land index conditions and other conditions. The photovoltaic land index conditions are the first photovoltaic land index, the second photovoltaic land index and the blue-red band ratio. The other conditions are slope, altitude, normalized building index and reflectivity of the first short-wave infrared band.

3. The large-area photovoltaic land classification method combining decision tree and deep learning as claimed in claim 2, characterized in that: The specific calculation methods of the first photovoltaic land index PVLI1, the second photovoltaic land index PVLI2, the blue-red band ratio BDRI, and the normalized building index NDBI are as follows: Among them, ρ Blue ,ρ Red ,ρ NIR ,ρ SWIR1 ,ρ SWIR2 They represent the reflectivity of the blue band, red band, near-infrared band, first short-wave infrared band, and second short-wave infrared band respectively.

4. A large-area photovoltaic land classification method combining decision tree and deep learning as claimed in claim 2 or 3, characterized in that: In step S2, the basis for coarse classification of the photovoltaic land sample dataset is set as follows: the first photovoltaic land index PVLI1∈[0.75,3.65], the second photovoltaic land index PVLI1∈[0.30,3.60], the blue-red band ratio BDRI∈[0.56,1.21], the slope SLOPE∈[0,20°], the altitude ALTITUDE∈[0,3500m], the normalized building index NDBI∈[-0.03,0.46] and the reflectivity ρ of the first short-wave infrared band are simultaneously satisfied. SWIR1 ∈[0.07,0.58].

5. The large-area photovoltaic land classification method combining decision tree and deep learning as claimed in claim 1, characterized in that: In step S3, the specific process of post-processing the coarse segmentation results is: first, the coarse segmentation results are subjected to spatial mean filtering, then the spatial mean filtering results are subjected to reclassification and denoising, and then the reclassification and denoising results in the form of a binary raster image are vectorized to obtain a vectorized binary image, and finally, all elements in the vectorized binary image whose area is larger than the preset area threshold are removed to obtain the post-processed coarse segmentation results.

6. The large-area photovoltaic land classification method combining decision tree and deep learning as claimed in claim 1, characterized in that: In the photovoltaic land classification model of step S5, the specific processing flow is: based on the rough classification area image of the photovoltaic land after slicing, a multispectral image and an index image are generated, the multispectral image is input into the input module of the main branch to obtain a first spectral feature map, the index image is input into the input module of the sub-branch to obtain a first index feature map, the first index feature map is input into the first index space attention module of the sub-branch to obtain a first index feature space weight matrix, the first index feature space weight matrix and the first spectral feature map are multiplied to obtain a first fusion feature map; the first fusion feature map is input into the first residual block of the main branch encoder to obtain a second spectral feature map, and the first index feature map is input into the sub-branch. In the first residual block of the encoder, a second exponential feature map is obtained, the first exponential feature map and the second exponential feature map are densely connected to obtain a first densely connected feature map, the first densely connected feature map is input into the second exponential space attention module of the sub-branch to obtain a second exponential feature space weight matrix, the second exponential feature space weight matrix and the second spectral feature map are multiplied to obtain a second fused feature map; the second fused feature map is input into the second residual block of the main branch encoder to obtain a third spectral feature map, the first densely connected feature map is input into the second residual block of the sub-branch encoder to obtain a third exponential feature map, the third exponential feature map and the first densely connected feature map are densely connected, Obtain a second densely connected feature map, input the second densely connected feature map into the third exponential space attention module of the sub-branch to obtain a third exponential feature space weight matrix, multiply the third exponential feature space weight matrix and the third spectral feature map to obtain a third fusion feature map; input the third fusion feature map into the third residual block of the main branch encoder to obtain a fourth spectral feature map, input the second densely connected feature map into the third residual block of the sub-branch encoder to obtain a fourth exponential feature map, densely connect the fourth exponential feature map and the second densely connected feature map to obtain a third densely connected feature map, input the third densely connected feature map into the fourth exponential space attention module of the sub-branch to obtain the fourth spectral feature map. Four exponential feature space weight matrices, multiplying the fourth exponential feature space weight matrix and the fourth spectral feature map to obtain a fourth fused feature map; inputting the fourth fused feature map into the fourth residual block of the main branch encoder to obtain a fifth spectral feature map, inputting the third densely connected feature map into the fourth residual block of the sub-branch encoder to obtain a fifth exponential feature map, densely connecting the fifth exponential feature map and the third densely connected feature map to obtain a fourth densely connected feature map, inputting the fourth densely connected feature map into the fifth exponential space attention module of the sub-branch to obtain a fifth exponential feature space weight matrix, and multiplying the fifth exponential feature space weight matrix and the fifth spectral feature map to obtain a fifth fused feature map;Input the fifth fused feature map into the dilute spatial pyramid pooling module to obtain a dilute spatial pooling feature map, multiply the dilute spatial pooling feature map and the fifth exponential feature space weight matrix to obtain a sixth fused feature map, input the sixth fused feature map into the fourth residual block of the main branch decoder to obtain a sixth spectral feature map, densely connect the fifth exponential feature space weight matrix and the fourth exponential feature space weight matrix to obtain a first dense connection weight matrix, and multiply the sixth spectral feature map and the first dense connection weight matrix to obtain a seventh fused feature map; Input the seventh fused feature map into the third residual block of the main branch decoder to obtain a seventh spectral feature map, densely connect the first dense connection weight matrix and the third exponential feature space weight matrix to obtain a second dense connection weight matrix, and multiply the seventh spectral feature map and the second dense connection weight matrix to obtain an eighth fused feature map; input the eighth fused feature map into the second residual block of the main branch decoder to obtain an eighth spectral feature map, densely connect the second dense connection weight matrix and the second exponential feature space weight matrix to obtain a third dense connection weight matrix, and multiply the eighth spectral feature map and the third dense connection weight matrix to obtain a ninth fused feature map; input the ninth fused feature map into the first residual block of the main branch decoder to obtain a ninth spectral feature map, densely connect the third dense connection weight matrix and the first exponential feature space weight matrix to obtain a fourth dense connection weight matrix, and multiply the ninth spectral feature map and the fourth dense connection weight matrix to obtain a tenth fused feature map; The tenth fused feature map is input into the output module of the main branch to obtain the segmentation result output by the photovoltaic land classification model.

7. The large-area photovoltaic land classification method combining decision tree and deep learning as claimed in claim 6, characterized in that: The multispectral image is obtained in the following manner: In the coarse classification area image of the photovoltaic land after slicing, the images of the blue band, green band, red band and the first short-wave infrared band are synthesized to form a multispectral image.

8. The large-area photovoltaic land classification method combining decision tree and deep learning as claimed in claim 6, characterized in that: The method for obtaining the index image is specifically as follows: The red band, near-infrared band, first short-wave infrared band and second short-wave infrared band images are taken from the sliced ​​photovoltaic land coarse classification area image and calculated according to the calculation formula of the first photovoltaic land index to obtain the first photovoltaic land index image. The blue band, red band, near-infrared band, first short-wave infrared band and second short-wave infrared band images are taken from the sliced ​​photovoltaic land coarse classification area image and calculated according to the calculation formula of the second photovoltaic land index to obtain the second photovoltaic land index image. The first photovoltaic land index image and the second photovoltaic land index image are band-synthesized to form an index image.

9. A large-area photovoltaic land classification system combining decision trees and deep learning, characterized in that: include: The data acquisition module is used to select data of representative photovoltaic land areas from the global photovoltaic land location dataset and classification dataset, and construct a photovoltaic land sample dataset that includes remote sensing images, geographic coordinates, imaging time, underlying land cover type, and photovoltaic array arrangement; The coarse classification module is used to input the photovoltaic land sample data set into the constructed decision tree classifier for coarse classification, and obtain the coarse segmentation result in the form of a binary raster map; A post-processing module is used to post-process the coarse segmentation results, set a buffer zone of a preset size around the post-processed coarse segmentation results, frame an outer envelope rectangle outside the buffer zone, and form a coarse classification area image of the photovoltaic land from the pixels within the outer envelope rectangle in the remote sensing image; The dataset construction module is used to generate a true photovoltaic land label for each photovoltaic land coarse classification area image through human-computer interaction, and to slice the photovoltaic land coarse classification area image and the photovoltaic land true label. The sliced ​​photovoltaic land coarse classification area image and the photovoltaic land true label constitute the photovoltaic land classification dataset; Model training module, used to train a photovoltaic land classification model that integrates multi-scale photovoltaic land features on the photovoltaic land classification dataset; The result acquisition module is used to input the coarse classification area image of the photovoltaic land to be classified into the trained photovoltaic land classification model to obtain the segmentation result of the coarse classification area image of the photovoltaic land, thereby realizing the fine identification of photovoltaic land; The photovoltaic land classification model has a main-sub dual-branch structure. The main branch includes an input module, an encoder, a dilated spatial pyramid pooling module, a decoder, and an output module. The sub-branch includes an input module, an encoder, and five exponential spatial attention modules. The encoder, encoder, and decoder of the main branch each contain four residual blocks. The main branch has a U-shaped symmetrical structure. Each residual block of the main branch decoder is jump-connected with each residual block of the same level of the main branch encoder. The input and output of each residual block of the sub-branch encoder are densely connected. The exponential feature space weight generated by each exponential spatial attention module in the sub-branch is densely connected with each residual block of the same level of the main branch encoder. The exponential feature space weight generated by each exponential spatial attention module in the sub-branch is densely connected with the exponential feature space weight generated by the subsequent exponential spatial attention module, and then densely connected with each residual block of the same level of the main branch decoder. The input module of the main branch is a multispectral image generated based on the coarse classification area image of the photovoltaic land after slicing. The input module of the sub-branch is an index image generated based on the coarse classification area image of the photovoltaic land after slicing.

Images