Low-efficiency industrial land two-dimensional fusion identification and evaluation method based on deep learning

Abstract

The invention discloses a low-efficiency industrial land two-dimensional fusion identification and evaluation method based on deep learning. The method comprises the following steps: S1, acquiring industrial land data and enterprise data of a research city, and associating the industrial land data and the enterprise data to obtain an industrial land enterprise database; s2, constructing a multi-dimensional index system to carry out weighted evaluation on the industrial land enterprise database and screening out low-efficiency industrial land units; s3, constructing a waste building recognition model for performing waste building recognition training through the street view image sample data, obtaining street view image data associated with the low-efficiency industrial land unit, inputting the street view image data into the waste building recognition model, and extracting and recognizing a waste building entity and a boundary region in the waste building recognition model; and S4, correspondingly displaying the boundary area of the low-efficiency industrial land unit in the research urban road map. According to the method, the low-efficiency industrial land is screened and re-checked and identified by fusing the multi-dimensional indexes including economy, income and the like with the two-dimensional fusion of the real scene features of the building, so that the accurate identification and classification judgment of the low-efficiency industrial land are realized.

Description

Technical Field

[0001] The present invention relates to the field of industrial land planning, and in particular to a two-dimensional fusion identification and evaluation method for inefficient industrial land based on deep learning. Background Art

[0002] With the acceleration of urbanization and the increasing scarcity of land resources, the existence of inefficient industrial land has become a prominent bottleneck hindering high-quality urban development. Inefficient industrial land generally refers to land with low unit output, low land utilization rates, or a chaotic functional layout. This type of land not only wastes land resources but can also lead to a series of chain reactions, such as an inefficient industrial structure and insufficient public facilities. Therefore, accurately identifying inefficient industrial land has become a core challenge urgently needed in urban renewal and land preparation. While traditional field survey methods can capture the true state of land parcels, they suffer from high labor costs and long survey cycles, making them difficult to scale, especially across large areas. This limits both identification efficiency and coverage. Intelligent recognition methods based on remote sensing imagery are limited to a vertical perspective and lack the ability to directly assess building details, spatial relationships, and other information. Currently, there is an urgent need for a recognition path technology that integrates multiple perspectives, improves recognition accuracy, and balances efficiency and intelligence to more scientifically support land resource integration. Summary of the Invention

[0003] The purpose of the present invention is to provide a two-dimensional fusion identification and evaluation method for inefficient industrial land based on deep learning, which screens inefficient industrial land through a multi-dimensional indicator system, and performs abandoned building entity identification processing through a constructed abandoned building identification model. The two-dimensional fusion screening and review of multi-dimensional indicators such as economy and income and real-life building features are integrated to identify and verify inefficient industrial land, thereby realizing accurate identification and classification of inefficient industrial land.

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

[0005] A two-dimensional fusion identification and evaluation method for inefficient industrial land based on deep learning includes:

[0006] S1. Obtain industrial land data and enterprise data of the research city and associate them according to geographic location information to obtain an industrial land enterprise database;

[0007] S2. Construct a multi-dimensional indicator system to conduct a weighted evaluation of the industrial land enterprise database and screen out inefficient industrial land units. The inefficient industrial land units contain interrelated industrial land data and enterprise data;

[0008] S3. Construct an abandoned building recognition model trained on street view image sample data. The abandoned building recognition model is built based on the YOLOv11 structure and integrates the convolutional additive self-attention module. Obtain street view image data associated with inefficient industrial land units and input it into the abandoned building recognition model to extract and identify abandoned building entities and boundary areas.

[0009] S4. The boundary areas of the low-efficiency industrial land units are displayed in the study city road map.

[0010] In order to better implement the present invention, in method S1, the industrial land data is the land area block, and the enterprise data includes the enterprise land boundaries, buildings, per mu tax revenue, per mu added value, unit energy consumption added value, unit emission added value, total labor productivity and R&D expense ratio. The enterprise data is associated and aggregated according to the geographic location information of the land area block, and the industrial land enterprise database is associated and aggregated according to the land area block and enterprise data level.

[0011] Preferably, the multidimensional indicator system includes indicators and indicator weights, and the indicators of the multidimensional indicator system include per-mu tax revenue, per-mu added value, unit energy consumption added value, unit emission added value, total labor productivity and R&D expense ratio; the method for obtaining the assessment value corresponding to the land area block in the industrial land enterprise database is as follows:

[0012] S21. Obtain the average value of the indicator data of the geographical location j and indicator i in the land area block in the past T years before the research date and multiply the average value by the coefficient M. i As the reference value P ij , add and calculate the benchmark value P of index i i ;

[0013] S22. Calculate the evaluation value S of the area block, which is expressed as follows:

[0014] Where n is the total number of indicators of land area blocks, E i is the total data corresponding to the indicator i in the land area block, K i is the indicator weight of indicator i in the land area block;

[0015] The assessed values of all land area blocks in the industrial land enterprise database are calculated in sequence.

[0016] Preferably, an industrial classification standard is constructed to divide several industrial categories, each industrial category contains at least four grade intervals from high to low according to the evaluation value, and the industrial classification standard is used to classify the land area blocks into industrial categories and grade intervals under the industrial categories in turn, and the two grade intervals at the bottom are screened out as low-efficiency industrial land units.

[0017] Preferably, the street view image sample data includes labels of abandoned buildings and normal buildings. The abandoned building recognition model first extracts building features from the street view image sample data for building recognition, and then distinguishes and recognizes abandoned buildings from normal buildings based on the fusion of spatial morphological features and image features. The spatial morphological features include structural damage and functional abandonment features, and the image features include atlas features such as roof collapse and vegetation cover.

[0018] Preferably, the abandoned building recognition model is an improved structure based on the YOLOv11 architecture. The abandoned building recognition model includes a feature extraction Backbone part, a neck network Neck part, and a head decision network Head part. The feature extraction Backbone part includes several convolution modules, several convolutional additive self-attention AdditiveBlock modules, a spatial pyramid pooling SPPF module and an attention mechanism C2PSA module; the neck network Neck part is used for feature fusion and enhancement, and the neck network Neck part includes several convolutional additive self-attention AdditiveBlock modules, several upsampling modules and several splicing modules; the head decision network Head part is used to output the abandoned building recognition result, and the abandoned building recognition result is the abandoned building entity and boundary area.

[0019] Preferably, the convolutional additive self-attention AdditiveBlock module includes an integrated sub-network, a CATM and an MLP module connected in sequence, wherein the integrated sub-network is composed of three 3×3 depth-separable convolutional layers activated by ReLU, and the integrated sub-network is used to expand the receptive field of the street view feature image and integrate local information in the street view feature image; the CATM module is used to abandon the global information interaction of building features and perform feature fusion through space and channels; the MLP module performs feature processing and output.

[0020] Preferably, the feature extraction Backbone part includes four convolutional additive self-attention AdditiveBlock modules connected in sequence through convolution modules, and the first convolutional additive self-attention AdditiveBlock module is connected in series with two convolution modules in front; the four convolutional additive self-attention AdditiveBlock modules are connected in sequence to the spatial pyramid pooling SPPF module and the attention mechanism C2PSA module; the neck network Neck part includes four convolutional additive self-attention AdditiveBlock modules, the first convolutional additive self-attention AdditiveBlock module upsamples the output features of the attention mechanism C2PSA module and splices the third convolutional additive self-attention AdditiveBlock module in the feature extraction Backbone part. The feature output of the block, the second convolutional additive self-attention AdditiveBlock module of the neck network Neck part upsamples the output features of the first convolutional additive self-attention AdditiveBlock module and splices the feature output of the second convolutional additive self-attention AdditiveBlock module in the feature extraction Backbone part, the third convolutional additive self-attention AdditiveBlock module splices the feature outputs of the first and second convolutional additive self-attention AdditiveBlock modules, and the fourth convolutional additive self-attention AdditiveBlock module splices the feature output of the attention mechanism C2PSA module in the feature extraction Backbone part with the feature output of the third convolutional additive self-attention AdditiveBlock module.

[0021] Preferably, in method S4, the road map includes boundary lines including administrative boundaries, roads, water areas, and parks. The abandoned building entities and boundary areas in each inefficient industrial land unit are first displayed on the road map, and then the boundaries of the abandoned building entities are extended to the boundary lines in the road map.

[0022] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0023] (1) The present invention screens inefficient industrial land through a multi-dimensional indicator system, identifies abandoned building entities through a constructed abandoned building identification model, and integrates a two-dimensional fusion system including multi-dimensional indicators such as economy and benefits with the real-life characteristics of buildings to screen and review and identify inefficient industrial land, thereby achieving accurate identification and classification of inefficient industrial land.

[0024] (2) The present invention constructs and utilizes a high-precision abandoned building recognition model to achieve intelligent perception and precise discrimination of building status; the abandoned building recognition model of the present invention is an improved structure based on the YOLOv11 architecture, and the model adopts the convolutional additive self-attention AdditiveBlock module to enhance the model's global modeling ability and multi-scale feature expression ability; the abandoned building recognition model has lower classification loss, better training and fitting effect, stronger generalization ability, and better indicators such as accuracy and mean average precision (mAP).

[0025] (3) The multi-dimensional indicator system of the present invention improves the scientificity and feasibility of the initial screening of inefficient plots. Combined with the weighted scoring mechanism, the evaluation values of all land use areas are quantified, which can scientifically identify industrial land with relatively low economic benefits. The multi-dimensional indicator method improves the systematicness and accuracy of the method for determining land efficiency, and provides a high-quality data basis for subsequent land redevelopment. The present invention integrates and analyzes the preliminary identification results based on economic benefits with the output results of the abandoned building identification model, and finally identifies inefficient industrial land through dual-dimensional fusion screening, which helps to avoid misjudgment and wrong judgment, and promotes the industrial land redevelopment policy to focus more on the truly "inefficient" plots, and realize the optimal allocation and intensive utilization of land resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 This is a flow chart of the method for dual-dimensional fusion identification and evaluation of inefficient industrial land of the present invention;

[0027] Figure 2 Schematic diagram of the improved YOLOv11 structure in the abandoned building recognition model in the embodiment;

[0028] Figure 3 Schematic diagram of the principle structure of the convolutional additive self-attention AdditiveBlock module in the embodiment;

[0029] Figure 4 This is a schematic diagram of building identification as an example in the embodiment;

[0030] Figure 5 This is a schematic diagram of abandoned building feature recognition using street view image sample data as an example in the embodiment;

[0031] Figure 6 This is a graph comparing the loss and accuracy of the abandoned building recognition model tested under the number of training rounds in the embodiment;

[0032] Figure 7 The spatial distribution map of inefficient industrial land implemented in a research city is given as an example in the embodiments. DETAILED DESCRIPTION

[0033] Below in conjunction with embodiment, the present invention is described in further detail:

[0034] Example

[0035] like Figure 1 As shown in the figure, a two-dimensional fusion identification and evaluation method for inefficient industrial land based on deep learning is proposed, which includes:

[0036] S1. Obtain the industrial land data and enterprise data of the research city and associate them according to the geographic location information to obtain an industrial land enterprise database. In method S1, the industrial land data is a land area block, and the enterprise data includes enterprise land boundaries, buildings, per-mu tax revenue, per-mu added value, unit energy consumption added value, unit emission added value, total labor productivity and R&D expense ratio. The enterprise data includes all data or calculated data such as enterprise land boundaries, buildings, per-mu tax revenue, per-mu added value, unit energy consumption added value, unit emission added value, total labor productivity ratio and R&D expense ratio (i.e., R&D expense expenditure as a percentage of operating income) (the calculated data is the data calculated from the source data, such as per-mu tax revenue is the tax revenue of the land area block divided by the number of acres occupied). Enterprise data is associated and aggregated according to the geographic location information of the land area block, and the industrial land enterprise database is associated and aggregated according to the land area block and enterprise data level. Preferably, the industrial land data and enterprise data in the industrial land enterprise database are first subjected to data cleaning and spatial matching operations.

[0037] S2. Construct a multi-dimensional indicator system to conduct weighted evaluation of the industrial land enterprise database and screen out inefficient industrial land units. The inefficient industrial land units contain interrelated industrial land data and enterprise data.

[0038] In some embodiments, the multi-dimensional indicator system includes indicators and indicator weights. The indicators of the multi-dimensional indicator system include per-mu tax revenue, per-mu added value, per-unit energy consumption added value, per-unit emission added value, total labor productivity ratio, and R&D expense ratio (i.e., the ratio of R&D expense to operating income). The method for obtaining the assessment value corresponding to the land area block in the industrial land enterprise database is as follows:

[0039] S21, obtain the index data of the geographical location j and index i in the land area block in the past T years of the study date, calculate the average value and multiply the average value by the coefficient Mi as the benchmark value Pij, and add the sum to calculate the benchmark value Pi of index i. Preferably, set the ratio of the index data corresponding to the calculated index i to the benchmark value The limit is 2, and ratios greater than 2 are counted as 2.

[0040] S22. Calculate the evaluation value S of the area block, which is expressed as follows:

[0041] Where n is the total number of indicators of land area blocks, E iis the total data corresponding to the indicator i in the land area block, K i The weight of indicator i in the land area block is set according to the policy planning of the research city (fine-tuning is performed in different cities). A simple example of this embodiment is as follows: the weight of the per-mu tax share is 40%, the weight of the per-mu added value share is 20%, the weight of the unit energy consumption added value share is 10%, the weight of the unit emission added value share is 10%, the weight of the total labor productivity share is 10%, and the weight of the proportion of R&D expenditure to operating income is 10%. The evaluation values of all land area blocks in the industrial land enterprise database are calculated in sequence according to the above method.

[0042] In some embodiments, an industrial classification standard is constructed to divide several industrial categories, each industrial category contains at least four grade intervals from high to low according to the evaluation value, and the industrial classification standard is used to classify the land area blocks into industrial categories and the grade intervals under the industrial categories in turn, and the two grade intervals at the bottom are screened out as inefficient industrial land units. A simple example of this embodiment is as follows: the industrial classification standard divides the industrial categories into high-energy-consuming industries and general industries, and the land area blocks are classified into industrial categories in turn, and then classified according to the grade intervals of the evaluation values under the industrial categories. For example, the grade intervals are divided into four grade intervals of A, B, C, and D. Finally, the locations of enterprises with evaluation grades of C and D (i.e., land area blocks) are preliminarily screened and identified as economically inefficient industrial land.

[0043] For example, Ningbo is selected as the research city. The industrial classification standard divides the industry into high-energy-consuming industries and general industries, and the grade intervals are A, B, C, and D. The details are as follows:

[0044] Table 1 Industrial classification ratio

[0045]

[0046] Finally, the locations of enterprises with evaluation grades of C and D (i.e., land use blocks) will be preliminarily screened and identified as economically inefficient industrial land.

[0047] S3. Construct an abandoned building recognition model that uses street view image sample data to train abandoned building recognition. The abandoned building recognition model is built based on the YOLOv11 structure and integrates the convolutional additive self-attention module. Obtain street view image data associated with inefficient industrial land units and input it into the abandoned building recognition model to extract and identify the abandoned building entities and boundary areas. The abandoned building recognition model first performs building recognition (such as Figure 4 Then, the abandoned building features are identified.

[0048] In some embodiments, the street view image sample data includes labels for abandoned buildings and normal buildings. The labels clearly mark the outline of each building and its usage status, and the buildings are divided into two categories: "Abandoned-building" and "Normal-building". The abandoned building recognition model first extracts building features from the street view image sample data for building recognition, and then distinguishes and identifies abandoned buildings ("Abandoned-building") from normal buildings ("Normal-building") based on the fusion of spatial morphological features and image features. Spatial morphological features include structural damage and functional abandonment features, and image features include atlas features such as roof collapse and vegetation cover. The abandoned building recognition model is an improved structure based on the YOLOv11 architecture. Preferably, the street view images input to the abandoned building recognition model are uniformly scaled to a resolution of 640×640 pixels to meet the model input tensor dimension requirements; the experimental platform uses the Windows 11 operating system, and hardware acceleration relies on the NVIDIA GeForce RTX 4060 GPU (8GB video memory), in conjunction with the CUDA 12.2 parallel computing architecture and cuDNN acceleration library; the model optimization uses the stochastic gradient descent (SGD) algorithm, the initial learning rate is set to 0.01, the training process lasts for 300 iterations, the batch size is set to 16 to adapt to the GPU video memory capacity, and the data cache mechanism is enabled (cache=True) to improve data loading efficiency. In order to avoid potential compatibility issues with multi-threaded data loading under Windows systems, set the workers=0 parameter to turn off multi-process acceleration. Figure 2 As shown, the abandoned building recognition model includes the feature extraction Backbone part ( Figure 2 Backbone corresponding to the dotted box), the neck network Neck part ( Figure 2 The dotted box corresponding to Neck), the head decision network Head part ( Figure 2 The feature extraction backbone part includes several convolution modules (i.e. Figure 2 conv module in ), several convolutional additive self-attention AdditiveBlock modules (i.e. Figure 2 AdditiveBlock module in the spatial pyramid pooling SPPF module (i.e. Figure 2 SPPF module in) and attention mechanism C2PSA module (i.e. Figure 2 The C2PSA module in the neck network is used for feature fusion and enhancement. The neck network Neck part includes several convolutional additive self-attention AdditiveBlock modules (i.e. Figure 2C2PSA module in ), several upsampling modules (i.e. Figure 2 Upsample module in ) and several splicing modules (i.e. Figure 2 The head decision network outputs the abandoned building recognition results, which include the abandoned building entity and its boundary area. The street view image sample data in the abandoned building recognition model is divided into a training set, a validation set, and a test set in an 8:1:1 ratio. These sets are used for model training, model convergence determination, and performance evaluation of the trained model, respectively. The present invention uses stratified sampling for data partitioning to ensure that the proportions of different building types in each dataset are similar.

[0049] Table 2 Classification of abandoned building dataset

[0050]

[0051] The abandoned building recognition model uses indicators such as precision (P), recall (R), and mean average precision (mAP) of all class labels when the intersection-over-union ratio threshold is 50% to evaluate the model, in order to quantify the model performance and verify its applicability. In this example, the abandoned building recognition model of the present invention is compared with the traditional YOLOv11 algorithm model. The results are shown in Figure 2. Figure 6 shown; from Figure 5 (a) As can be seen from the classification loss curve, the classification loss value of the abandoned building recognition model of the present invention is significantly lower, indicating that the abandoned building recognition model of the present invention has a better fit to the training data and helps to improve the performance of the model on the test set. Figure 5 (b) is a graph of the mean average precision of the validation set. It can be seen that the mean average precision of the abandoned building recognition model of the present invention is significantly higher, which improves the expression ability and perception of the model and achieves better detection performance. Among them, the precision rate (Precision) is improved by 8.9%, the recall rate (Recall) is improved by 4.4%, and the mean average precision (mAP) is improved by 5.6%. While reducing the missed detection rate, the improved model further improves the accuracy of detection, and its detection performance on the street view image building dataset is significantly optimized. The total mAP (mean average precision) of the abandoned building recognition model of the present invention reaches 80.1%, indicating that it has a high classification accuracy in the building recognition task of street view images.

[0052] In the implementation example of Ningbo, the research city is as follows: Figure 5 As shown in the figure, the spatial morphological features and image features in the street view image sample data label are integrated with the architectural features as follows: Figure 5The building in a has a dilapidated appearance and only the building frame remains (abandoned building characteristics). Figure 5 b) unfinished abandoned buildings, such as those without doors or windows installed (buildings under construction are not considered abandoned); Figure 5 c. Doors and windows are damaged or missing and have not been maintained for a long time; Figure 5 d. Doors and windows are blocked or sealed by debris for a long time; Figure 5 In e, there are no signs of production, such as weeds blocking the entrance, rubble blocking the entrance, etc.; Figure 5 In f, the building facade has cracked, partially demolished or collapsed walls; Figure 5 In g, the roof collapsed or was removed; in Figure 5 In h, there are words such as "demolish" on the wall of the building.

[0053] In some embodiments, as Figure 3 As shown, the convolutional additive self-attention AdditiveBlock module includes sequentially connected integrated sub-networks (i.e. Figure 3 integration module in), CATM module (i.e. Figure 3 CATM module in ) and MLP module (i.e. Figure 3 The integrated sub-network consists of three 3×3 depth-wise separable convolutional layers activated by ReLU. The integrated sub-network is used to expand the receptive field of street view feature images and integrate local information in street view feature images; the CATM module is used to interact with the global information of abandoned building features and perform feature fusion through space and channels; the MLP module performs feature processing and output.

[0054] The feature extraction Backbone part includes four convolutional additive self-attention AdditiveBlock modules connected in sequence through convolutional modules, and the first convolutional additive self-attention AdditiveBlock module is connected in series with two convolutional modules; the four convolutional additive self-attention AdditiveBlock modules are connected in sequence to the spatial pyramid pooling SPPF module and the attention mechanism C2PSA module; the neck network Neck part includes four convolutional additive self-attention AdditiveBlock modules, the first convolutional additive self-attention AdditiveBlock module upsamples the output features of the attention mechanism C2PSA module and splices the features of the third convolutional additive self-attention AdditiveBlock module in the feature extraction Backbone part. The second convolutional additive self-attention AdditiveBlock module of the neck part of the neck network upsamples the output features of the first convolutional additive self-attention AdditiveBlock module and splices the feature output of the second convolutional additive self-attention AdditiveBlock module in the feature extraction Backbone part. The third convolutional additive self-attention AdditiveBlock module splices the feature outputs of the first and second convolutional additive self-attention AdditiveBlock modules. The fourth convolutional additive self-attention AdditiveBlock module splices the feature output of the attention mechanism C2PSA module in the feature extraction Backbone part and the feature output of the third convolutional additive self-attention AdditiveBlock module.

[0055] S4. The boundary area of the low-efficiency industrial land unit is displayed in the road map of the research city. In the implementation example of the research city being Ningbo, Figure 7 As shown in the figure, the accurate identification of inefficient industrial land in the city was achieved, and finally 118 inefficient industrial land plots in the central urban area of Ningbo were identified. Figure 7 This is the result of identifying inefficient industrial land in the central urban area of Ningbo. Combining economic benefit evaluation results with building condition identification results, economically inefficient land parcels are re-identified: if the model identifies an "abandoned building" entity in the street view imagery of the parcel, it is considered inefficient industrial land. This dual-dimensional fusion judgment mechanism significantly improves identification accuracy and practical application value.

[0056] In some preferred embodiments, the road map includes boundary lines including administrative boundaries, roads, water areas, and parks. The abandoned building entities and boundary areas in each inefficient industrial land unit are first displayed on the road map, and then the boundaries of the abandoned building entities are extended to the boundary lines in the road map.

[0057] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A deep learning-based dual-dimensional fusion identification and assessment method for inefficient industrial land, characterized by: The methods include: S1. Obtain industrial land data and enterprise data of the research city and associate them according to geographic location information to obtain an industrial land enterprise database; S2. Construct a multi-dimensional indicator system to conduct a weighted evaluation of the industrial land enterprise database and screen out inefficient industrial land units. The inefficient industrial land units contain interrelated industrial land data and enterprise data; S3. Construct an abandoned building recognition model trained on street view image sample data. The abandoned building recognition model is built based on the YOLOv11 structure and integrates the convolutional additive self-attention module. Obtain street view image data associated with inefficient industrial land units and input it into the abandoned building recognition model to extract and identify abandoned building entities and boundary areas. S4. The boundary areas of the low-efficiency industrial land units are displayed in the study city road map.

2. The deep learning-based dual-dimensional fusion identification and assessment method for inefficient industrial land according to claim 1 is characterized by: In method S1, industrial land data are land area blocks, and enterprise data include enterprise land boundaries, buildings, per-mu tax revenue, per-mu added value, unit energy consumption added value, unit emission added value, total labor productivity and R&D expense ratio. Enterprise data are associated and aggregated according to geographic location information based on land area blocks, and the industrial land enterprise database is associated and aggregated according to land area blocks and enterprise data levels.

3. The deep learning-based dual-dimensional fusion identification and assessment method for inefficient industrial land according to claim 2 is characterized by: The multi-dimensional indicator system includes indicators and indicator weights. The indicators of the multi-dimensional indicator system include per-mu tax revenue, per-mu added value, per-unit energy consumption added value, per-unit emission added value, total labor productivity and R&D expense ratio. The method for obtaining the assessment value corresponding to the land area block in the industrial land enterprise database is as follows: S21. Obtain the average value of the indicator data of the geographical location j and indicator i in the land area block in the past T years before the research date and multiply the average value by the coefficient M. i As the reference value P ij , add and calculate the benchmark value P of index i i ; S22. Calculate the evaluation value S of the area block, which is expressed as follows: Where n is the total number of indicators of land area blocks, E i is the total data corresponding to the indicator i in the land area block, K i is the indicator weight of indicator i in the land area block; The assessed values of all land area blocks in the industrial land enterprise database are calculated in sequence.

4. The deep learning-based dual-dimensional fusion identification and assessment method for inefficient industrial land according to claim 3 is characterized by: Construct an industrial classification standard that divides several industrial categories. Each industrial category contains at least four grade intervals ranked from high to low according to the assessed value. Use the industrial classification standard to classify the land area blocks into industrial categories and grade intervals under the industrial categories in turn, and select the two grade intervals at the bottom as low-efficiency industrial land units.

5. The deep learning-based dual-dimensional fusion identification and assessment method for inefficient industrial land according to claim 1 is characterized by: The street view image sample data includes labels for abandoned buildings and normal buildings. The abandoned building recognition model first extracts building features from the street view image sample data for building recognition, and then distinguishes and recognizes abandoned buildings from normal buildings based on the fusion of spatial morphological features and image features. The spatial morphological features include structural damage and functional abandonment features, and the image features include atlas features such as roof collapse and vegetation cover.

6. The deep learning-based dual-dimensional fusion identification and assessment method for inefficient industrial land according to claim 1 is characterized by: The abandoned building recognition model is an improved structure based on the YOLOvll architecture. The abandoned building recognition model includes a feature extraction Backbone part, a neck network Neck part, and a head decision network Head part. The feature extraction Backbone part includes several convolution modules, several convolutional additive self-attention AdditiveBlock modules, a spatial pyramid pooling SPPF module and an attention mechanism C2PSA module; the neck network Neck part is used for feature fusion and enhancement, and the neck network Neck part includes several convolutional additive self-attention AdditiveBlock modules, several upsampling modules and several splicing modules; the head decision network Head part is used to output the abandoned building recognition result, and the abandoned building recognition result is the abandoned building entity and boundary area.

7. The deep learning-based dual-dimensional fusion identification and assessment method for inefficient industrial land according to claim 6 is characterized by: The convolutional additive self-attention AdditiveBlock module includes an integrated sub-network, a CATM and an MLP module connected in sequence. The integrated sub-network is composed of a 3×3 depth-separable convolution layer activated by ReLU. The integrated sub-network is used to expand the receptive field of the street view feature image and integrate local information in the street view feature image; the CATM module is used to abandon the global information interaction of building features and perform feature fusion through space and channels; the MLP module performs feature processing and output.

8. The deep learning-based dual-dimensional fusion identification and assessment method for inefficient industrial land according to claim 6 is characterized by: The feature extraction Backbone part includes four convolutional additive self-attention AdditiveBlock modules connected in sequence through convolutional modules, and the first convolutional additive self-attention AdditiveBlock module is connected in series with two convolutional modules; the four convolutional additive self-attention AdditiveBlock modules are connected in sequence to the spatial pyramid pooling SPPF module and the attention mechanism C2PSA module; the neck network Neck part includes four convolutional additive self-attention AdditiveBlock modules, the first convolutional additive self-attention AdditiveBlock module upsamples the output features of the attention mechanism C2PSA module and splices the feature output of the third convolutional additive self-attention AdditiveBlock module in the feature extraction Backbone part, and the first convolutional additive self-attention AdditiveBlock module in the neck network Neck part The lock module upsamples the output features of the first convolutional additive self-attention AdditiveBlock module and splices the feature output of the first convolutional additive self-attention AdditiveBlock module in the feature extraction Backbone part. The third convolutional additive self-attention AdditiveBlock module splices the feature outputs of the first and second convolutional additive self-attention AdditiveBlock modules. The fourth convolutional additive self-attention AdditiveBlock module splices the feature output of the attention mechanism C2PSA module in the feature extraction Backbone part and the feature output of the third convolutional additive self-attention AdditiveBlock module.

9. The deep learning-based dual-dimensional fusion identification and assessment method for inefficient industrial land according to claim 1 is characterized by: In method S4, the road map includes boundary lines including administrative boundaries, roads, water areas, and parks. The abandoned building entities and boundary areas in each inefficient industrial land unit are first displayed on the road map, and then the boundaries of the abandoned building entities are extended to the boundary lines in the road map.

Images