Construction initial-stage land parcel detection method fusing remote sensing time-space spectrum characteristics

Abstract

The invention provides an initial construction land parcel detection method fusing remote sensing time-space spectrum characteristics, which comprises the following steps: carrying out wave band operation on different time phase remote sensing images by utilizing a land cleaning land parcel detector, and realizing primary screening detection on potential change land parcels based on spectral characteristic variation caused by land cleaning; a double-branch deep convolutional neural network of a convolutional neural network based on an object-scene is used for extracting double spatial features, a pseudo-change region caused by farming activities and vegetation phenological changes is eliminated through decision-making level fusion of feature spaces, and finally high-precision initial development plot pattern spots of the construction land are output. According to the invention, the spatial features of the target object and the scene thereof can be synchronously extracted by using the double-branch deep network, and high-precision detection of the initial development plot of the construction land is realized; according to the method, high-precision extraction of the initial development plot of the construction land is realized, the problem of false detection caused by crop rotation and vegetation phenological change is solved, and the accuracy and reliability of a detection result are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of remote sensing detection technology, and in particular to a method for detecting land parcels in the initial stage of construction by integrating remote sensing spatiotemporal spectral features. Background Art

[0002] Illegal construction land development has become a major threat to the sustainable use of land, leading to the loss of arable land resources, ecological damage, and inefficient land use. Existing remote sensing-based construction land monitoring systems primarily rely on the identification of completed buildings. However, this end-point monitoring model suffers from a significant lag: by the time buildings are detected, development activities have largely completed, and the resulting ecological damage and social costs are often irreversible. Therefore, monitoring should be shifted to the construction land development phase, particularly the initial land clearance phase. This phase is typically characterized by the removal of surface cover (including vegetation, water bodies, and existing structures), resulting in bare land awaiting construction. Compared to established buildings, bare land in the early stages of development is smaller in scale and more reversible. By developing detection technology tailored to these early development sites, early warning monitoring of illegal construction activities can be achieved, providing a critical window for land law enforcement and effectively curbing the spread of illegal construction.

[0003] In response to the temporal spectral change characteristics caused by land leveling in the early stage of construction land development, researchers have proposed a detection method based on the Land Clearing Index (LCI). The core of this method is to construct LCI through band operations of two-phase images to characterize the typical spectral change characteristics of the process in which the surface cover changes from vegetation, water bodies or buildings to bare land to be built; on this basis, adaptive threshold segmentation technology is used to realize the automatic extraction of early development plots of construction land. The LCI method has the advantages of simple algorithm and high computational efficiency, and shows good applicability to different sensor data (such as Landsat, Sentinel, etc.). However, this method still has significant limitations in practical applications, mainly manifested in its sensitivity to agricultural production activities and seasonal changes in vegetation. Specifically: (1) Bare farmland formed after crop harvest has similar spectral characteristics to bare land to be built; (2) The seasonal growth and decline of natural vegetation (such as lush in summer and withered in winter) will also produce spectral change patterns similar to land clearing. These interference factors lead to a high false alarm rate in the detection results based on LCI, affecting the detection accuracy of the method. Summary of the Invention

[0004] In response to the shortcomings of the existing technology, the present invention provides a method for detecting plots of land in the early stages of construction that integrates remote sensing spatiotemporal spectral features. The present invention constructs a detection model through temporal spectral change features and spatial morphological features, which not only achieves high-precision extraction of plots of land for initial development of construction land, but also effectively overcomes the false detection problem caused by crop rotation and vegetation phenological changes, thereby significantly improving the accuracy and reliability of the detection results.

[0005] The technical solution of the present invention is: a method for detecting land parcels in the initial stage of construction by integrating remote sensing spatiotemporal spectral features, comprising the following steps:

[0006] S1) Obtaining temporal remote sensing images of the land parcel;

[0007] S2) Using the LCD of the land clearing plot detector to calculate the bands of remote sensing images of different phases, based on the spectral feature variation caused by land clearing, to achieve preliminary screening of potential change plots;

[0008] S3), constructing a scene image for each detected object;

[0009] S4), building a land parcel detection model based on an object-scene convolutional neural network and training the land parcel detection model;

[0010] S5) Use a dual-branch deep convolutional neural network based on an object-scene convolutional neural network to extract dual spatial features, eliminate pseudo-change areas caused by agricultural activities and vegetation phenological changes through decision-level fusion of feature space, and finally output a high-precision map of initial development plots of construction land.

[0011] Preferably, in step S2), the land clearing plot detector LCD simultaneously captures three types of areas with similar spectral responses during the initial screening detection of potential change plots: 1. initial development plots of construction land; 2. exposed farmland formed by crop harvesting; 3. areas with seasonal changes in natural vegetation.

[0012] Preferably, in step S2), the land clearing plot detector LCD performs preliminary screening of potential construction land development plots by calculating two remote sensing images of different phases, and defines a bare soil index RBLI according to the spectral characteristics of the bare soil, namely:

[0013]

[0014] Where, Band red Indicates the top of atmosphere reflectance or surface reflectance of the red band; Band blue Represents the top-of-atmosphere reflectance or surface reflectance in the blue band.

[0015] Preferably, in step S2), during the initial development of the construction land, the original surface cover is cleared and the land cover type is converted to bare soil. This change is reflected in the remote sensing image as the bare soil index (RBLI) of the second phase image is significantly higher than that of the first phase image. Therefore, the LCI value is defined as:

[0016]

[0017] Where, RBLI t1 and RBLI t2 The bare earth index was calculated from the remote sensing images of the first period and the second period respectively.

[0018] Preferably, in step S3), a minimum bounding rectangle is constructed for each detected object, and then the minimum bounding rectangle is expanded by introducing a scale parameter N to generate scene rectangles of different spatial ranges.

[0019] Preferably, in step S5), the dual spatial features include object features and scene features, wherein the object features are texture and morphological features of the detected plot, and the scene features are the spatial relationship between the detected plot and surrounding objects.

[0020] Preferably, in step S5), the object-scene convolutional neural network-based land parcel detection model OSN includes three branches; the three branches are used to process object images, resampling ratios, and scene images, respectively; the outputs of the three branches are fused through a cascade layer and then a binary classification result is generated through a fully connected layer to determine whether the object is an initial development plot of construction land, and a dropout layer is added to the fully connected layer to reduce overfitting.

[0021] Preferably, in step S5), the branch for processing the object image includes a convolution block, four residual blocks, and a spatial pyramid pooling SPP; the branch for processing the scene image includes a convolution block, seven residual blocks, and a spatial pyramid pooling SPP.

[0022] Preferably, in step S5), the convolution block includes a 3×3 convolution layer and a maximum pooling layer, and the 3×3 convolution layer and the maximum pooling layer are processed by batch normalization BN and ReLU activation function.

[0023] Preferably, in step S5), the residual block includes two 3×3 convolutional layers, followed by batch normalization BN and ReLU activation function, and the residual block introduces skip connection to enable the network to learn residual mapping.

[0024] Preferably, in step S5), the expression of the residual block is:

[0025] H(x)=x+F(x) (3)

[0026] Where H(x) represents the output of the residual block, x represents the input, and F(x) represents the residual function.

[0027] Preferably, in step S5), the expression of the residual function F(x) is:

[0028] F(x)=ReLU(BN(Conv3×3(ReLU(BN(Conv3×3(x)))))) (4)

[0029] Where Conv3×3(x) represents a 3×3 convolutional layer that processes the input x, BN represents batch normalization, and ReLU is the activation function.

[0030] Preferably, in step S5), the spatial pyramid pooling (SPP) divides the feature map into multiple grids of different scales, and performs a pooling operation on each grid.

[0031] Preferably, in step S5), the resampling ratio branch is processed by a flattening layer and calculated based on the relationship between the original size of the object image and the resampling size when input to the object-scene based convolutional neural network. The resampling ratio is expressed as a pair of resampling factors (R x ,R y ), the resampling factor (R x ,R y ) is expressed as:

[0032]

[0033] Where h i 、w i The height and width H of the minimum bounding rectangle of the object O shape are respectively input 、

[0034] W input They represent the resampled height and width when input to the object-scene based convolutional neural network.

[0035] The beneficial effects of the present invention are:

[0036] 1. The present invention adopts a dual-branch deep network architecture, which can simultaneously extract the spatial features of the target object and its scene, and work in conjunction with the LCD module to achieve high-precision detection of early development plots of construction land;

[0037] 2. The present invention constructs a detection model through the temporal spectral change characteristics and spatial morphological characteristics, which not only achieves high-precision extraction of early development plots of construction land, but also effectively overcomes the false detection problem caused by crop rotation and vegetation phenological changes, thereby significantly improving the accuracy and reliability of the detection results. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is a flow chart of the method of the present invention;

[0039] Figure 2 This is a structural framework diagram of the object-scene convolutional neural network of the present invention;

[0040] Figure 3 Schematic diagram of LCI values calculated from dual-temporal images of six cities in Example 2 of the present invention;

[0041] Figure 4 The scene images with different spatial ranges in Example 2 of the present invention;

[0042] Figure 5 Schematic diagram of the impact of the spatial range of scene images on OSN performance in Example 2 of the present invention;

[0043] Figure 6 1 is a comparison chart of the detection results of the method of Example 1 of the present invention and the comparative method LCI;

[0044] Figure 7 This is a comparison chart of the improvement effect of the false alarm rate in Example 2 of the present invention;

[0045] Figure 8 This is a comparison diagram of the impact of scene images on the detection of initial development plots of construction land in Example 2 of the present invention. DETAILED DESCRIPTION

[0046] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings:

[0047] Example 1

[0048] like Figure 1 As shown, the present invention provides a method for detecting land parcels in the initial stage of construction by integrating remote sensing spatiotemporal spectral features, comprising the following steps:

[0049] S1) Obtaining temporal remote sensing images of the land parcel;

[0050] S2) Using the LCD of the land clearing plot detector to perform band calculations on remote sensing images of different phases, based on the variation of spectral characteristics caused by land clearing, to achieve preliminary screening of potential change plots;

[0051] In this embodiment, the land clearing plot detector LCD simultaneously captures three types of areas with similar spectral responses during the initial screening detection of potential change plots: 1. Initial development plots of construction land; 2. Bare farmland formed by crop harvesting; and 3. Areas with seasonal changes in natural vegetation.

[0052] The land clearing plot detector LCD performs preliminary screening of potential construction land development plots by calculating two remote sensing images of different phases, and defines the bare soil index RBLI according to the spectral characteristics of bare soil, namely:

[0053]

[0054] Where, Band red Indicates the top of atmosphere reflectance or surface reflectance of the red band; Band blue Represents the top-of-atmosphere reflectance or surface reflectance in the blue band.

[0055] In this example, during the initial development of construction land, the original surface cover was cleared and the land cover type was converted to bare soil. This change is reflected in the remote sensing image as the bare soil index (RBLI) in the second phase of imagery is significantly higher than that in the first phase of imagery. Therefore, the LCI value is defined as:

[0056]

[0057] Where, RBLI t1 and RBLI t2 The bare earth index was calculated from the remote sensing images of the first period and the second period respectively.

[0058] Higher LCI values indicate a greater likelihood of land clearing, e.g. Figure 3 This example uses a threshold method to detect land clearing plots from LCI imagery. However, the detection results not only include construction land development plots but may also be affected by agricultural activities (such as crop harvesting) and seasonal vegetation changes (such as natural leaf fall and snowmelt), resulting in the misdetection of additional changed plots.

[0059] S3), constructing a scene image for each detected object;

[0060] This embodiment constructs a minimum bounding rectangle MBR for each detected object, and then expands the minimum bounding rectangle MBR by introducing a scale parameter N to generate scene rectangles of different spatial ranges, such as Figure 4 As shown; Figure 4 (a) shows the scene images of the initial development plot of construction land with scale parameters N ranging from 1 to 10; Figure 4 (b) Scene images showing the newly added bare land caused by crop harvesting with scale parameters N ranging from 1 to 10.

[0061] By systematically adjusting the value of N, a series of scene images with different spatial coverage are obtained.

[0062] S4) Construct a land parcel detection model OSN based on object-scene convolutional neural network, such as Figure 2 As shown, the land parcel detection model is trained;

[0063] In this embodiment, during the model training phase, each detection object provides two types of input data: an object image and its corresponding scene image. This embodiment sets 15 sets of training data sets, one of which uses only object images (no scene images). The training goal is to accurately distinguish between early development plots and false detection targets by fusing the two types of image information. The model performance is evaluated using four standard indicators: accuracy, precision, recall, and F1 score. Figure 5 As shown in the figure, when the scale parameter N increases from 0 to 9, all evaluation indicators show significant improvement; however, when the scale parameter N>9, the performance decreases significantly. This phenomenon indicates that: (1) scene context information can effectively improve the discrimination ability of OSN (optimal performance is achieved when N=9); (2) too large a spatial range (N>9) will introduce redundant information, causing the model to lose attention. Experiments based on Sentinel-2 images show that the scale parameter N=9 is the optimal choice for this method. At this time, the scene image can provide sufficient contextual support while avoiding the interference of irrelevant noise.

[0064] S5) Use the dual-branch deep convolutional neural network based on the object-scene convolutional neural network to extract dual spatial features, eliminate the pseudo-change areas caused by agricultural activities and vegetation phenological changes through decision-level fusion of feature space, and finally output a high-precision map of the initial development plots of construction land.

[0065] In this embodiment, the dual spatial features include object features and scene features, wherein the object features are texture and morphological features of the detection block, and the scene features are the spatial relationship between the detection block and surrounding objects.

[0066] like Figure 2As shown, the object-scene convolutional neural network (OSN)-based land parcel detection model consists of three branches: one for processing the object image, the other for processing the resampling scale, and the other for processing the scene image. The outputs of these three branches are fused through a cascade layer and then passed through a fully connected layer to generate a binary classification result, determining whether the object is an initial development parcel for construction. A dropout layer is added to the fully connected layer to reduce overfitting. The "object" refers to the land clearing parcel detected by the LCD. To extract different feature representations, the object image and the scene image are processed separately through two independent encoding branches.

[0067] In this embodiment, the branch for processing the object image includes a convolution block, four residual blocks, and a spatial pyramid pooling SPP; the branch for processing the scene image includes a convolution block, seven residual blocks, and a spatial pyramid pooling SPP.

[0068] The convolution block includes a 3×3 convolution layer and a maximum pooling layer, and the 3×3 convolution layer and the maximum pooling layer are processed by batch normalization BN and ReLU activation function.

[0069] The residual block consists of two 3×3 convolutional layers, followed by batch normalization (BN) and ReLU activation functions, and the residual block introduces skip connections to enable the network to learn residual mapping. The residual block is expressed as:

[0070] H(x)=x+F(x) (3)

[0071] Where H(x) represents the output of the residual block, x represents the input, and F(x) represents the residual function.

[0072] The expression of the residual function F(x) is:

[0073] F(x)=ReLU(BN(Conv3×3(ReLU(BN(Conv3×3(x)))))) (4)

[0074] Where Conv3×3(x) represents a 3×3 convolutional layer that processes the input x, BN represents batch normalization, and ReLU is the activation function.

[0075] The spatial pyramid pooling (SPP) divides the feature map into multiple grids of different scales and performs a pooling operation on each grid.

[0076] In this embodiment, the resampling ratio branch is processed by the flattening layer and calculated based on the relationship between the original size of the object image and the resampled size when input to the object-scene based convolutional neural network. The resampling ratio is expressed as a pair of resampling factors (Rx ,R y ), the resampling factor (R x ,R y ) is expressed as:

[0077]

[0078] Where h i 、w i The height and width H of the minimum bounding rectangle of the object O shape are respectively input 、

[0079] W input They represent the resampled height and width when input to the object-scene based convolutional neural network.

[0080] Example 2

[0081] This example tests the existing LCI method and the method proposed in Example 1 in six areas, including Beijing, Guangzhou, Nanjing, Bangkok, Ho Chi Minh City, and Kuala Lumpur. The test data is Sentinel-2 optical remote sensing images. Sentinel-2 is a component of the European Space Agency's Copernicus program. It consists of two identical satellites, Sentinel-2A and Sentinel-2B, which have an image resolution of 10 meters and a revisit period of 5 days. Table 1 lists the acquisition dates of the reference and test images used in each city. By visually interpreting the reference images and test images, we selected verification samples of developed plots and non-developed plots for evaluating the detection results. Table 2 lists the distribution of the number of samples in the training set, verification set, and test set used in the training method. Table 3 lists the number of verification samples selected for each detection.

[0082] Table 1 Remote sensing image datasets used for training and validation of the proposed method

[0083]

[0084] Table 2 shows the training samples collected from remote sensing image pairs in Guangzhou and Beijing. “Other changes” refer to land cover changes that are easily misidentified as initial development sites for construction, including bare land caused by crop harvesting and seasonal vegetation changes (unit: patch).

[0085]

[0086] Table 3 Verification samples collected from remote sensing image pairs of different cities (unit: pixel)

[0087]

[0088] The traditional LCI method and the newly proposed LCD-OSN method were used to extract and compare the initial development plots of construction land in the study area. The following three indicators were used for accuracy evaluation: (1) detection accuracy (DA): the proportion of correctly identified development plots to actual development plots; (2) false alarm rate (FR): the proportion of undeveloped plots mistakenly identified as development plots; and (3) overall accuracy (OA): the proportion of correctly classified plots among all validation samples. The accuracy evaluation results are shown in Table 5.

[0089] Table 4 False positive pixels in the detection of initial development plots of construction land: LCD-OSN vs. LCI method

[0090]

[0091] Table 5 Evaluation of detection results of initial development plots of construction land: LCD-OSN vs. LCI method

[0092]

[0093] Results show that while the LCI method performs well in terms of detection sensitivity, it suffers from a high false alarm rate in practical applications. In contrast, the LCD-OSN method proposed in Example 1 achieves significant performance improvements by combining temporal spectral variations with spatial features of ground objects: while maintaining high detection accuracy, it significantly reduces the false alarm rate. As shown in Table 4, the LCD-OSN method proposed in Example 1 can reduce false alarms by up to 96.1%, demonstrating a significant advantage over the LCI method. Empirical analysis of four representative regions—Beijing, Guangzhou, Kuala Lumpur, and Ho Chi Minh City—demonstrates significant optimization effects. The false alarm rate in Beijing decreased by as much as 7.7%, while the overall accuracy in Kuala Lumpur improved most significantly, with a relative increase of 6.33%. In-depth analysis shows that the LCD-OSN method proposed in Example 1 effectively overcomes the problem of false change signal interference caused by factors such as crop harvesting and seasonal vegetation changes in traditional methods by deeply integrating the spatial features of objects and scenes. This detection mechanism, which integrates multiple temporal, spatial, and spectral features, not only significantly reduces the false alarm rate but also achieves higher overall detection accuracy.

[0094] like Figure 6As shown in Figure 1, the effectiveness of the LCD-OSN method proposed in Example 1 in controlling the false alarm rate is further verified. The spatiotemporal distribution characteristics of construction land development in the study area are as follows: the development in Beijing, Nanjing and Guangzhou is concentrated in the urban fringe areas, which is characterized by the rapid conversion of farmland to construction land; Bangkok and Ho Chi Minh City are mainly characterized by the transformation of vegetation and farmland within the city; Kuala Lumpur has a small and scattered development plot due to the short image time interval. Although the LCI-based method can detect early development plots, its false alarms mainly come from three types of interference factors: (1) agricultural activities (such as rice field rotation), (2) seasonal vegetation changes, and (3) suspended sediments in water bodies. By fusing object and scene level information, the LCD-OSN method effectively suppresses the above interference: first, it reduces the false detection caused by farmland dynamics in Guangzhou and Ho Chi Minh City; second, it improves the detection robustness of areas affected by seasonal vegetation fluctuations in Beijing and Nanjing; and finally, it significantly alleviates the false alarm problem caused by suspended sediments and river bank exposure in Kuala Lumpur and Bangkok.

[0095] Analysis of the detection results based on the LCI method in the six study areas shows that agricultural activities are the main source of false positives. Specifically, during the crop growth cycle, agricultural activities such as harvesting and crop rotation can temporarily expose farmland. Because the spectral characteristics of exposed farmland are highly similar to those of initially developed land, the LCI method often misidentifies it as construction land development. In addition to Kuala Lumpur, the study areas such as Beijing, Nanjing, Bangkok and Ho Chi Minh City ( Figure 6 ) have large agricultural areas, and agricultural activities in these areas significantly increase the false alarm rate. In addition, seasonal vegetation changes are also an important factor leading to false alarms. This phenomenon is particularly prominent in areas with significant seasonal changes. Vegetation is dense during the growing season, but in winter, the biomass decreases sharply due to leaf fall. Its spectral change characteristics are similar to those of early development plots of construction land, which leads to false detection. For example, seasonal forest changes in some areas of Kuala Lumpur were mistakenly identified as early development plots ( Figure 6 ). Changes in the water environment will also interfere with the detection results. This is mainly reflected in two aspects: First, changes in the concentration of suspended sediments will affect the spectral characteristics of the water body. For example, due to the high concentration of suspended sediments in the Pearl River Estuary in Guangzhou, its spectral characteristics are similar to those of bare land, resulting in false alarms ( Figure 6 ); secondly, the water level drop will expose the riverbank area. For example, some riverbanks in Beijing were mistakenly identified as early development areas by the LCI method due to the water level drop ( Figure 6 ). These factors together increase the false detection rate of LCI-based methods.

[0096] The LCD-OSN method proposed in Example 1 shows significant advantages in reducing the false alarm rate. Figure 7As shown. The method of Example 1 effectively alleviated the interference caused by agricultural activities and seasonal vegetation changes in Beijing and Bangkok, significantly reduced the misjudgment caused by rice field rotation and water sediment changes in Guangzhou, Nanjing and Ho Chi Minh City, and improved the detection accuracy in mountainous and water areas in Kuala Lumpur. This performance improvement is due to the OSN module's ability to collaboratively extract and fuse the characteristics of the target object and its surrounding scenes, achieving accurate distinction between initial development plots and other land cover changes. Specifically, bare land from different sources has distinguishable texture features: agricultural bare land usually presents regular textures such as ridges and terraces, while construction land development areas show relatively uniform surface features. More importantly, construction land development has a unique spatial distribution pattern, which tends to be distributed around existing roads and buildings. This feature is significantly different from other land cover changes around natural landscapes, farmland or water bodies. The LCD-OSN method proposed in Example 1 processes the target image and scene image separately through parallel deep learning encoders, effectively capturing the subtle texture features of the target object and its spatial correlation with the surrounding environment. This dual-branch architecture not only enhances the recognition capability of early development plots of construction land, but also successfully solves the misjudgment problem caused by spectral similarity.

[0097] The scene image plays a key role in improving the detection performance of the LCD-OSN method of Example 1. By providing rich spatial context information, it helps to accurately distinguish the spectral confusion between the early development plots and the bare land after crop harvest. Figure 8 As shown in the comparative experiments, compared with using only object images, Figure 8 (a) and Figure 8 (b) shows that combining the object and the optimal scale scene image significantly improves the detection accuracy, as shown in Figure 8 As shown in (c)-(d), this is especially true in suburban areas where agricultural activities and vegetation dynamics have a significant impact. This performance improvement is mainly due to the spatial distribution feature information provided by the scene image. Construction land development usually shows spatial proximity to existing roads and buildings, while bare land caused by agricultural activities is often surrounded by farmland or natural vegetation. Traditional LCI methods, which rely only on spectral features, have difficulty distinguishing between these situations with very different spatial distribution patterns but similar spectral characteristics. For example, bare land caused by seasonal vegetation changes and initially developed plots may have highly similar temporal spectral changes, but there are obvious differences in spatial distribution. The LCD-OSN method can accurately identify these complex situations by deeply fusing object and scene spatial features, thereby significantly reducing the false detection rate and improving detection reliability.

[0098] The above embodiments and descriptions are only for explaining the principles and best embodiments of the present invention. Without departing from the spirit and scope of the present invention, the present invention may be subject to various changes and improvements, which shall fall within the scope of the invention to be protected.

Claims

1. A method for detecting land parcels in the initial stage of construction by integrating remote sensing spatiotemporal spectral features, characterized in that: The steps include: S1) Obtaining temporal remote sensing images of the land parcel; S2) Using the LCD of the land clearing plot detector to calculate the bands of remote sensing images of different phases, based on the spectral feature variation caused by land clearing, to achieve preliminary screening of potential change plots; S3), constructing a scene image for each detected object; S4), building a land parcel detection model based on an object-scene convolutional neural network and training the land parcel detection model; S5) Use a dual-branch deep convolutional neural network based on an object-scene convolutional neural network to extract dual spatial features, eliminate pseudo-change areas caused by agricultural activities and vegetation phenological changes through decision-level fusion of feature space, and finally output a high-precision map of initial development plots of construction land.

2. The method for detecting land parcels in the initial stage of construction by integrating remote sensing spatiotemporal spectral features according to claim 1, characterized in that: In step S2), the land clearing plot detector LCD performs preliminary screening of potential construction land development plots by calculating two remote sensing images of different phases, and defines the bare soil index RBLI according to the spectral characteristics of the bare soil, namely: Where, Band red Indicates the top of atmosphere reflectance or surface reflectance of the red band; Band blue Represents the top-of-atmosphere reflectance or surface reflectance in the blue band.

3. The method for detecting land parcels in the initial stage of construction by integrating remote sensing spatiotemporal spectral features according to claim 2, characterized in that: In step S2), during the initial development of the construction land, the original surface cover is cleared and the land cover type is converted to bare soil. This change is reflected in the remote sensing image as the bare land index (RBLI) of the second phase image is significantly higher than that of the first phase image. Therefore, the LCI value is defined as: Where, RBLI t1 and RBLI t2 The bare earth index was calculated from the remote sensing images of the first period and the second period respectively.

4. The method for detecting land parcels in the initial stage of construction by integrating remote sensing spatiotemporal spectral features according to claim 1, characterized in that: In step S3), a minimum bounding rectangle is constructed for each detected object, and then the minimum bounding rectangle is expanded by introducing a scale parameter N to generate scene rectangles of different spatial ranges.

5. The method for detecting land parcels in the initial stage of construction by integrating remote sensing spatiotemporal spectral features according to claim 1, characterized in that: In step S5), the object-scene convolutional neural network-based land parcel detection model OSN includes three branches; the three branches are used to process the object image, the resampling ratio, and the scene image, respectively; the outputs of the three branches are fused through the cascade layer and then generated into a binary classification result through the fully connected layer to determine whether the object is an initial development plot of construction land, and a dropout layer is added to the fully connected layer to reduce overfitting.

6. The method for detecting land parcels in the initial stage of construction by integrating remote sensing spatiotemporal spectral features according to claim 5, characterized in that: In step S5), the branch for processing the object image includes a convolution block, four residual blocks, and a spatial pyramid pooling SPP; the branch for processing the scene image includes a convolution block, seven residual blocks, and a spatial pyramid pooling SPP.

7. The method for detecting land parcels in the initial stage of construction by integrating remote sensing spatiotemporal spectral features according to claim 6, characterized in that: In step S5), the convolution block includes a 3×3 convolution layer and a maximum pooling layer, and the 3×3 convolution layer and the maximum pooling layer are processed by batch normalization BN and ReLU activation function.

8. The method for detecting land parcels in the initial stage of construction by integrating remote sensing spatiotemporal spectral features according to claim 6, characterized in that: In step S5), the residual block includes two 3×3 convolutional layers, followed by batch normalization (BN) and ReLU activation functions, and the residual block introduces skip connections to enable the network to learn residual mapping; The expression of the residual block is: H(x)=x+F(x) (3) Where H(x) represents the output of the residual block, x represents the input, and F(x) represents the residual function; The expression of the residual function F(x) is: F(x)=ReLU(BN(Conv3×3(ReLU(BN(Conv3×3(x))))))) (4) Where Conv3×3(x) represents a 3×3 convolutional layer that processes the input x, BN represents batch normalization, and ReLU is the activation function.

9. The method for detecting land parcels in the initial stage of construction by integrating remote sensing spatiotemporal spectral features according to claim 6, characterized in that: In step S5), the spatial pyramid pooling (SPP) divides the feature map into multiple grids of different scales and performs a pooling operation on each grid.

10. The method for detecting land parcels in the initial stage of construction by integrating remote sensing spatiotemporal spectral features according to claim 5, characterized in that: In step S5), the resampling ratio branch is processed by the flattening layer and calculated based on the relationship between the original size of the object image and the resampling size when input to the object-scene based convolutional neural network. The resampling ratio is expressed as a pair of resampling factors (R x ,R y ), the resampling factor (R x ,R y ) is expressed as: Where h i 、w i The height and width H of the minimum bounding rectangle of the object O shape are respectively input 、W input They represent the resampled height and width when input to the object-scene based convolutional neural network.

Images