Rural house idling intelligent discrimination method and system based on multi-source data fusion

Abstract

The invention relates to the field of image recognition, in particular to an intelligent rural house idling judgment method and system based on multi-source data fusion. The method comprises the following steps: obtaining a multi-source remote sensing satellite image, carrying out multi-time-point fitting, and constructing a multi-time-phase image set; performing farm house target identification and refined contour segmentation on the multi-temporal image set, and marking a plurality of farm house image frames; performing time sequence tracking segmentation based on the plurality of rural house image frames to obtain a multi-temporal image frame of each image frame; carrying out time sequence idle evolution mining based on the multi-temporal image frame, and generating an idle evolution track of each farm house; and carrying out idle state trend analysis according to the idle evolution track, and carrying out intelligent discrimination to obtain an idle house identification result. According to the invention, idle rural houses can be quickly and accurately identified, and the rural house resource management efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the field of image recognition, and in particular to a method and system for intelligently identifying idle farmhouses based on multi-source data fusion. Background Art

[0002] Compared with traditional rural housing survey methods, remote sensing image recognition technology offers advantages such as high efficiency, wide coverage, non-contact, and periodic updates. It can automatically extract and analyze changes in the spatial status of rural housing over large scales. Especially with the increasing availability of multi-source remote sensing data, the integration of multi-type and multi-temporal image data, such as optical satellite imagery, high-resolution imagery, multispectral imagery, and historical time series data, not only allows for more comprehensive acquisition of rural housing spatial morphology information, but also allows for deeper exploration of its temporal variations, providing more reliable technical support for identifying rural housing usage.

[0003] However, the field of identifying vacant rural houses still faces numerous technical bottlenecks. On the one hand, traditional methods often rely on manual field surveys or image classification based on a single time point. Their results are easily influenced by subjective experience and data timeliness, making dynamic updates and large-scale promotion difficult. On the other hand, existing image recognition models mostly focus solely on target detection or single-phase classification, lacking the ability to model the temporal evolution of rural house status. This makes it difficult to address the complex, slowly changing, and easily confused realities of rural house usage. Furthermore, the diverse morphology of rural houses, complex roof materials, and the susceptibility to interference from the surrounding environment pose significant challenges to traditional image analysis methods. Summary of the Invention

[0004] In order to solve the above technical problems, the present invention proposes a method and system for intelligently identifying idle rural houses based on multi-source data fusion to solve at least one of the above technical problems.

[0005] To achieve the above objectives, the present invention provides an intelligent identification method for idle farmhouses based on multi-source data fusion, comprising the following steps: Step S1: Acquire multi-source remote sensing satellite images, perform multi-time point fitting, and construct a multi-temporal image set; Step S2: Perform farmhouse target recognition and refined contour segmentation on the multi-temporal image set, and mark multiple farmhouse image frames; Step S3: performing temporal tracking segmentation based on multiple farmhouse image frames to obtain multi-temporal image frames for each image frame; Step S4: performing time-series idleness evolution mining based on the multi-temporal image frames to generate an idleness evolution trajectory of each farmhouse; Step S5: performing idle state trend analysis based on the idle evolution trajectory and performing intelligent discrimination to obtain idle house identification results.

[0006] In this specification, a system for intelligently identifying vacant rural houses based on multi-source data fusion is provided, which is used to execute the above-mentioned intelligent method for identifying vacant rural houses based on multi-source data fusion, including: Image processing module, used to acquire multi-source remote sensing satellite images, perform multi-time point fitting, and construct a multi-temporal image set; The target recognition module is used to identify farmhouse targets and perform refined contour segmentation on multi-temporal image sets, marking multiple farmhouse image frames; The template segmentation module is used to perform temporal tracking segmentation based on multiple farmhouse image frames to obtain multi-temporal image frames for each image frame; an idle evolution analysis module, configured to perform time-series idle evolution mining based on the multi-temporal image frames to generate an idle evolution trajectory for each farmhouse; The intelligent discrimination module is used to perform idle state trend analysis based on the idle evolution trajectory and perform intelligent discrimination to obtain idle house identification results.

[0007] The beneficial effects of the present invention are specifically as follows: by fusing high-resolution optical images, full-color images and multispectral images, not only the spatial resolution and the recognizability of land objects are improved, but also the ability to express the physical characteristics of land objects is enriched, which helps to finely extract details such as the form of farmhouses, roof materials, and traces of use. The multi-temporal image set formed by multi-time point fitting realizes the temporal continuity modeling of the state of farmhouses, so that the model has the ability to capture "changing trends" rather than "single-point judgments", thereby significantly improving the timeliness and accuracy of the recognition of the use status of farmhouses. By performing target detection and contour segmentation in multi-temporal images, the system can automatically mark the location, size, shape, orientation and other information of each farmhouse, avoiding the subjective bias and inefficiency caused by manual drawing. The introduction of refined contour segmentation greatly improves the accuracy of house boundary recognition, and solves the problem that traditional rough frame selection is difficult to cope with complex scenes such as densely distributed farmhouses and irregular shapes. This not only improves the positioning accuracy of individual houses, but also provides accurate spatial anchor points for subsequent change tracking and state evolution; By performing temporal tracking and segmentation of farmhouses within each image frame, the system ensures that each temporal image frame represents the same house, avoiding issues such as target mismatch and recognition drift, and significantly improving the continuity and reliability of temporal analysis. Furthermore, the construction of temporal image frames provides a complete temporal information chain for each house, allowing the system to accurately identify changes in house performance across years, seasons, and even imaging conditions, providing the raw data foundation for subsequent trajectory analysis. By analyzing key indicators such as brightness, texture, vegetation cover, and structural changes in multi-temporal image frames, the system can extract temporal trends in house usage. For example, a continuous decrease in roof reflectivity and an increase in surrounding vegetation may indicate "gradual vacancy." A sudden change in roof color or disappearance of the structure may indicate "sudden demolition" or "complete abandonment." These dynamic trajectories not only provide a more robust basis for identifying the current state but also reflect the impact of macro-context factors such as government policies, rural population migration, and land use adjustments on farmhouse usage.

[0008] The results of the entire data processing chain are transformed into decision-making outputs, achieving an intelligent transition from "image information" to "behavior identification." By analyzing the trends of rural housing evolution, the system not only identifies whether a house is currently idle, but also determines its degree of vacancy (e.g., temporary, long-term, or impending) and its mode of vacancy (e.g., intermittent use, structural damage). The intelligent identification component can predict status based on machine learning classifiers (e.g., decision trees, support vector machines) or more advanced time series models (e.g., long-term LSTM, Transformer), making the identification results interpretable and quantifiable. The identified vacant houses are then output as spatial distribution maps and attribute tables, serving rural housing resource management, policy implementation, and grassroots governance. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 This is a schematic flow chart of the steps of an intelligent method for identifying idle farmhouses based on multi-source data fusion according to the present invention; Figure 2 Detailed implementation flow chart of step S1; Figure 3 Detailed implementation flow chart of step S2; Figure 4 Schematic diagram of the detailed implementation steps of step S3. DETAILED DESCRIPTION

[0010] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0011] This application provides a method and system for intelligently identifying vacant farmhouses based on multi-source data fusion. The execution entities of this method and system include, but are not limited to, the following: mechanical equipment, data processing platforms, cloud server nodes, network upload devices, etc., which can be considered general computing nodes of this application. The data processing platform includes, but is not limited to, at least one of an audio and video management system, an information management system, and a cloud data management system.

[0012] See also Figures 1 to 4 , the present invention provides a method comprising the following steps: Step S1: Acquire multi-source remote sensing satellite images, perform multi-time point fitting, and construct a multi-temporal image set; Step S2: Perform farmhouse target recognition and refined contour segmentation on the multi-temporal image set, and mark multiple farmhouse image frames; Step S3: performing temporal tracking segmentation based on multiple farmhouse image frames to obtain multi-temporal image frames for each image frame; Step S4: performing time-series idleness evolution mining based on the multi-temporal image frames to generate an idleness evolution trajectory of each farmhouse; Step S5: performing idle state trend analysis based on the idle evolution trajectory and performing intelligent discrimination to obtain idle house identification results.

[0013] In the embodiment of the present invention, see Figure 1 , is a flowchart of the steps of a method for intelligently identifying idle farmhouses based on multi-source data fusion according to the present invention. In this example, the steps of the method for intelligently identifying idle farmhouses based on multi-source data fusion include: Step S1: Acquire multi-source remote sensing satellite images, perform multi-time point fitting, and construct a multi-temporal image set; In this embodiment, multi-source remote sensing satellite imagery data is acquired for the study area (e.g., a target rural area) to ensure that the spatial detail, spectral information, and temporal coverage meet the requirements for identifying vacant rural housing. Multi-source data typically includes high-resolution optical imagery, multispectral imagery, and panchromatic imagery. High-resolution optical imagery can clearly depict the shape and structure of houses, multispectral imagery can provide spectral reflectance characteristics of features such as vegetation and water bodies, and panchromatic imagery offers higher resolution in spatial detail. Data acquisition should ensure that the time span covers the target analysis period, for example, quarterly or monthly data acquisition for 2-3 consecutive years. During the data preprocessing phase, the images undergo radiometric calibration, atmospheric correction, and geometric correction to eliminate brightness differences and geometric distortion between images at different times. Subsequently, the images at each time point are aligned according to the spatial registration results, and a temporally continuous and spatially consistent multi-temporal image set is generated using multi-time point fitting methods (such as time series image fusion and interpolation reconstruction).

[0014] Step S2: Perform farmhouse target recognition and refined contour segmentation on the multi-temporal image set, and mark multiple farmhouse image frames; In this example, after constructing a multi-temporal image set, it is necessary to identify the location of farmhouses and accurately determine their outlines. Specifically, a deep learning object detection algorithm (such as the YOLO series or Faster R-CNN) can be used, trained on a large number of labeled farmhouse samples, enabling the model to automatically identify farmhouses of varying shapes and roof materials. The detection output is a rectangular bounding box (image frame) for the farmhouse, each containing coordinates and a confidence score. To improve spatial accuracy, the detection results require refined contour segmentation. Semantic segmentation networks (such as U-Net or DeepLab) can be used to perform pixel-level segmentation of the farmhouse area to obtain accurate house boundaries. In experiments, optical images with a spatial resolution better than 1 meter are used as input to ensure clear visibility of roof structural features. Furthermore, duplicate detection results for the same farmhouse in images taken at different times should be deduplicated to avoid repeated labeling. The final output is a farmhouse target set containing spatial location, boundary shape, and confidence information, covering all identified farmhouses in the study area.

[0015] Step S3: performing temporal tracking segmentation based on multiple farmhouse image frames to obtain multi-temporal image frames for each image frame; In this embodiment, after obtaining multiple farmhouse image frames, their positions within the multi-temporal image set need to be aligned and tracked to extract image frames of the same farmhouse at different times. The core of temporal tracking segmentation is to establish a correspondence between spatial position and temporal sequence. Common methods include center point position-based matching, IoU (Intersection over Union) overlap detection, optical flow, or feature point matching. In experiments, a method combining center point position and area stability is preferred. Specifically, if the farmhouse detection frame at a certain time point is within 2 pixels of the center point at the previous moment and the area change does not exceed 20%, it is considered the same object. After temporal tracking, the image frames of the same farmhouse at different times are cropped to form a temporal sequence of image frames. This process also requires handling occlusion and missing frames. For example, when cloud cover or shadows affect recognition, interpolation or frame infilling can be used to maintain sequence continuity. Ultimately, each farmhouse corresponds to a temporally ordered multi-temporal image frame sequence, providing direct input for subsequent state change analysis.

[0016] Step S4: performing time-series idleness evolution mining based on the multi-temporal image frames to generate an idleness evolution trajectory of each farmhouse; In this embodiment, after obtaining multi-temporal image frames for each farmhouse, the next step is to analyze their temporal variation characteristics to determine whether there are signs of long-term vacancy. Temporal vacancy evolution mining typically involves three aspects: analysis of brightness changes, texture feature changes, and surrounding activity changes. Brightness changes can reflect information such as roof material aging and changes in light reflection; texture changes can capture changes in the house's surface structure or traces of human activity in the environment; and surrounding activity reflects changes in farming, vehicle access, and other conditions around the house. In experiments, metrics such as average brightness, gray-level co-occurrence matrix features, and vegetation index are calculated for each temporal image frame. These metrics are then subjected to time series fitting and trend calculation. If a farmhouse exhibits very low changes over a long period of time and lacks signs of use, it may exhibit signs of vacancy. By concatenating the analysis results at each time point, a vacancy evolution trajectory for each farmhouse can be generated: a curve or data series reflecting the changes in the house's status over time.

[0017] Step S5: performing idle state trend analysis based on the idle evolution trajectory and performing intelligent discrimination to obtain idle house identification results.

[0018] In this embodiment, trend analysis of idleness evolution trajectories is performed, combined with an intelligent discrimination model, to ultimately output vacant housing identification results. The goal of trend analysis is to extract long-term patterns from the evolution trajectories and distinguish between short-term periods of unoccupied space and true, long-term vacancy. In experiments, the trajectories are analyzed on a sliding scale, either quarterly or semi-annually, to calculate the intensity of the vacancy trend for each time period. If a farmhouse maintains a high vacancy trend intensity for multiple consecutive time periods, with no significant decrease in the trend, it is considered to be a highly likely vacant house. To further improve discrimination accuracy, a machine learning classifier (such as a random forest, support vector machine, or lightweight neural network) can be introduced. Features such as brightness change rate, texture stability, and changes in surrounding activity are input into the model, which outputs a classification result for vacant or non-vacant housing. The final result is accompanied by a confidence score, which is combined with a preset time threshold to screen out vacant houses that truly meet the identification criteria. The output dataset will contain information such as house ID, spatial location, vacancy duration, and vacancy trend, which can be directly used for subsequent visualization and management.

[0019] In this embodiment, refer to Figure 2 , is a flowchart of the detailed implementation steps of step S1. In this embodiment, the detailed implementation steps of step S1 include: Acquiring multi-source remote sensing satellite images of the target rural area at different times, wherein the multi-source remote sensing satellite images include high-resolution optical images, multispectral images, and panchromatic images; Calculating the pixel grayscale values ​​of the multi-source remote sensing satellite image to obtain a grayscale histogram; Performing grayscale value distribution stretching processing on the grayscale histogram to obtain a stretched grayscale density function; Performing global histogram equalization processing on the multi-source remote sensing satellite image according to a stretched grayscale density function to construct a brightness enhanced image; Multi-time point fitting is performed on the brightness enhanced image to construct a multi-temporal image set.

[0020] In this example, multi-source remote sensing satellite data with a long time span is acquired for the study area (e.g., a typical rural area in a province) to dynamically assess the status of vacant farmhouses. "Multi-source" refers to the simultaneous use of imagery from different satellite platforms or sensors, such as high-resolution optical imagery (e.g., GF-2 and WorldView-3, which provide visible light data at 0.5 m or higher resolution), multispectral imagery (e.g., Sentinel-2 MSI, which provides multispectral information with 1-3 bands and a resolution of 10-60 m), and panchromatic imagery (e.g., Gaofen-1 or IKONOS, which provides 0.5 m panchromatic data). The recommended time span is at least 2-3 years, covering multiple key agricultural seasons (e.g., spring plowing, summer harvest, and post-autumn harvest) to observe seasonal and annual trends. When acquiring imagery, cloud cover should be controlled (typically, less than 10%). Imagery should be collected over the same region, within the same season, or in similar months whenever possible to minimize the impact of seasonal vegetation changes on the results. Image download sources include the China Resources Satellite Application Center (CRESDA), ESA's Copernicus Data Center, and commercial data providers. After data acquisition, a unified projection coordinate system (such as WGS-84 / UTM 50N) and geometric correction are required to ensure accurate spatial alignment of multi-temporal images. This phase provides the basic data foundation for subsequent grayscale calculation and brightness enhancement. Grayscale value distribution information for pixels is extracted from the imagery. For optical and multispectral imagery, band selection and radiometric calibration are first required to normalize the raw values ​​acquired from different sources and sensors so that they can be compared within the same grayscale range. For example, 8-bit imagery has a grayscale value range of 0 to 255; 16-bit imagery requires scaling or normalization to map it to this range. Multispectral imagery typically calculates grayscale values ​​independently for each band, while panchromatic imagery uses the brightness values ​​of a single band. A grayscale histogram is then generated by counting the number of pixels corresponding to each grayscale value in the image. The horizontal axis of the histogram represents grayscale levels, while the vertical axis represents the corresponding number of pixels or frequency. This statistical process can be performed on the entire image or divided into multiple regions or blocks, for example, each 256×256 pixel unit. This allows for local brightness enhancement in subsequent processing. By observing the histogram, we can intuitively understand the image brightness distribution. For example, farmland and vegetation in rural areas tend to be concentrated in the darker grayscale range, while hard objects such as roofs and roads are generally concentrated in the brighter grayscale range.

[0021] The purpose of grayscale value stretching is to expand the dynamic range of image brightness, making images with limited contrast visually clearer and making it easier to distinguish different features during computation. The grayscale histogram is first analyzed to identify the lower and upper grayscale boundaries. Pixel values ​​below the lower limit are then raised to the darkest end, while those above the upper limit are compressed to the brightest end. Pixel values ​​in the middle are then evenly distributed across the available grayscale range. For example, if an image has a large concentration of pixels between grayscales 40 and 210, with very few pixels below 40 or above 210, the pixels in the middle can be stretched evenly across the entire brightness range, starting at 40 for the dark portion and ending at 210 for the bright portion. In multispectral image processing, this operation is typically performed on each band separately to prevent color distortion. For panchromatic imagery, the brightness values ​​of a single band are directly stretched. Experiments can also employ segmented or nonlinear stretching to enhance detail in either dark or bright areas. After stretching, the image's brightness distribution becomes more uniform, enhancing the contrast between features and providing more suitable input for the next step, histogram equalization. The core purpose of global histogram equalization is to improve the overall contrast of the image, balancing the number of pixels across different grayscale ranges and thus enhancing detail. Based on the stretched grayscale distribution, the image's brightness values ​​are redistributed, boosting dark areas, compressing bright areas, and achieving smoother overall brightness transitions. This approach first calculates the frequency of occurrence of each grayscale value in the image, then maps the original brightness values ​​to new brightness levels based on the cumulative frequency, resulting in a more uniform brightness distribution in the output image. To avoid color bias in multispectral image processing, the image can first be converted to a color space (such as HSI or Lab), then equalized only the brightness component before being recombined with the original chrominance components. Compared to local equalization, global equalization is more suitable for large areas with more uniformly distributed features, such as rural areas. It can improve overall contrast without causing local over-enhancement. After global equalization, the brightness differences of different landforms such as house roofs, roads, farmlands, and vegetation are clearer, and the boundaries are easier to identify, providing a more reliable visual and data basis for the identification of vacant rural houses based on time series changes.

[0022] Brightness-enhanced images at different time points are processed into a time series to form a multi-temporal image set, which is used to analyze the temporal trends of rural housing. The spatial registration accuracy of the images must be ensured, ensuring perfect pixel-level alignment across all temporal images. This typically requires a registration error of no more than one pixel. Algorithms based on feature point matching or cross-correlation methods based on image block similarity can be used during the registration process. After spatial registration, radiometric normalization is required to eliminate brightness differences between images at different times. This can be achieved by selecting targets with stable reflective properties (such as roads or concrete pavement) as references. Subsequently, a time series is created for each pixel position, recording the brightness changes at each point in time. Stable trends can be extracted through smoothing or fitting methods (such as trend analysis or sliding window averaging). The resulting multi-temporal image set not only contains the enhanced images at each time point but also generates a difference map reflecting brightness changes, which can be used to identify areas with minimal long-term brightness changes. These areas may correspond to vacant or long-unused rural housing. This result directly serves as important input data for the intelligent rural housing vacancy identification model.

[0023] In this embodiment, the specific steps of performing multi-time point fitting on the brightness enhanced image to construct a multi-temporal image set are as follows: Perform atmospheric blur recognition on brightness-enhanced images and mark atmospheric blur areas; Perform dark target visual analysis on brightness-enhanced images and extract multiple dark targets in the images; Calculating digital pixel values ​​of dark targets in the image; averaging the digital pixel values ​​and performing distribution analysis to obtain atmospheric scattering parameters; Perform atmospheric scattering compensation on the brightness enhanced image according to the atmospheric scattering parameters to construct a scattering compensated image; Identify the geometric distortion of the scatter compensation image and perform adaptive precision correction to construct a distortion-corrected image; Multi-time point fitting is performed on the distortion-corrected images to construct a multi-temporal image set.

[0024] In this embodiment, even after brightness enhancement, images may still be affected by atmospheric conditions. This is particularly true for multi-source, multi-temporal data. Different capture times and climate conditions can lead to blurring caused by airborne moisture, suspended particles, and other factors. This can reduce image sharpness and contrast, thereby impacting the accuracy of subsequent analysis. By identifying and marking areas affected by atmospheric blurring, targeted compensation can be performed in subsequent processing. The identification process can first detect the degree of blurring using spatial frequency analysis. For example, after high-pass filtering the image, a local clarity index (such as a clarity score based on gradient amplitude) is calculated. Areas below a set threshold are identified as blurry. Furthermore, contrast analysis can be combined to calculate the brightness variation range of local areas. Excessively small variations may indicate a smoothing effect caused by blurring. The threshold setting can be dynamically adjusted based on the statistical results of a reference clear image. For example, in a high-resolution image with a resolution of 0.5 meters, if the local contrast falls below 70% of the normal value, the image is marked as blurry. This marking method typically generates a binary mask image, with white representing blurry areas and black representing clear areas. This mask serves as a reference weight in the subsequent atmospheric scattering compensation process, enabling stronger compensation correction in blurry areas.

[0025] Atmospheric scattering compensation requires reference to "dark target" information within the image. Dark targets are objects with extremely low reflectivity in the image, theoretically indicating near-zero brightness. These include deep water, shadowed areas, and shadows cast by dense vegetation. In optical remote sensing images, these areas are normally very dim, but due to atmospheric scattering, they still appear bright. This brightness shift is a characteristic of atmospheric scattering. Therefore, dark target analysis is necessary within the brightness-enhanced image. This approach combines manual experience with automated detection. For example, a brightness threshold is used to identify pixels with brightness values ​​within the lowest 1% or 2% as preliminary dark target candidates. Morphological filtering is then used to remove small noise points. To avoid confusion between shadows and artificial structures, multispectral band information can be combined to eliminate shadow artifacts caused by high-reflectivity materials. For 0.5-meter resolution images of rural areas, at least dozens of dark targets distributed across each image are extracted to ensure spatial representativeness of the sample distribution. These dark targets will be used to estimate atmospheric scattering parameters, so their reflectivity stability and location diversity are crucial.

[0026] After extracting the dark target areas, the digital pixel values ​​in these areas need to be accurately measured. This quantifies the actual brightness offset of these pixels, which theoretically should be close to zero brightness, in the image. During the calculation process, each pixel in the dark target area is first extracted individually, and its grayscale value or radiometric brightness value in the image is recorded. For multispectral imagery, separate statistics are required for each band, as atmospheric scattering has different effects in different bands. For example, scattering in shortwave bands is generally more significant than in longwave bands. To reduce the impact of individual anomalous pixels, the dark target area can first be smoothed using a 3×3 or 5×5 sliding window, with the median of the pixel values ​​within the window taken as the representative value. The pixel values ​​of all dark targets will form a dataset for averaging and distribution analysis.

[0027] After acquiring a dataset of dark object pixel values, these values ​​need to be statistically analyzed to estimate the intensity and distribution characteristics of atmospheric scattering. First, the average brightness value of all dark objects is calculated. This value represents the overall brightness offset introduced by atmospheric scattering in the image. Next, a distribution analysis is performed, using metrics such as the standard deviation and range of the brightness values ​​to determine the uniformity of the scattering effect across different regions. A large standard deviation indicates that the atmospheric scattering effect is unevenly distributed within the image, necessitating the use of a spatially variable correction factor during compensation. A small standard deviation allows for a globally uniform compensation factor. Dark objects are typically analyzed separately at different locations in the image to detect differences in local atmospheric conditions. For example, valleys and open plains may exhibit different scattering characteristics. The atmospheric scattering parameters obtained in this way are directly used to compensate for atmospheric scattering in the image, thereby restoring brightness values ​​that are closer to the true reflectance of the ground objects. The goal of atmospheric scattering compensation is to remove the atmospheric offset in the image brightness, making the image more similar to the true reflectance characteristics of the ground objects. Based on the atmospheric scattering parameters obtained in the previous step, the brightness value of each pixel in the image can be corrected. The specific approach is to deduct the atmospheric brightness offset reflected by the dark target from all pixels in the image. If the scattering parameters in different areas are significantly different, different compensation values ​​are applied according to the spatial position. In multispectral images, the compensation values ​​of different bands may be different, so it is necessary to calculate and apply compensation parameters for each band separately. After compensation, a comparative verification will be performed immediately to confirm whether the compensation effect is reasonable by comparing it with images taken in cloudless and clear weather. The compensated images will have a significant improvement in overall brightness and contrast, especially in long-distance or low-contrast areas. The boundaries of the objects will be clearer, providing a higher quality data basis for geometric correction and subsequent analysis.

[0028] Geometric distortion is a common problem in remote sensing image processing and can be caused by a variety of factors, including sensor imaging geometry, satellite attitude variations, and terrain undulations. After scatter compensation is completed, the images need to be further examined for geometric distortion, especially in multi-temporal comparative analysis, where any pixel-level spatial misalignment can lead to biased results. Methods for identifying geometric distortion include comparison with ground control points (GCPs), pixel-level matching with high-precision reference images, and the use of feature point matching algorithms to detect nonlinear deformations. After distortion is detected, adaptive precision correction techniques are employed, selecting an appropriate geometric transformation model based on the type and distribution of the distortion, such as a polynomial transformation, an affine transformation, or a local thin plate spline transformation. The requirement for corrected registration accuracy is better than one pixel, meaning that for 0.5-meter resolution imagery, the spatial error should not exceed 0.5 meters. Images corrected for geometric distortion have higher spatial accuracy and comparability with other temporal data. After atmospheric compensation and geometric correction, images from different temporal phases need to be fitted and integrated to form a multi-temporal image set. This process requires high consistency in spatial position and radiometric properties between the images to facilitate time series analysis. First, the images of each temporal phase are sorted by capture time, ensuring that all images completely overlap spatially. During the temporal fitting process, a sliding time window method can be used to smooth the brightness variations of a pixel over time, reducing the influence of random noise on the variation trend. To identify vacant farmhouses in rural areas, the brightness variation curves of the built-up areas can be analyzed. Areas with no significant changes over a long period of time may indicate vacancy. The resulting multi-temporal image set includes both high-quality images for each phase and brightness trend data, providing direct input support for intelligent analysis algorithms.

[0029] In this embodiment, refer to Figure 3 , is a flowchart of the detailed implementation steps of step S2. In this embodiment, the detailed implementation steps of step S2 include: Perform pixel-level visual recognition on multi-temporal image sets and classify soil use to obtain soil types in the region; The land types within the said area include farmland, residential land, transportation roads and water vegetation; Detect residential land boundaries based on soil types within the area and extract residential land boundary areas; Perform edge noise filtering on the boundary area of ​​residential land to obtain the boundary area of ​​filtered optimized land; The farmhouse targets are identified in the filtered optimized boundary area, and refined contour segmentation is performed to mark multiple farmhouse image frames.

[0030] In this embodiment, a pixel-level visual recognition method is used to combine spectral, textural, and temporal information to accurately distinguish different land features. Machine learning algorithms based on supervised classification can be selected, such as support vector machines (SVM), random forests (RF), or deep learning convolutional neural networks (CNN). 3D-CNNs with multi-temporal feature inputs are particularly effective in utilizing temporal information. Training samples should cover all expected classification types, including farmland, residential land, roads, water bodies, and vegetation. The number of samples per category should be balanced and distributed across the study area to ensure generalization. Classification input features include not only multispectral reflectance values ​​from a single image, but also derived features such as the Normalized Difference Vegetation Index (NDVI) and the Normalized Water Index (NDWI), as well as image texture features (such as contrast and entropy of the Gray Level Co-occurrence Matrix (GLCM)). Classification accuracy can be assessed using confusion matrices and the Kappa coefficient, with an overall accuracy requirement of at least 85% to ensure the reliability of subsequent residential land extraction. After land use classification is complete, the residential land category can be filtered out from the classification results, and boundary detection can be performed to determine the precise extent of residential land. Pixels marked as residential land in the classification map are extracted to form a binary mask, where residential land pixels are marked as 1 and non-residential land pixels are marked as 0. Because the classification results may contain noise or small areas of misclassification, morphological processing is required first. This includes opening operations to remove small isolated pixel blocks and closing operations to fill small holes within residential land. Connected domain analysis is then used to identify continuous residential land parcels, and noisy areas that are too small are filtered out based on an area threshold. For 0.5-meter resolution imagery, the minimum area for identifying residential land can be set to 50 square meters to avoid misidentifying individual small structures or temporary sheds as residential land. Boundary detection can use contour extraction algorithms, such as Canny edge detection combined with contour tracing, to ultimately obtain a closed boundary polygon for each residential land parcel.

[0031] Because image classification and boundary detection are affected by the complexity of object textures, image noise, and lighting conditions, the extracted residential land boundaries may exhibit jagged, burred, or partially misaligned noise. Boundaries are smoothed and optimized to obtain continuous, regular boundaries that more closely resemble the actual object contours. Vertex simplification can be performed on boundary polygons, for example, using the Douglas-Peucker algorithm to reduce redundant nodes while retaining key shape features. Next, the boundary lines are smoothed using median filtering or Gaussian smoothing to eliminate jitter caused by high-frequency noise. To avoid loss of detail due to over-smoothing, the smoothing radius is typically kept within a range of 1 to 3 pixels. Furthermore, a boundary optimization method based on region growing can be used, incorporating brightness gradient information from the original image to locally adjust the boundary so that it more closely matches the edge of the actual building or plot. The specific location and shape of the farmhouses within the area can then be identified. Farmhouse object recognition can be achieved using an object-oriented approach: first, the image within the residential land is segmented into several image objects with similar spectral and textural characteristics, and then classified based on shape, color, and texture features. In experiments, farmhouses typically exhibit regular geometric shapes, high brightness (rooftops have strong reflectivity in full-color images), and low vegetation index values. These characteristics can serve as a basis for identification. If using deep learning methods, detection models such as Faster R-CNN and YOLOv5 can be selected and trained on labeled farmhouse samples. After detecting farmhouse targets, refined contour segmentation is required. This step can use edge-detection-based active contour models (Snake models) or graph cut methods to ensure that the segmentation results closely match the actual farmhouse boundaries. Finally, each identified farmhouse is marked with a rectangular or polygonal box, and its location, area, and outline information is recorded.

[0032] In this embodiment, reference Figure 4 , is a flowchart of the detailed implementation steps of step S3. In this embodiment, the detailed implementation steps of step S3 include: Recognize repeated frame selection based on multiple farmhouse image frames; Perform non-maximum suppression on repeatedly selected image frames to obtain multiple high-confidence image frames; Calculating the spatial position coordinates of the high-confidence image frame; The multi-temporal image set is subjected to time-series tracking processing based on the spatial position coordinates, and time-series segmentation is performed on the same position to obtain a multi-temporal image frame for each image frame.

[0033] In this embodiment, after completing the identification of farmhouses, there may be a large number of repeated frame selection areas in the preliminary detection results, that is, the detection model may give multiple overlapping detection boxes for the same farmhouse target. This repetition is more common when using deep learning detection algorithms (such as YOLO, Faster R-CNN), because the model may detect the same target at different feature layers and different anchor box sizes. The goal of this step is to identify those repeated frame selection image boxes from multiple candidate boxes for subsequent de-duplication optimization. The specific method is first to group all detection boxes by category and detection confidence in the same image, and only the detection boxes of the farmhouse category are judged for repetition. Then, calculate the overlap (intersection over union, IOU) of any two detection boxes, if the IOU is greater than the set repetition judgment threshold (0.5 to 0.7 in the experiment, adjust according to the accuracy requirements of the detection task), it is considered that they belong to the same farmhouse repeated frame selection result. In this process, in order to reduce false positives, the center point distance and area difference of the detection box can be combined to avoid misjudging two adjacent independent farmhouses as repeated frames. The non-maximum suppression (NMS) method is used to select the optimal frame from multiple overlapping detection boxes. Non-maximum suppression is a commonly used post-processing method for target detection, its basic idea is: among multiple candidate boxes of the same target, keep the one with the highest confidence, remove other boxes with lower confidence and high overlap. The specific method is to sort the repeated boxes by detection confidence from high to low, first select the box with the highest confidence as the reserved box, then calculate the overlap (IOU) with the remaining boxes, if the IOU exceeds the set suppression threshold (0.5 or 0.6 in the experiment), the box is removed from the candidate set, repeat this process until all boxes are processed. In order to avoid suppressing the actual small target, the suppression threshold can be appropriately relaxed for small area boxes.

[0034] The pixel position parameters of each detection frame in the image are obtained, including the pixel row and column numbers of the top-left corner and the bottom-right corner. Then, the pixel coordinates are accurately converted into geographic coordinates or projection coordinates (such as WGS-84 latitude and longitude or UTM coordinates) using the geographic reference information of the image (usually provided in the metadata or external registration file of remote sensing data, including spatial resolution, projection coordinate system, and geographic transformation parameters). For high-resolution images with a resolution of 0.5 meters, each pixel represents a real ground area of 0.5x0.5 meters, so the accuracy of coordinate conversion can reach sub-meter level. At the same time, it is recommended to record the center point coordinates and boundary polygon coordinates of the detection frame, which is not only beneficial for position matching, but also convenient for subsequent time series segmentation and area change analysis. With the accurate spatial coordinates of the rural house detection frame, position matching and tracking analysis can be performed in a multi-temporal image set. This step first extracts the corresponding image area in each image based on the spatial position coordinates. This process can be achieved through spatial query, i.e., finding the pixel range containing the location in different time images. To ensure temporal consistency, it is necessary to ensure that all images are geometrically fully registered, with a spatial error of no more than one pixel. During the extraction process, the boundary can be appropriately expanded based on the original detection frame, such as increasing 1-2 pixels in each direction, to ensure complete coverage of the rural house and its surrounding area. The extracted image areas are arranged in chronological order to form a multi-temporal image frame set for the rural house. These image frames can be used to analyze the brightness, texture, and shape changes of the rural house at different times, such as by comparing the image frames at different times to determine whether the rural house has been in a long-term unchanged state, thereby inferring its idle state.

[0035] In this embodiment, step S4 includes the following steps: Based on the multi-temporal image frames, the spectral changes of the drying area, the traces of the courtyard soil, and the degree of road wear in front of the door are analyzed to obtain the life trace features of each rural house; The multi-temporal image frames are subjected to multi-frame calculation of the aging rate of the roof material to obtain the aging rate; According to the aging rate, the house maintenance frequency is calculated, and the resident attention level is evaluated to obtain the resident attention level evaluation value; Based on the multi-temporal image frames, the green state around the rural house is generated to generate a vegetation time series index curve; Based on the vegetation time series index curve, the resident attention level evaluation value, and the life trace features, the time series idle evolution is mined to generate the idle evolution trajectory of each rural house.

[0036] In this embodiment, in the multi-temporal image frame analysis, the life traces are an important basis for judging the use activity of the farm house. The possible drying area in the image is located, which usually appears in the courtyard, roof or front yard. Using the reflectivity characteristics of visible light and near-infrared band of multispectral image, the spectral change characteristics of drying objects (such as food, crops, clothes) in different time periods can be detected, for example, the reflectivity decreases in blue and red light bands, and the near-infrared reflectivity slightly rises. Then, the texture analysis and color change detection are performed on the courtyard soil area, the texture feature extraction method based on gray level co-occurrence matrix is used, and the multi-temporal color index change comparison is used to identify the soil compaction, color darkening or texture smoothing phenomenon caused by frequent trampling. For the detection of road wear degree in front of the door, the brightness stability and roughness change of the road area are used to judge the situation of high frequency of vehicles or people flow. In the experiment, a multi-temporal sequence with a time interval of 1 month can be set for detection, and at least one year of change data can be combined for stability analysis. Finally, the drying activity frequency, the significance of courtyard trampling traces, and the road wear change amplitude are integrated to form the life trace feature vector of the farm house, which provides basic information for idle discrimination.

[0037] The aging rate of the roof material is an important indicator for evaluating the use and maintenance of the house. In the multi-temporal image frame, by analyzing the spectral reflectivity change, color shift trend and texture degradation degree of the roof area, the aging speed of the roof can be calculated. First, the roof area boundary is extracted by using high-resolution optical image, and the color mean value, color saturation and brightness distribution characteristics of each frame image are calculated. Material aging usually shows color fading, brightness decrease and surface roughness increase, such as tile darkening, metal roof losing reflectivity. The near-infrared and short-wave infrared bands in multispectral image can more sensitively reflect the structure change of material surface. In the experiment, the roof features at each time point are differentially calculated with the initial state, and the change curve is fitted according to the time sequence, so as to obtain the aging rate of the material (such as the annual brightness decrease ratio, color shift amount). If the aging rate is obviously high and lacks periodic recovery signs, it may mean that the house lacks regular maintenance. This result will be directly used for subsequent maintenance frequency calculation and inhabitant attention evaluation.

[0038] After determining the roof degradation rate, the frequency of house maintenance can be estimated. Specifically, roof maintenance events are identified by observing material state changes in a multi-temporal sequence (e.g., significant brightness increase, color restoration, or smoothing of surface texture). For example, in a 12-month image series, if the roof's color significantly recovers in the seventh month, and its spectral reflectance approaches that of new material, this is considered a maintenance event. Counting the number of maintenance events over the entire analysis period yields a maintenance frequency metric. Maintenance frequency is then combined with the degradation rate to assess resident attention to the house: a high maintenance frequency and a low degradation rate indicate high resident attention; a low aging rate indicates low attention. In experiments, attention levels are categorized as high (>0.8), medium (0.5-0.8), and low (<0.5). The score is based on the normalized maintenance frequency and the inverse of the degradation rate. This resident attention assessment is not only an important indicator for identifying vacant rural houses, but can also be combined with traces of habitation and vegetation status to improve identification accuracy.

[0039] The status of greenery around farmhouses can reflect the management and use of the surrounding environment. Using multispectral imagery, the differences in vegetation reflectance in the red and near-infrared bands can be used to analyze greenery status. The specific method is as follows: First, a buffer zone (e.g., a 10-meter radius) is extracted around the farmhouse to identify vegetation areas within it. Then, for each frame of the multitemporal image, spectral indicators of this vegetation area, such as red reflectance and near-infrared reflectance, are calculated to generate a time-series curve reflecting vegetation growth vitality. For experiments, it is recommended to select data spanning at least one year to cover the complete vegetation growth cycle. For inhabited houses, the greenery around them typically shows signs of periodic pruning and replanting, resulting in a relatively stable vegetation vitality curve with regular peak cycles. However, for long-term vacant houses, the vegetation around them may exhibit irregular rapid growth, withering, or even become covered by weeds, resulting in large fluctuations and a lack of evidence of human intervention. The resulting vegetation time-series index curve serves as an important input for vacancy evolution trajectory analysis.

[0040] The vegetation time-series index curve, household attention assessment values, and life trace characteristics are integrated to conduct temporal vacancy evolution mining. Specifically, the system maps these three features onto a timeline, constructing a multidimensional time-series feature matrix with monthly or quarterly time units. Trend analysis methods (such as sliding window averaging, time series clustering, and change point detection) are then used to identify patterns of change in the status of rural houses. For example, if the frequency of life trace activity decreases, the vegetation curve shows signs of neglect, and household attention remains low for a long time, then the rural house's vacancy trend is significantly increasing. In the experiment, a threshold for the intensity of the vacancy trend can be set. For example, if the comprehensive score remains below 0.4 for six consecutive months, the house is marked as potentially vacant. The resulting vacancy evolution trajectory reflects the changes in use and vacancy of each rural house throughout the entire analysis period, providing data support for subsequent vacancy identification and spatial distribution analysis.

[0041] In this embodiment, the specific steps of step S5 are: Performing multi-time window decomposition on the idle evolution trajectory to obtain multiple time window trajectory segments; Perform idle state trend analysis on multiple time window trajectory segments to generate idle state trends; Conduct progressive vacancy trend enhancement analysis on vacancy status trends and mark houses as potentially vacant; Calculate the idle time of potential vacant houses to obtain the idle time; Comparative analysis of idle time based on preset idle house time thresholds is performed to obtain idle house identification results; Based on the vacant house identification results, highlighted visual rendering is performed to build a visual view of the vacant house distribution to complete the task of identifying vacant rural houses.

[0042] In this embodiment, after extracting multi-temporal farmhouse image frames, each farmhouse corresponds to a time-varying "idleness evolution trajectory." This trajectory records the farmhouse's status characteristics at different time points, such as brightness changes, texture changes, and usage index. Because farmhouse usage status changes may not occur uniformly, this long series needs to be decomposed into multiple shorter time periods for analysis. This segmentation approach is called multi-temporal window decomposition. Specifically, the time window length is set based on the time scale of the research task and the data sampling frequency. For example, a three-month window can be used, or time units such as quarters or half-years can be used. Decomposition ensures that each window is both independent and comparable. Therefore, a sliding window approach can be used, where adjacent windows have a certain degree of overlap (e.g., one month) to capture cross-period status changes. For 2-3 years of multi-temporal image data, a common decomposition strategy is to set a 6-month window and slide it forward 12 months. This approach captures seasonal variations while preserving medium-term trends. After multi-temporal window decomposition, each farmhouse's status trajectory is divided into several segments. Each segment can be analyzed individually, allowing for more detailed capture of the periodic changes between idle and non-idle status. After obtaining multiple temporal window segments, idle status trend analysis is performed on each segment to determine whether the farmhouse exhibited idle characteristics during that period. This analysis can incorporate various characteristic metrics, such as image brightness fluctuations, changes in roof reflectivity, and the level of surrounding activity (e.g., vegetation changes, the presence of temporary shelters). Typical characteristics of idleness are often characterized by small brightness changes, long-term stability in texture features, and a lack of significant changes in signs of use. The rate of change of these features can be calculated for each segment and then compared to a pre-set idle status threshold. For example, an idle trend is identified when both brightness changes fall below a certain percentage and texture changes fall below a fixed value. To improve accuracy, consistency verification can be performed by combining the results from multiple temporal windows. Specifically, if the same farmhouse exhibits an idle trend in two or more consecutive temporal window segments, its idle trend is more reliable. Ultimately, a trend curve of idle status will be generated for each farmhouse to describe its status evolution during the entire study period.

[0043] A progressive idleness trend enhancement analysis is performed. By tracking the continuity and enhancement of rural housing status trends over multiple time windows, farmhouses with a genuine tendency to long-term idleness are screened. Specifically, the intensity of the idleness trend (i.e., the significance of idleness characteristics across each time window) is cumulatively analyzed. If a farmhouse shows a gradual increase in idleness intensity over time, it is identified as potentially idle. A threshold can be set. For example, if the idleness trend index increases or remains stable at a high level over three consecutive time windows, the farmhouse is marked as potentially idle. This progressive analysis helps distinguish between seasonal unoccupied housing and long-term idleness, reducing misjudgments. Farmhouses screened in this step will enter the next stage of idleness duration calculation.

[0044] Traversing multiple time-window trajectory segments for each rural house, the duration of those segments identified as showing an idle trend is summed to obtain the total idle time. If there are gaps between trajectory segments, these gaps should be excluded or counted separately. To ensure calculation accuracy, the duration unit should be consistent with the time window length. For example, if the time window is six months, the idle time is calculated as six months. If a sliding time window is used, the overlapped time should be deducted. For a three-year study period, rural houses with an idle time of more than one year and distributed across multiple time windows are often more likely to be long-term idle. The calculated idle time will serve as one of the key criteria for the final vacant housing identification and will be compared with the time threshold for comparative analysis. After calculating the idle time for each potential vacant house, it is compared with the preset time threshold to determine the final vacant house identification result. This time threshold can be set based on research objectives and policy requirements. For example, setting it to 12 months or 18 months means that only rural houses with an idle time exceeding this threshold will be considered vacant. During comparative analysis, all potentially vacant houses can be categorized as "exceeding a threshold" or "not exceeding a threshold," and an identification label is generated for each farmhouse. To enhance the robustness of the results, a confidence score can be introduced, combining the duration of vacancy and the intensity of the vacancy trend to calculate a score. Only farmhouses with scores exceeding both thresholds are ultimately designated as vacant. A list of vacant houses, including information such as spatial location, vacancy duration, and confidence level, is generated, which can be directly used by management departments for decision-making and subsequent remediation measures. This step completes the entire intelligent identification process for vacant farmhouses based on multi-source remote sensing data.

[0045] In this embodiment, the specific steps of performing highlight visualization rendering based on the vacant house identification results and constructing a vacant distribution visualization view to complete the idle farmhouse identification task are as follows: Calculate the spatial coordinates of each vacant house based on the vacant house identification results; Performing idle distribution analysis based on the spatial coordinates to obtain a regional idle distribution map; Perform highlighted visual rendering on the regional idle distribution map and construct a visual view of idle distribution to complete the idle farmhouse identification task.

[0046] In this embodiment, after vacant houses are identified, a unique identification tag is created for each farmhouse determined to be vacant. The pixel locations of these farmhouses in the image are converted to precise geographic coordinates. The bounding box information for each vacant house in the final detected image is read, including the row and column numbers of the pixels at the upper left and lower right corners, as well as the center point position of the detection box. The pixel coordinates are then converted to projected or geographic coordinates using the remote sensing image's georeferencing information (typically including metadata such as spatial resolution, projection coordinate system, and geographic transformation parameters). For example, for a high-resolution optical image with a spatial resolution of 0.5 meters, the center point pixel coordinates can be accurately determined to within 0.5 meters of the actual location after georeferencing. To improve accuracy, orthorectified image data is preferred, and small-scale geometric error correction is performed during the calculation process to ensure spatial consistency between images taken at different times. Ultimately, each vacant farmhouse will be assigned unique spatial coordinates (center point coordinates) and a corresponding spatial extent (polygonal boundary), providing basic spatial data for the subsequent vacancy distribution analysis. Import all coordinate points into a geographic information system (GIS) platform or spatial analysis software, and perform spatial statistical analysis based on the study area's boundary data. Common methods include nearest neighbor distance analysis (to determine clustering), kernel density estimation (to generate continuous spatial density surfaces), and spatial autocorrelation analysis (such as Moran's I coefficient, which determines spatial correlation). If the study area is large and contains a large number of houses, kernel density estimation can be used, with a search radius of 200 or 500 meters to reflect the concentration of vacant houses within the village or district. Once the analysis is complete, the generated results can be used to generate a regional vacancy distribution map through spatial interpolation or density mapping.

[0047] In a GIS platform or remote sensing visualization software, the results of kernel density or distribution analysis are rendered in a color-coded manner according to numerical ranges. For example, areas with high vacancy density are represented by bright red or orange, medium density by yellow, and low density by light green, highlighting hotspots. Five rendering levels can be set, with the top 20% of density highlighted as the first level. Actual farmhouse locations are overlaid on the map to visualize the correspondence between specific farmhouses and density hotspots. To further enhance readability, background elements such as village boundaries, roads, and water bodies can be added, and zooming in on local details is supported. The resulting visualization can not only be viewed interactively on a computer but can also be exported as a static map or online map service for direct use by relevant departments, township planners, or in scientific research reports, thus completing the visualization of farmhouse vacancy identification.

[0048] In the embodiment, a farmland house idle intelligent discrimination system based on multi-source data fusion is provided, which is used for executing the farmland house idle intelligent discrimination method based on multi-source data fusion as described above, and comprises: An image processing module is configured to acquire multi-source remote sensing satellite images, perform multi-time point fitting, and construct a multi-time image set. A target recognition module is configured to perform farmland house target recognition and fine contour segmentation on the multi-time image set, and mark a plurality of farmland house image frames. A template segmentation module is configured to perform time sequence tracking segmentation based on the plurality of farmland house image frames, and obtain multi-time image frames of each image frame. An idle evolution analysis module is configured to perform time sequence idle evolution mining based on the multi-time image frames, and generate an idle evolution track of each farmland house. An intelligent discrimination module is configured to perform idle state trend analysis according to the idle evolution track, and perform intelligent discrimination, and obtain an idle house recognition result.

[0049] Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting, the scope of the present application is defined by the appended claims rather than the above description, and therefore all changes falling within the meaning and scope of the equivalent elements of the application file are intended to be included in the present application.

[0050] The above description is only a specific embodiment of the present application, enabling those skilled in the art to understand or implement the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein are implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to these embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An intelligent identification method for idle farmhouses based on multi-source data fusion, characterized by: The following steps are involved: Step S1: Acquire multi-source remote sensing satellite images, perform multi-time point fitting, and construct a multi-temporal image set; Step S2: Perform farmhouse target recognition and refined contour segmentation on the multi-temporal image set, and mark multiple farmhouse image frames; Step S3: performing temporal tracking segmentation based on multiple farmhouse image frames to obtain multi-temporal image frames for each image frame; Step S4: performing time-series idleness evolution mining based on the multi-temporal image frames to generate an idleness evolution trajectory of each farmhouse; Step S5: performing idle state trend analysis based on the idle evolution trajectory and performing intelligent discrimination to obtain idle house identification results.

2. The intelligent identification method for idle rural houses based on multi-source data fusion according to claim 1 is characterized in that: The specific steps of step S1 are: Acquiring multi-source remote sensing satellite images of the target rural area at different times, wherein the multi-source remote sensing satellite images include high-resolution optical images, multispectral images, and panchromatic images; Calculating the pixel grayscale values ​​of the multi-source remote sensing satellite image to obtain a grayscale histogram; Performing grayscale value distribution stretching processing on the grayscale histogram to obtain a stretched grayscale density function; Performing global histogram equalization processing on the multi-source remote sensing satellite image according to a stretched grayscale density function to construct a brightness enhanced image; Multi-time point fitting is performed on the brightness enhanced image to construct a multi-temporal image set.

3. The intelligent identification method for idle rural houses based on multi-source data fusion according to claim 2 is characterized in that: The specific steps of performing multi-time point fitting on the brightness enhanced image to construct a multi-temporal image set are as follows: Perform atmospheric blur recognition on brightness-enhanced images and mark atmospheric blur areas; Perform dark target visual analysis on brightness-enhanced images and extract multiple dark targets in the images; Calculating digital pixel values ​​of dark targets in the image; averaging the digital pixel values ​​and performing distribution analysis to obtain atmospheric scattering parameters; Perform atmospheric scattering compensation on the brightness enhanced image according to the atmospheric scattering parameters to construct a scattering compensated image; Identify the geometric distortion of the scatter compensation image and perform adaptive precision correction to construct a distortion-corrected image; Multi-time point fitting is performed on the distortion-corrected images to construct a multi-temporal image set.

4. The intelligent identification method for idle rural houses based on multi-source data fusion according to claim 1 is characterized in that: The specific steps of step S2 are: Perform pixel-level visual recognition on multi-temporal image sets and classify soil use to obtain soil types in the region; The land types within the said area include farmland, residential land, transportation roads and water vegetation; Detect residential land boundaries based on soil types within the area and extract residential land boundary areas; Perform edge noise filtering on the boundary area of ​​residential land to obtain the boundary area of ​​filtered optimized land; The farmhouse targets are identified in the filtered optimized boundary area, and refined contour segmentation is performed to mark multiple farmhouse image frames.

5. The intelligent identification method for idle rural houses based on multi-source data fusion according to claim 1 is characterized in that: The specific steps of step S3 are: Recognize repeated frame selection based on multiple farmhouse image frames; Perform non-maximum suppression on repeatedly selected image frames to obtain multiple high-confidence image frames; Calculating the spatial position coordinates of the high-confidence image frame; The multi-temporal image set is subjected to time-series tracking processing based on the spatial position coordinates, and time-series segmentation is performed on the same position to obtain a multi-temporal image frame for each image frame.

6. The intelligent identification method for idle rural houses based on multi-source data fusion according to claim 1 is characterized in that: The specific steps of step S4 are: Based on the multi-temporal image frames, the spectrum changes in the drying area, the trampling marks on the courtyard soil, and the wear degree of the road in front of the door are analyzed to obtain the life trace characteristics of each farmhouse; Performing multi-frame calculation of the roof material aging rate on the multi-temporal image frames to obtain the aging rate; Calculate the frequency of house maintenance based on the aging rate and evaluate the resident attention to obtain the resident attention evaluation value; The greening state around the farmhouse is analyzed based on the multi-temporal image frames to generate a vegetation time series index curve; Based on the vegetation time-series index curve, the residents' attention evaluation value and the characteristics of the living traces, the time-series idle evolution mining is carried out to generate the idle evolution trajectory of each farmhouse.

7. The intelligent identification method for idle rural houses based on multi-source data fusion according to claim 1 is characterized in that: The specific steps of step S5 are: Performing multi-time window decomposition on the idle evolution trajectory to obtain multiple time window trajectory segments; Perform idle state trend analysis on multiple time window trajectory segments to generate idle state trends; Conduct progressive vacancy trend enhancement analysis on vacancy status trends and mark houses as potentially vacant; Calculate the idle time of potential vacant houses to obtain the idle time; Comparative analysis of idle time based on preset idle house time thresholds is performed to obtain idle house identification results; Based on the vacant house identification results, highlighted visual rendering is performed to build a visual view of the vacant house distribution to complete the task of identifying vacant rural houses.

8. The intelligent identification method for idle farmhouses based on multi-source data fusion according to claim 7 is characterized in that: The specific steps of performing highlight visualization rendering based on the idle house identification results and constructing an idle distribution visualization view to complete the idle farmhouse identification task are as follows: Calculate the spatial coordinates of each vacant house based on the vacant house identification results; Performing idle distribution analysis based on the spatial coordinates to obtain a regional idle distribution map; Perform highlighted visual rendering on the regional idle distribution map and construct a visual view of idle distribution to complete the idle farmhouse identification task.

9. An intelligent identification system for idle farmhouses based on multi-source data fusion, characterized by: The method for executing the intelligent identification method of idle rural houses based on multi-source data fusion according to claim 1 comprises: Image processing module, used to acquire multi-source remote sensing satellite images, perform multi-time point fitting, and construct a multi-temporal image set; The target recognition module is used to identify farmhouse targets and perform refined contour segmentation on multi-temporal image sets, marking multiple farmhouse image frames; The template segmentation module is used to perform temporal tracking segmentation based on multiple farmhouse image frames to obtain multi-temporal image frames for each image frame; an idle evolution analysis module, configured to perform time-series idle evolution mining based on the multi-temporal image frames to generate an idle evolution trajectory for each farmhouse; The intelligent discrimination module is used to perform idle state trend analysis based on the idle evolution trajectory and perform intelligent discrimination to obtain idle house identification results.

Images