A method for monitoring farmland non-agriculturalization based on remote sensing technology

Abstract

The application discloses a kind of based on remote sensing technology's cultivated land non-agricultural monitoring method, it is related to cultivated land non-agricultural monitoring technical field, method includes: obtaining the current remote sensing image of monitoring area and historical remote sensing image of different period, after respectively boundary segmentation detection, extract the boundary continuous region in each image;Through the semantic segmentation of pre-trained semantic segmentation model to the boundary continuous region of extracted and marked new in the current remote sensing image, identify the boundary continuous region belonging to cultivated land type;For the boundary continuous region not belonging to cultivated land type after semantic segmentation, obtain corresponding sub remote sensing image calculation vegetation index, identify the boundary continuous region belonging to cultivated land type.The application improves the monitoring precision of small-scale, scattered cultivated land in plateau by dynamic history comparison, detailed segmentation analysis and vegetation index statistical calculation.

Description

Technical Field

[0001] This invention relates to the field of monitoring the non-agricultural use of arable land, specifically a method for monitoring the non-agricultural use of arable land based on remote sensing technology. Background Technology

[0002] Currently, monitoring the non-agriculturalization of arable land is primarily achieved through the identification of remote sensing images. In remote sensing image identification, existing methods typically rely on semantic segmentation techniques to segment and classify the boundary areas between arable land and non-agricultural land. This method demonstrates high accuracy in plains areas, especially for large, continuously distributed arable land with distinct spectral characteristics, uniform texture variations, and strong spatial continuity. The semantic segmentation model can effectively distinguish arable land from other land types, exhibiting strong identification capabilities and stability.

[0003] However, this method has significant limitations for farmland in high-altitude mountainous areas. Taking the high-altitude mountainous areas of Qinghai Province as an example, Qinghai Province is located in the heart of the Qinghai-Tibet Plateau, and its farmland is characterized by small-scale and scattered distribution, which contrasts sharply with the large-scale, contiguous farmland in the eastern plains. These farmlands are usually distributed along river valleys, mountain terraces, or slopes, exhibiting strip-like, block-like, or even fragmented distribution patterns. This makes it difficult to accurately capture the boundaries of farmland in remote sensing images, and existing boundary segmentation-based algorithms are ill-suited to this complex distribution pattern. Furthermore, the boundaries of farmland in the high-altitude mountainous areas of Qinghai Province are often blurred, and the farmland and surrounding natural vegetation, such as grasslands and shrubs, have high similarity in spectral characteristics.

[0004] This problem is particularly pronounced during the crop's non-growing season. Due to the low vegetation cover of arable land in Qinghai Province, its reflectance characteristics are similar to bare land or surrounding natural vegetation (such as grasslands and sandy areas). This makes semantic segmentation models prone to misclassifying arable land as non-agricultural areas. Furthermore, the strong plateau climate effects in Qinghai (such as strong sunlight and large diurnal temperature variations) further increase the variability of spectral reflectance characteristics on vegetation growth cycles. The combined effect of these factors makes existing methods face a higher risk of misclassification in the complex geographical environment of Qinghai's plateau and mountainous regions. Summary of the Invention

[0005] 1) Technical problems to be solved This invention provides a method for monitoring the non-agriculturalization of arable land based on remote sensing technology. It achieves multi-level identification of remote sensing images based on the characteristics of arable land distribution in plateau and mountainous areas, thereby solving the problem of low accuracy in identifying non-agriculturalization of arable land in plateau and mountainous environments.

[0006] (ii) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: a method for monitoring the non-agriculturalization of arable land based on remote sensing technology, comprising: Acquire historical remote sensing images of the monitored area at different time points, as well as the latest current remote sensing images; For each of the remote sensing images, region identification is performed to extract regions with continuous boundaries in each remote sensing image; By comparing the continuous boundary regions extracted from the current remote sensing image with those from historical remote sensing images, newly added changed regions are marked in the current remote sensing image. These newly added changed regions, together with all regions detected in the current remote sensing image, constitute a region set. Using a pre-trained semantic segmentation model, regions belonging to the cultivated land type are identified from the set of regions in the current remote sensing image; For areas that do not belong to the cultivated land type after semantic segmentation, they are marked as areas to be refined, and sub-remote sensing images of each area to be refined are obtained; For each sub-remote sensing image, it is divided into multiple sub-regions of fixed size. For each sub-region, the mean and standard deviation of the vegetation index are calculated. The number of sub-regions that meet the set vegetation index threshold is counted and their proportion to the total number is calculated. If the proportion of sub-regions that meet the set vegetation index threshold meets the preset threshold, the corresponding area of ​​the sub-remote sensing image is marked as cultivated land; otherwise, it is marked as non-agricultural land.

[0007] Furthermore, before performing region identification on each of the remote sensing images, each remote sensing image needs to be calibrated and standardized to ensure the consistency of images from different time periods.

[0008] Furthermore, for the current and historical remote sensing images after correction and standardization, a filtering algorithm is first used for noise reduction and contrast enhancement. Then, based on semantic segmentation technology, region identification is performed on each remote sensing image to extract the regions with continuous boundaries in each remote sensing image.

[0009] Furthermore, the newly added changed regions are combined with all regions detected in the current remote sensing image to form a region set. Specifically, this is done as follows: The current remote sensing image is aligned with historical remote sensing images using image registration technology to ensure that the spatial position of each remote sensing image is consistent before regional comparison is performed. Based on the regional comparison results, mark the newly added changed areas in the current remote sensing image; All detected regions in the current remote sensing image, along with the newly added changed regions, are marked to form a region set.

[0010] Furthermore, the process involves using a pre-trained semantic segmentation model to identify regions belonging to the cultivated land type from the region set in the current remote sensing image; specifically, the operation is as follows: Load the pre-trained semantic segmentation model; The set of regions in the current remote sensing image is input into the semantic segmentation model. The semantic segmentation model classifies each region in the set of regions based on the spectral reflectance characteristics, texture features and shape features of the regions, and outputs the classification result of each region and its corresponding classification confidence. A confidence threshold is set in the semantic segmentation model. If the classification confidence of a region output by the semantic segmentation model is higher than the confidence threshold, the region is identified as farmland. If the semantic segmentation model outputs a region with a classification confidence score lower than the confidence score threshold, then that region is determined to be a region that needs further refinement.

[0011] Furthermore, for the region to be refined, a sub-remote sensing image of each region to be refined is acquired, wherein the resolution of the sub-remote sensing image is higher than the resolution of the current remote sensing image.

[0012] Furthermore, the process of dividing each sub-remote sensing image into multiple fixed-size sub-regions, and for each sub-region, calculating the mean and standard deviation of the vegetation index, specifically involves: Each of the aforementioned sub-remote sensing images is divided into multiple consecutive sub-regions of fixed size; For each sub-region, the red, green, and blue bands of the visible light spectrum and the near-infrared band are selected as key bands; Based on the selected key spectral bands, the vegetation index value of each sub-region is calculated according to the vegetation index formula, and a vegetation index statistical map of each sub-region is generated. Calculate the mean and standard deviation of each vegetation index in the statistical chart.

[0013] Furthermore, the vegetation index statistical map is provided with an index mean threshold and a standard deviation threshold. If there are multiple adjacent sub-regions exceeding a set proportion, wherein the vegetation index mean of each sub-region is greater than the mean threshold and the standard deviation is less than the standard deviation threshold, then the region corresponding to the sub-remote sensing image is determined to be cultivated land; otherwise, the regions corresponding to other sub-remote sensing images are non-agricultural land.

[0014] (iii) Beneficial effects: Compared with the prior art, this invention has the following beneficial effects: This invention uses a pre-trained semantic segmentation model to perform preliminary classification of regions in the current remote sensing image, quickly identifying cultivated land types. For regions that do not belong to cultivated land types, those with low confidence are divided into regions to be refined. By dividing the sub-remote sensing images of the regions to be refined into sub-regions of fixed size, the mean and standard deviation of vegetation indices are calculated for each sub-region, and the proportion of sub-regions that meet the set threshold is statistically analyzed. This helps to accurately identify complex terrain and scattered cultivated land areas in plateau regions, thereby better adapting to the characteristics of plateau cultivated land. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating a method for monitoring the non-agricultural use of arable land based on remote sensing technology, provided in an embodiment of the present invention. Figure 2 This is a flowchart illustrating the process of extracting a set of regions from a current remote sensing image in a method for monitoring the non-agriculturalization of cultivated land based on remote sensing technology, provided in an embodiment of the present invention. Figure 3 A flowchart illustrating the process of dividing the area to be refined in a remote sensing-based method for monitoring the non-agriculturalization of cultivated land provided in an embodiment of the present invention. Figure 4 This is a schematic diagram of the overall process of a method for monitoring the non-agricultural use of arable land based on remote sensing technology, provided in an embodiment of the present invention. Figure 5 This is a schematic diagram of a set of regions in a current remote sensing image provided in an embodiment of the present invention, wherein the regions marked in red are the regions detected in the current remote sensing image, and the regions marked in blue are the newly added changed regions after comparison; Figure 6 The embodiments of the present invention provide Figure 5 The diagram shows how the A2 region is classified as the region to be refined and divided into multiple sub-regions of fixed size. Detailed Implementation

[0016] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0017] In the description of this invention, it should be understood that the terms "longitudinal", "lateral", "length", "width", "thickness", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0018] Furthermore, the terms "first" and "second" are used only to distinguish descriptions and should not be interpreted as indicating or implying relative importance.

[0019] It should be noted that, where there is no conflict, the features in the embodiments of the present invention can be combined with each other.

[0020] Currently, monitoring the conversion of arable land to non-agricultural uses mainly relies on the identification of remote sensing images. A commonly used method is semantic segmentation, which uses models to classify remote sensing images at the pixel level, thereby identifying the boundaries between arable land and non-agricultural areas. However, this approach is more suitable for large, continuously distributed arable land areas in plains, because the spectral characteristics and texture patterns of these areas are relatively consistent, allowing classification models to more accurately identify and segment arable land boundaries.

[0021] However, in high-altitude mountainous environments, such as Qinghai Province, the distribution of arable land is far more complex. The characteristics of arable land in these areas differ significantly from those in plains. In Qinghai Province, arable land is typically located in valleys, slopes, or along riverbanks, exhibiting a small, fragmented pattern, far less continuous than in plains areas. This distribution characteristic makes it difficult to discern uniform boundary features in remote sensing images. While semantic segmentation models can easily identify large-scale, continuous arable land, they can easily misclassify these scattered, small-scale arable lands as background areas or non-agricultural land.

[0022] Typically, due to the small and scattered nature of arable land, existing model segmentation methods may misidentify adjacent grassland or bare soil areas as arable land. Furthermore, because scattered arable land areas are small in size or have indistinct spectral characteristics, existing model segmentation methods may directly classify them as non-agricultural areas, thus failing to record them as arable land.

[0023] Taking Qinghai Province as an example, in some areas with frequent agricultural activities, such as the Huangshui River Basin or Huangnan Tibetan Autonomous Prefecture, local arable land is often distributed in mountain valleys, small in scale and scattered. In traditional semantic segmentation models, these areas are easily misclassified as non-agricultural land due to the lack of significant continuous features. In addition, the natural degradation problems in Qinghai (such as grassland degradation and desertification) further complicate the accurate determination of arable land.

[0024] Therefore, existing semantic segmentation-based technologies face significant limitations in high-altitude mountainous environments, and it is necessary to further combine them with other methods to improve the monitoring capabilities of small-scale, scattered farmland, thereby more accurately determining the non-agricultural status of farmland.

[0025] To address the problem of existing semantic segmentation models misidentifying small-scale, scattered farmland in high-altitude mountainous environments, this paper combines... Figures 1 to 6 As shown, this invention proposes a method for monitoring the non-agriculturalization of arable land based on remote sensing technology. This method combines multiple steps of regional segmentation, semantic segmentation, and refined analysis. First, by identifying continuous boundary regions, it ensures that even small-scale arable land areas are included in the analysis, avoiding missed reports due to unclear initial features. During semantic segmentation of the boundary-segmented regions, the model's ability to quantify classification confidence is utilized to distinguish between high-confidence and low-confidence regions, allowing for more refined subsequent processing of small-scale, scattered areas and reducing false positives and false negatives. Subsequently, a comprehensive analysis combining model output and vegetation indices provides an independent discriminative perspective for areas with low classification confidence, thereby reducing background interference and classification bias. This method can more effectively monitor the trend of arable land non-agriculturalization, providing more accurate data support for regional land use planning and ecological protection.

[0026] For details, please refer to Figure 1 The flowchart of a method for monitoring the non-agriculturalization of cultivated land based on remote sensing technology provided by the embodiment of the present invention is shown. First, step S10 is performed: acquiring historical remote sensing images of the monitoring area at different time points, as well as the latest current remote sensing image.

[0027] In the process of monitoring the conversion of arable land to non-agricultural uses, it is necessary to comprehensively collect current and historical remote sensing image data. Regarding the acquisition of remote sensing images, optical or radar images of the target area can be collected using multispectral or high-resolution remote sensing satellites. Furthermore, regarding historical remote sensing images, it is necessary to select historical images with coverage areas consistent with the current remote sensing images, prioritizing data from key periods (such as seasonal transitions and planting seasons) to more accurately analyze changes.

[0028] To accurately compare remote sensing images from different periods, these images need to be corrected and standardized. Specifically, because remote sensing images are affected by factors such as satellite orbit, viewing angle, and sensor characteristics, they may be distorted, offset, or scaled. Geometric correction is used to eliminate geometric distortions in remote sensing images and ensure that images from different times have consistent geographic coordinates. In some embodiments of this invention, known geographic coordinate points (such as landmarks, road intersections, etc.) are used for geometric correction to align the image to a standard map projection coordinate system. Geometric correction algorithms can include, but are not limited to, nearest neighbor interpolation, bilinear interpolation, and higher-order polynomial fitting.

[0029] Because remote sensing images acquired at different times are affected by clouds, aerosols, atmospheric scattering, etc., their spectral reflectance values ​​deviate, directly affecting subsequent analysis results. Radiometric correction is used to eliminate the influence of the atmosphere, light intensity, and sensor characteristics, ensuring consistent spectral information in the images. In some feasible embodiments of this invention, atmospheric effects are corrected using radiative transfer models (such as MODTRAN) or the dark object subtraction (DOS) method based on ground features, adjusting the image reflectance values ​​to compensate for illumination differences caused by variations in solar altitude or azimuth angles.

[0030] In addition, it is important to ensure that the spatial resolution of current and historical remote sensing images is consistent, so as to ensure that the physical size of each pixel is consistent.

[0031] It can be understood that the above steps can effectively reduce the analytical bias caused by differences in remote sensing image acquisition conditions, laying a reliable data foundation for subsequent monitoring of farmland non-agriculturalization.

[0032] After acquiring the current and historical remote sensing images, step S20 is performed: region identification is performed on each remote sensing image to extract the regions with continuous boundaries in each remote sensing image.

[0033] To accurately extract continuous boundary regions in remote sensing images, it is necessary to perform filtering and noise reduction, contrast enhancement, boundary detection, and boundary segmentation processes step by step, ultimately extracting the continuous boundary regions in each remote sensing image.

[0034] Specifically, preprocessing is performed on the current and historical remote sensing images after correction and standardization. The preprocessing steps include filtering for noise reduction and contrast enhancement. Remote sensing images often contain noise (such as sensor noise, atmospheric interference, etc.), and noise reduction can improve image quality. In some feasible embodiments of this invention, Gaussian filtering or median filtering can be used to remove noise. Enhancing contrast makes edges more prominent, facilitating subsequent processing. Histogram equalization or adaptive contrast enhancement methods can be used to enhance the contrast between light and dark areas.

[0035] After remote sensing image preprocessing, edge detection is performed on each image to determine the boundary positions between the ground and objects, providing initial information for boundary segmentation. Edge detection algorithms can include, but are not limited to, Canny edge detection or the Sobel operator.

[0036] After edge detection is performed on the remote sensing image to identify the boundary, the continuous boundary regions in the image are segmented based on the boundary detection results.

[0037] Specifically, in some feasible embodiments of the present invention, segmentation is performed based on the watershed algorithm. The edge detection results are regarded as "mountains," and filling and marking are performed starting from low values ​​("valleys") to divide different regions. More specifically, a distance transformation is performed on the edge detection results to determine the center point of the potential region. The segmented region is gradually filled using the marked points until the edge is reached. For the marked map of the generated region, the continuous boundary region of the segmented area is extracted.

[0038] Extract the continuous boundary regions that meet the criteria from the current remote sensing image and historical remote sensing images. The output information of the continuous boundary regions includes: the contour boundary coordinates of each region, the morphological feature parameters of each region, and the multispectral information of the boundary regions, so as to be used for subsequent refinement detection.

[0039] Understandably, through the aforementioned steps of filtering, boundary detection, and boundary segmentation, continuous boundary regions can be extracted from remote sensing images, laying an accurate regional foundation for subsequent monitoring of farmland conversion to non-agricultural uses. This process fully considers the combined effects of noise interference, spectral characteristics, and regional morphological features, making it suitable for detecting small-scale, scattered farmland in complex plateau terrain.

[0040] After extracting the continuous boundary regions in each remote sensing image, step S30 is performed: compare the continuous boundary regions extracted from the current remote sensing image with those extracted from historical remote sensing images, mark the newly added changed regions in the current remote sensing image, and the newly added changed regions together with all regions detected in the current remote sensing image constitute a region set.

[0041] refer to Figure 2 First, S301 is performed: image registration technology is used to align the current remote sensing image with historical remote sensing images to ensure spatial consistency before regional comparison. This is because remote sensing images may have geometric distortions or positional differences due to factors such as shooting angle, satellite orbit, and terrain undulations. Image registration uses geometric transformations to align remote sensing images acquired at different times, ensuring they are compared under the same spatial reference frame.

[0042] The specific steps for image registration are as follows: First, key feature points are extracted from both the current and historical remote sensing images. These feature points include road intersections, river bends, and building boundaries. Nearest neighbor matching (such as the KNN algorithm) is used to match the extracted feature points between the two images, finding potentially corresponding pairs. To avoid mismatches, the RANSAC (Random Sample Consensus) algorithm can be applied to filter outliers and retain the best match. Based on the matched feature point pairs, the geometric transformation relationship between the images (such as affine or perspective transformation matrices) is estimated. The current remote sensing image is then resampled using the transformation matrix and aligned to the reference frame of the historical remote sensing image. Interpolation methods (such as bilinear interpolation or nearest neighbor interpolation) are used to process the registered image to ensure that image details are not lost. Finally, the current and historical remote sensing images, with consistent spatial locations, are output to provide accurate input for subsequent regional comparisons.

[0043] Step S302: Based on the regional comparison results, mark the newly added changed areas in the current remote sensing image. By comparing the continuous boundary regions in the current remote sensing image and historical remote sensing images through the aligned remote sensing images, identify the newly added boundary regions in the current image.

[0044] Understandably, by acquiring historical and recent remote sensing images of the monitored area and comparing them with the extracted regions, newly changed areas can be identified. This temporal change analysis overcomes the limitations of analyzing remote sensing images at a single time point, helping to capture dynamic changes in cultivated land in plateau regions and improve monitoring accuracy.

[0045] Specifically, in some embodiments of the present invention, the boundary continuous region of the current remote sensing image and the historical remote sensing image is binarized to generate a region mask, where 1 represents the inside of the region and 0 represents the outside.

[0046] Boolean operations are used to compare two binary mask images using logical operations. In some feasible embodiments of this invention, Boolean operations are employed: mask Wherein: Current mask: the region with continuous boundaries identified in the current remote sensing image; Historical mask: the region with continuous boundaries identified in historical remote sensing images; The logical AND operation represents the portion of the current mask that belongs to the continuous region of the color boundary being identified. The logical NOT operation indicates a part that does not belong to the history mask.

[0047] Morphological operations (such as dilation and erosion) are performed on the newly added region mask to remove noise and connect broken areas, improving the integrity of the region. The outline of the newly added region is extracted and marked as a polygon on the current remote sensing image. The newly added boundary continuous change region is output, which provides a basis for merging the subsequent region set.

[0048] Step S303: Mark all detected regions and newly added changed regions in the current remote sensing image to form a region set.

[0049] In summary, a logical OR operation is performed between the boundary continuous region mask of the current remote sensing image and the newly added region mask: mask That is, all regions with continuous boundaries are marked in the same mask.

[0050] In some embodiments, the superimposed mask image is optimized by filling blank points inside the region to ensure the integrity of the boundary region, using an area threshold to remove regions smaller than a specified area (isolated points) to avoid false detection, and using Gaussian filtering or morphological opening operations to eliminate jagged edges at the region edges.

[0051] The optimized mask is converted into a vector polygon format, such as GeoJSON or Shapefile, for subsequent finer detection.

[0052] In summary, it is important to note that due to the complex distribution of arable land in plateau and mountainous areas (strip-like, block-like, fragmented), the boundary areas may contain irregularly shaped regions. In flat areas or clearly defined arable land areas (such as field ridges), the boundaries are relatively regular, appearing rectangular or strip-like. Furthermore, some irregular boundaries originate from actual topographic features (such as hillsides or river valleys) and the representation of ground features in remote sensing images (such as the blurred boundary at the junction of arable land and grassland). The boundaries of regions are usually represented in vector form, such as polygonal or irregularly shaped outlines, which corresponds to the actual distribution of ground features and avoids forcibly dividing the area into regular grid shapes.

[0053] It can be understood that the delineated boundary areas are not regular shapes. These boundary areas provide a basis for subsequent identification of cultivated land and non-cultivated land, enabling more accurate location of cultivated land distribution while reducing the risk of misclassification, especially in complex plateau and mountainous environments.

[0054] Then proceed to step S40: using a pre-trained semantic segmentation model, identify regions belonging to the cultivated land type from the current remote sensing image region set.

[0055] For details, please refer to Figure 3 First, perform S401: load the pre-trained semantic segmentation model.

[0056] The semantic segmentation model is trained on a large number of labeled remote sensing datasets, and its goal is to distinguish different types of land features (such as farmland, buildings, and forest land). It is important to note that the model preprocesses the current remote sensing image data during loading, normalizing the band values ​​of the remote sensing image to match the model's input requirements.

[0057] Step S402: Input the set of regions in the current remote sensing image into the semantic segmentation model. The semantic segmentation model classifies each region in the set of regions based on the spectral reflectance characteristics, texture features and shape features of the regions, and outputs the classification result of each region and its corresponding classification reliability.

[0058] Each region in the current remote sensing image set has been extracted in the previous step, and these regions vary in shape and size. The model classifies each region individually using pixel spectral reflectance characteristics, texture features, and shape features. Spectral reflectance characteristics analyze the spectral reflectance of pixels within the region to distinguish between crops, shrubs, and bare land. For example, cultivated land has high reflectance in the near-infrared band. In some embodiments, texture features are calculated using the gray-level co-occurrence matrix (GLCM) to determine the texture uniformity and roughness of the region. Typically, cultivated land in high-altitude mountainous areas has texture features between smooth grassland and rough bare land. Shape features characterize the shape complexity of the region and can reflect its degree of human intervention.

[0059] It is important to note that each region in the current remote sensing image needs to be cropped into a separate small image, which can be understood as cropping into sub-images. These sub-images are then normalized to the input format required by the model. Specifically, sub-images are extracted from the original image based on boundary coordinates, and pixel values ​​are standardized (e.g., 0-1 or -1 to 1) to ensure consistency with the dynamic range of the model training data. The cropped sub-images are then adjusted according to the model input size (e.g., 256×256 pixels or 512×512 pixels).

[0060] The processed region sub-images are input into a semantic segmentation model for pixel-by-pixel classification. Specifically, the pixel values ​​of each region are used to extract feature vectors through the model's input layer, including spectral values, texture statistics, and shape parameters. After passing through a fully connected layer or a softmax classification layer, the model outputs the category and confidence value of each region.

[0061] In this context, classification reliability represents the confidence level of the model's region classification results. High reliability indicates that the region matches the characteristics of cultivated land samples in the training set, while low reliability indicates higher uncertainty. Reliability calculation is based on the probability distribution output of the model's last layer. For example, for the pixel values ​​of each input region, the model extracts features step-by-step through multiple convolutional layers and uses pooling layers to reduce the dimensionality and summarize the features, generating a feature vector. This feature vector is mapped to the probability distribution space of different categories through fully connected layers. For each region input to the model for classification, the model outputs its classification category and corresponding reliability value. The reliability value, representing the model's confidence in its classification results, is directly given by the probability value of the Softmax function.

[0062] Specifically, the semantic model calculates the classification confidence of the entire region based on the classification probabilities of all pixels in the region: After the model outputs the reliability value of the region, step S403 is performed: a reliability threshold is set in the semantic segmentation model. If the classification reliability of a region output by the semantic segmentation model is higher than the reliability threshold, the region is identified as farmland.

[0063] The reliability range is usually between 0 and 1. If the classification reliability of a region is higher than the preset threshold (such as 0.8), the region is considered to be reliably identified as farmland. High-reliability regions are directly labeled as "farmland" and output to the final classification result.

[0064] Step S404: If the classification confidence of a region output by the semantic segmentation model is lower than the confidence threshold, then the region is determined to be a region to be refined.

[0065] If the classification reliability of a region is lower than the preset threshold, the model's classification result for that region has high uncertainty, and it is judged as a "region to be refined". The region to be refined is marked as a region that needs further processing.

[0066] For example, refer to Figure 5 In the set of regions extracted from the current remote sensing image, the marked region A1 has strong spectral reflectance in the near-infrared band. Using the trained semantic segmentation model, based on comprehensive feature analysis, its confidence level is 0.72, which is higher than the set confidence threshold of 0.6. Therefore, region A1 is directly marked as farmland.

[0067] In the current set of regions extracted from the remote sensing image, another labeled region, A2, has spectral characteristics similar to grassland. Using a trained semantic segmentation model and based on comprehensive feature analysis, its confidence score is 0.52, which is lower than the set confidence threshold of 0.6. Therefore, it is marked as a region to be refined. Subsequent refinement analysis methods will confirm its classification.

[0068] In summary, it can be understood that a pre-trained semantic segmentation model is used to initially classify the region set, effectively utilizing spectral, textural, and spatial features to quickly identify cultivated land types. For regions where the model's output confidence is below a set threshold, they are classified as regions to be further refined, and subsequent classifications are carried out more precisely. This multi-level identification method reduces the risk of misclassification and improves overall classification accuracy.

[0069] Regarding the division of the areas to be refined, specifically, step S50 is performed: for areas that do not belong to the cultivated land type after semantic segmentation, they are marked as areas to be refined, and sub-remote sensing images of each area to be refined are obtained.

[0070] To address the uncertainty in identifying the refined regions using semantic segmentation models, it is necessary to acquire higher-resolution sub-remote sensing images for more detailed analysis. Specifically, the coordinate range of each region to be refined is determined, and high-resolution sub-remote sensing images of each region are acquired using satellite remote sensing platforms or UAVs. Based on the boundary coordinates or mask information of the region to be refined, the high-resolution images are cropped to the corresponding region range, ensuring that the edges of the cropped image match the boundaries of the region to be refined. It is important to note that the spatial resolution of the sub-remote sensing images must be significantly higher than the resolution of the current remote sensing image.

[0071] The reason for using higher-resolution remote sensing images for sub-images is that the natural environment in plateau regions is diverse, and there may be large amounts of natural grassland, shrubs, or bare soil around cultivated land. The spectral characteristics of these features are similar to those of crops in cultivated land, making them difficult to distinguish in low-resolution remote sensing images. Especially in the Qinghai Plateau, the crop growing season is short, vegetation cover is low, and soil exposure is high, further increasing the possibility of them being mixed with background features.

[0072] Through the above operations, sub-remote sensing images with higher resolution and smaller size than the current remote sensing image are obtained, providing a more accurate data foundation for the area to be refined, thereby improving the accuracy and reliability of farmland identification. This method lays the foundation for further vegetation index calculation and farmland classification.

[0073] Finally, proceed to step S60: For each sub-remote sensing image, divide it into multiple sub-regions of fixed size. For each sub-region, calculate the mean and standard deviation of the vegetation index, count the number of sub-regions that meet the set vegetation index threshold, and calculate their proportion of the total. If the proportion of sub-regions that meet the set vegetation index threshold meets the preset threshold, then mark the corresponding area of ​​the sub-remote sensing image as cultivated land; otherwise, mark it as non-agricultural land.

[0074] Specifically, firstly, each sub-remote sensing image is subdivided into smaller sub-regions. By dividing these into fixed-size sub-regions, fine-grained analysis of local characteristics can be performed, improving classification accuracy. It is important to note that each sub-region should be small enough to avoid losing detail while still covering the small cultivated land plots characteristic of plateau and mountainous areas. Regarding how to divide the sub-regions within the sub-remote sensing image, in some feasible embodiments of this invention, a grid partitioning algorithm, such as an equally spaced grid, is used to divide the sub-remote sensing image into continuous, fixed-size rectangular or square regions, with an appropriate resolution selected based on the actual application. For example, a remote sensing image with a 20-meter resolution can be divided into 5m × 5m sub-regions. See [reference needed] for details. Figure 6 , Figure 6 The embodiments of the present invention provide Figure 5 The A2 region was divided into areas to be refined. After acquiring its sub-remote sensing images, these sub-remote sensing images were further divided into multiple consecutive 5m×3m sub-regions. By dividing the region into smaller sub-regions, higher spatial resolution information is preserved, especially in small-scale farmland, where local features can be captured more accurately.

[0075] It is important to note that when using an evenly spaced grid to divide irregular regions, incomplete grids may occur at the region boundaries. To address incomplete grids, some embodiments weight the statistical values ​​according to the proportion of effective pixels within the grid. For example, for a certain boundary grid, its effective pixel coverage proportion is... Then its contribution weight value is The weighted mean of the vegetation index is calculated to represent the effective value of the entire grid. That is, in the subsequent calculation of the vegetation index, the pixel values ​​of the uncovered parts of the incomplete grid are considered invalid data and excluded from the calculation of the mean and standard deviation.

[0076] For each sub-region, its vegetation index is calculated. In some embodiments of the invention, visible light bands (red, green, and blue) and near-infrared bands (NIR) are used. The red band (R) reflects the strong absorption of red light by vegetation, the near-infrared band (NIR) reflects the strong reflectivity of vegetation, and the green and blue bands help distinguish between vegetation types and non-vegetated areas.

[0077] The Normalized Difference Vegetation Index (NDVI) is used to calculate the average reflectance of selected bands for each sub-region, and the NDVI value for each sub-region is calculated. This generates a vegetation index statistical map for each sub-region, visually displaying the vegetation distribution. The NDVI value ranges from -1 to 1, with a higher value indicating more abundant vegetation.

[0078] Introducing the blue band (B) into NDVI to reduce the effects of atmospheric scattering and improve sensitivity to sparse vegetation: in, This is the gain factor (usually taken as 0.25). and This is the atmospheric correction factor (generally taken as 6-7.5). This is the correction factor (usually set to 1).

[0079] Vegetation cover is calculated using the ratio of green to red light segments: Based on the above, NDVI is calculated to initially screen vegetated and non-vegetated areas. For areas where NDVI is determined to be a boundary value (e.g., areas where NDVI is between 0.2 and 0.3), EVI is further calculated to distinguish between sparse vegetation and non-vegetation. For areas where NDVI and EVI results are relatively ambiguous, the Green-Red Ratio (RGI) is calculated, which directly calculates the mean and distribution of G and B reflectance to enhance the judgment.

[0080] Based on the vegetation index values ​​within each sub-region, a vegetation index distribution map for that region is generated. Each sub-region can be considered a small statistical unit, and the map displays the spatial distribution of the vegetation index within that region. Statistical analysis is performed on the vegetation index within each sub-region, calculating the mean and standard deviation of the vegetation index for each sub-region. The mean is the average value of the vegetation index values ​​within each sub-region, and the standard deviation is the standard deviation of the vegetation index values ​​within each sub-region, reflecting the dispersion of the vegetation index within that region.

[0081] By analyzing the vegetation index distribution of known cultivated land, a mean threshold and a standard deviation threshold for the vegetation index are set. Generally, cultivated land has a higher vegetation index, so a suitable mean threshold can be set to determine whether a certain area has high vegetation cover. To exclude highly volatile non-cultivated land areas (such as areas with weeds and shrubs), a standard deviation threshold is used to determine whether the vegetation distribution is consistent. Areas with a standard deviation less than this threshold can be considered to have relatively stable vegetation, reducing the interference of high variability. This threshold is generally set based on the vegetation cover of the area.

[0082] Each subregion is assessed based on a mean threshold and a standard deviation threshold. To reduce false positives, the proportion of subregions that meet the set conditions (i.e., the mean and standard deviation of the vegetation index meet the thresholds) is statistically analyzed. : in, To meet the set conditions for the number of sub-regions, This represents the total number of sub-regions.

[0083] Based on sub-region proportions The value is used to determine the attribute of the corresponding region in the entire sub-remote sensing image. If so, then the area corresponding to the sub-remote sensing image is marked as cultivated land; if If so, the area corresponding to the sub-remote sensing image is marked as non-agricultural land.

[0084] In summary, it is understandable that by using dual threshold filtering of mean and standard deviation, combined with quantity ratio constraints, mixed areas with large spectral dispersion are excluded, thus ensuring the accuracy of farmland determination.

[0085] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. The scope of patent protection of the present invention shall be determined by the claims. Similarly, any equivalent structural changes made based on the description and drawings of the present invention shall also be included within the scope of protection of the present invention.

Claims

1. A method for monitoring the non-agricultural use of arable land based on remote sensing technology, characterized in that, include: Acquire historical remote sensing images of the monitored area at different time points, as well as the latest current remote sensing images; For each of the remote sensing images, region identification is performed to extract regions with continuous boundaries in each remote sensing image; the output information of the regions with continuous boundaries includes the contour boundary coordinates of each region, the morphological feature parameters of each region, and the multispectral information of the boundary regions. By comparing the continuous boundary regions extracted from the current remote sensing image with those from historical remote sensing images, newly added changed regions are marked in the current remote sensing image. These newly added changed regions, together with all regions detected in the current remote sensing image, constitute a region set. Using a pre-trained semantic segmentation model, regions belonging to the cultivated land type are identified from the set of regions in the current remote sensing image. The semantic segmentation model is based on the spectral reflectance characteristics, texture features, and shape features of the region to distinguish between cultivated land, shrubs, and bare land. Among them, the spectral reflectance characteristics are analyzed by analyzing the spectral reflectance of pixels in the region, the texture features are calculated by the gray-level co-occurrence matrix to determine the texture uniformity and roughness of the region, and the shape features characterize the shape complexity of the region. For areas that do not belong to the cultivated land type after semantic segmentation, they are marked as areas to be refined, and sub-remote sensing images of each area to be refined are obtained; For each sub-remote sensing image, it is divided into multiple sub-regions of fixed size. For each sub-region, the mean and standard deviation of the vegetation index are calculated. The number of sub-regions that meet the set vegetation index threshold is counted and their proportion to the total number is calculated. If the proportion of sub-regions that meet the set vegetation index threshold meets the preset threshold, the corresponding area of ​​the sub-remote sensing image is marked as cultivated land; otherwise, it is marked as non-agricultural land.

2. The method for monitoring the non-agriculturalization of arable land based on remote sensing technology according to claim 1, characterized in that, Before performing region identification on each of the remote sensing images, each remote sensing image needs to be calibrated and standardized to ensure the consistency of images from different time periods.

3. The method for monitoring the non-agriculturalization of arable land based on remote sensing technology according to claim 2, characterized in that, For the current and historical remote sensing images after correction and standardization, a filtering algorithm is first used to remove noise and enhance contrast. Then, based on semantic segmentation technology, region identification is performed on each remote sensing image to extract the regions with continuous boundaries in each remote sensing image.

4. The method for monitoring the non-agriculturalization of arable land based on remote sensing technology according to claim 1, characterized in that, The newly added changed regions are combined with all regions detected in the current remote sensing image to form a region set. The specific operation is as follows: The current remote sensing image is aligned with historical remote sensing images using image registration technology to ensure that the spatial position of each remote sensing image is consistent before regional comparison is performed. Based on the regional comparison results, mark the newly added changed areas in the current remote sensing image; All detected regions in the current remote sensing image, along with the newly added changed regions, are marked to form a region set.

5. The method for monitoring the non-agriculturalization of arable land based on remote sensing technology according to claim 1, characterized in that, The process involves using a pre-trained semantic segmentation model to identify regions belonging to the cultivated land type from the current remote sensing image's region set; specifically, the operation is as follows: Load the pre-trained semantic segmentation model; The set of regions in the current remote sensing image is input into the semantic segmentation model, each region in the set is classified, and the classification result and corresponding classification confidence of each region are output. A confidence threshold is set in the semantic segmentation model. If the classification confidence of a region output by the semantic segmentation model is higher than the confidence threshold, the region is identified as farmland. If the semantic segmentation model outputs a region with a classification confidence score lower than the confidence score threshold, then that region is determined to be a region that needs further refinement.

6. The method for monitoring the non-agriculturalization of arable land based on remote sensing technology according to claim 1, characterized in that, For the region to be refined, a sub-remote sensing image of each region to be refined is acquired, wherein the resolution of the sub-remote sensing image is higher than the resolution of the current remote sensing image.

7. The method for monitoring the non-agriculturalization of arable land based on remote sensing technology according to claim 1, characterized in that, The process involves dividing each sub-remote sensing image into multiple fixed-size sub-regions, and for each sub-region, calculating the mean and standard deviation of the vegetation index, specifically as follows: Each of the aforementioned sub-remote sensing images is divided into multiple consecutive sub-regions of fixed size; For each sub-region, the red, green, and blue bands of the visible light spectrum and the near-infrared band are selected as key bands; Based on the selected key spectral bands, the vegetation index value of each sub-region is calculated according to the vegetation index formula, and a vegetation index statistical map of each sub-region is generated. Calculate the mean and standard deviation of each vegetation index in the statistical chart.

8. A method for monitoring the non-agriculturalization of arable land based on remote sensing technology according to claim 7, characterized in that, The vegetation index statistical map is set with an index mean threshold and a standard deviation threshold. If there are multiple adjacent sub-regions exceeding a set proportion value, wherein the vegetation index mean of each sub-region is greater than the mean threshold and the standard deviation is less than the standard deviation threshold, then the area corresponding to the sub-remote sensing image is determined to be cultivated land; otherwise, the areas corresponding to other sub-remote sensing images are non-agricultural land.

Images