An automated field parcel segmentation method for complex landscapes

Abstract

The application discloses a zero-sample field plot segmentation method for complex farmland landscape, and relates to the technical field of remote sensing image intelligent interpretation and agricultural informatization. The method comprises the following steps: acquiring a target region RGB remote sensing image and cutting the image into a preset size image set, segmenting the image set through a SAM method to obtain a ground feature mask and a confidence score, constructing a field plot sample set, dividing the sample set to train an instance segmentation model, and finally applying the trained model to target region field plot segmentation. The application realizes zero-sample field plot segmentation, breaks through the limitations of traditional deep learning sample dependence and poor cross-region migration, supports regional modeling, and balances segmentation accuracy and efficiency, thereby providing a new technical solution for complex terrain region agricultural remote sensing application.

Description

Technical Field

[0001] This invention relates to the field of intelligent interpretation of remote sensing images and agricultural information technology, specifically to an automatic segmentation method for complex farmland landscape plots that integrates zero-sample segmentation model and instance segmentation model. Background Technology

[0002] With global climate change and food market fluctuations, accurately grasping the quantity, quality, and spatial structure of farmland resources has become an essential requirement for improving food self-sufficiency and the stability of agricultural production. Therefore, farmland not only has a production function but also bears the strategic mission of ensuring food security and sustainable development.

[0003] Farmland, composed of numerous and spatially complex plots, forms the basic unit of agricultural production management. Agricultural operations such as cultivation, pest and disease control, and operational path planning all rely on accurate plot boundary data. Especially with the rise of digital agriculture, plot-level information has become a crucial foundation for precise decision-making. Traditional methods of acquiring plot data rely on manual surveying and field investigations. While these methods offer high accuracy, they are costly, time-consuming, difficult to implement on a large scale, and unable to cope with the dynamic changes in agricultural production. Therefore, there is an urgent need to achieve automatic identification and efficient extraction of plot boundaries using modern technologies.

[0004] The rapid development of remote sensing technology has provided new avenues for acquiring large-scale, efficient field data. High-resolution satellite and UAV imagery can cover vast areas, and combined with multispectral data, they can effectively reflect the spectral characteristics of ground features, especially at meter- or sub-meter resolutions, clearly revealing field boundary details. With improved computing power, remote sensing image processing has gradually shifted from manual interpretation to automation and intelligence. Field segmentation technology has also evolved from spectral thresholding to object-oriented analysis, and then to methods based on machine learning and deep learning. In particular, the application of deep learning technologies (such as U-Net, DeepLab, and Mask R-CNN) in image segmentation has achieved good results in farmland segmentation tasks; however, its reliance on a large number of high-quality labeled samples, especially in complex terrain, remains a challenge.

[0005] Traditional methods perform poorly in this complex context. While deep learning methods can improve accuracy, their dependence on training samples and poor regional adaptability limit their widespread application. Therefore, to address these challenges, a method is needed to solve the problem of segmentation sample delineation, thereby improving the efficiency of deep learning methods, enabling independent training and segmentation of multiple small regions, and solving the problem of high-precision field segmentation over large areas in complex farmland backgrounds. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides an automated field segmentation method for complex landscapes. This invention employs a zero-shot segmentation model (SAM), combined with existing land cover products and random forests, to automatically generate a sample set of field segments for instance segmentation. Furthermore, it introduces an instance segmentation model to construct an automated field segmentation framework, achieving high-precision field segmentation.

[0007] The technical solution adopted in this invention is as follows:

[0008] An automated field segmentation method for complex landscapes includes the following steps:

[0009] Step 1: Acquire remote sensing images;

[0010] Acquire RGB remote sensing images of the target area and crop them into an image set of 512*512 pixels.

[0011] Step 2: Image segmentation;

[0012] The SAM method was used to segment the images in the image set, and the masks of all ground features in the images and their corresponding confidence scores were obtained.

[0013] Step 3: Construct a field sample set;

[0014] Step 4: Train the instance segmentation model;

[0015] The field sample set is divided into a training set and a validation set according to the proportion, and the instance segmentation model is trained.

[0016] Step 5: Apply the instance segmentation model trained in Step 4 to the field segmentation of the target area.

[0017] Furthermore, the specific method for step 3 is as follows:

[0018] Step 3.1: Acquire multi-band remote sensing images of the target area and calculate the NDVI and NDWI indices:

[0019] , ;

[0020] in, R represents the near-infrared band, and R represents the red band. Indicate the green light band; select the required NIR, NDVI, and NDWI bands, resample to the same resolution as the remote sensing image in step 1, and then crop to an image set of 512*512 pixels;

[0021] Step 3.2: Calculate features and construct a feature set;

[0022] For each mask obtained in step 2, the following features are calculated;

[0023] Calculate the mean and standard deviation of the R, G, and B bands of the remote sensing image in step 1;

[0024] Calculate the mean and standard deviation of the three bands (NIR, NDVI, and NDWI) of the multi-band remote sensing image in step 3.1;

[0025] Calculate the mask area, which is the number of pixels within the mask area;

[0026] Calculate the mean and standard deviation of the class uniformity index; the class uniformity index of the i-th pixel. The calculation formula is as follows:

[0027] ;

[0028] Where N is the number of adjacent pixels in the window excluding the center pixel; and Let i and k represent the class values ​​of the i-th cell (i.e., the center cell in the current window) and its k-th neighboring cell, respectively. When the classes are the same... It is 1; otherwise, it is 1. The value is 0; the range of HI is from 0 to 1, and the larger the value, the more uniform the category.

[0029] Calculate the aspect ratio:

[0030] ;

[0031] Where length is the projection length of the mask in the principal direction obtained by principal component analysis (PCA), and width is the projection width of the mask in the direction perpendicular to the principal direction obtained by PCA.

[0032] Computational boundary complexity:

[0033] ;

[0034] Where, perimeter represents the total length of the region boundaries obtained by solving the contour, and idealperimeter is the length of the ideal circular boundary calculated based on the area of ​​the region, and its calculation formula is:

[0035] .

[0036] By calculating the above features, a feature set for random forest classification is constructed.

[0037] Step 3.3: Extract the field cover;

[0038] First, the existing land cover products are used to assign initial categories to the masks. Then, the feature set obtained in step 3.2 is combined with the feature set and input into a random forest classifier for training. Finally, the trained random forest classifier is used to classify all the masks obtained in step 2, and only the masks of the preset categories are retained, thereby constructing a field sample set.

[0039] This invention provides an automated field segmentation framework for complex landscapes. First, RGB Google images of the target area at 0.6m spatial resolution are acquired, cropped to 512*512 pixels, and 250 images are randomly selected as training samples for the subsequent model. Then, the MoSAM method is used to automatically segment these 250 small images to obtain ground cover masks. Next, NIR, NDVI, and NDWI data are obtained by combining Sentinel-2 imagery, and features are calculated to construct a feature set to train a random forest classifier to extract the field cover masks. Finally, the training data is input into the Mask2Former model for instance segmentation training, and the trained model is used to achieve field segmentation.

[0040] The beneficial effects of this invention are as follows:

[0041] The beneficial effects of this invention are reflected in the fact that it compensates for the time-consuming and laborious sample delineation problem of the Mask2Former instance segmentation model by leveraging the zero-shot segmentation capability of the SAM model, and by using SinoLC-1 data and random forests to overcome the classless segmentation limitation of the SAM model, thus achieving higher-precision zero-shot field segmentation. In existing field segmentation methods, traditional methods are often affected by lighting conditions, terrain variations, and the complexity of field morphology, especially performing poorly in small-scale, fragmented, or poorly regular fields. Deep learning methods require large-scale, high-quality labeled samples, and the model's transferability across different regions is limited. This invention not only overcomes the limitations of traditional deep learning methods—"strong sample dependence and poor cross-regional transferability"—but also supports regional modeling for different terrain features, ensuring both segmentation accuracy and efficiency. It provides a new technical solution for agricultural remote sensing applications such as agricultural resource surveys, crop growth monitoring, and yield estimation in complex terrain areas. Attached Figure Description

[0042] Figure 1 This is a flowchart of a zero-sample field segmentation method for complex farmland landscapes according to the present invention;

[0043] Figure 2 A structural block diagram of the zero-sample field segmentation method provided in an embodiment of the present invention. Detailed Implementation

[0044] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments and accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0045] This embodiment takes Sichuan Province as the research area, and selects 30 512*512 pixel images from each of the 17 prefecture-level cities (excluding Ganzi, Aba, Liangshan Autonomous Prefecture and Panzhihua) as a test set for illustration, including the following steps:

[0046] Step 1: Acquire remote sensing images;

[0047] A high-resolution Google image of approximately 23,000×15,000 pixels (0.6m spatial resolution) was randomly selected from each of the 17 prefecture-level cities and cropped into an image set of 512×512 pixels.

[0048] For each Google image, 250 images are randomly selected from the cropped image set as the source of subsequent training samples.

[0049] Step 2: Image segmentation;

[0050] The SAM method was used to segment the images in the image set, and the masks of all ground features in the images and their corresponding confidence scores were obtained.

[0051] Step 3: Construct a field sample set;

[0052] Step 3.1: For 17 prefecture-level cities, acquire Sentinel-2 imagery of the target area and calculate the NDVI and NDWI indices:

[0053] , ;

[0054] in, R represents the near-infrared band, and R represents the red band. This indicates the green light band. Select the desired NIR, NDVI, and NDWI bands, resample them to the same resolution as Google Imagery, then crop them into a 512*512 pixel image set, and select the corresponding 250 images from it.

[0055] Step 3.2: Calculate features and construct a feature set;

[0056] For each mask obtained in step 2, the following features are calculated;

[0057] Calculate the mean and standard deviation of the R, G, and B bands for high-resolution Google imagery.

[0058] Calculate the mean and standard deviation of the three bands (NIR, NDVI, and NDWI) in Sentinel-2 imagery.

[0059] Calculate the mask area, which is the number of pixels within the mask area.

[0060] Calculate the mean and standard deviation of the class uniformity index; the class uniformity index of the i-th pixel. The calculation formula is as follows:

[0061] ;

[0062] Where N is the number of adjacent pixels in the window excluding the center pixel; and These represent the category values ​​of the center pixel and the kth neighboring pixel, respectively. When the categories are the same... It is 1; otherwise, it is 1. The value is 0; the range of HI is 0 to 1, and the larger the value, the more uniform the category.

[0063] Calculate the aspect ratio:

[0064] ;

[0065] Where length is the projection length of the mask in the principal direction obtained by principal component analysis (PCA), and width is the projection width of the mask in the direction perpendicular to the principal direction obtained by PCA.

[0066] Computational boundary complexity:

[0067] ;

[0068] Wherein, perimeter is the sum of the lengths of the region boundaries obtained by solving the contour using the cv2.arcLength() function, and ideal perimeter is the length of the ideal circular boundary calculated based on the area of ​​the region, and its calculation formula is as follows:

[0069] .

[0070] By calculating the above features, a feature set for random forest classification was constructed.

[0071] Step 3.3: Extract the field cover;

[0072] The SinoLC-1 dataset is used to assign initial categories to the masks. Then, the feature set obtained in step 3.2 is combined with the random forest classifier for training. Finally, the trained random forest classifier is used to classify all the masks obtained in step 2, and only the masks with the category "farmland" are retained to construct the field sample set.

[0073] Step 4: For each prefecture-level city, randomly select 200 images from the field sample set obtained in Step 3 as the training set and 50 images as the validation set to train the instance segmentation model Mask2Former.

[0074] Step 5: Apply the instance segmentation model Mask2Former trained in Step 4 to the field segmentation of the target region.

[0075] The zero-sample field segmentation method of this invention was tested in a test set of 17 prefecture-level cities, and compared with the original SAM method and RSPrompter method. The final results show that the field segmentation accuracy of this invention reached an average F1 score of 0.8764, an average OA of 0.9023, and an average mIoU of 0.7908. Compared with the original SAM method, the method of this invention improved by 20%, 11%, and 19%, respectively; compared with the RSPrompter method, the method of this invention improved by 14%, 12%, and 15%, respectively.

[0076] In summary, the zero-shot field segmentation method for complex farmland landscapes proposed in this invention demonstrates higher segmentation accuracy and field integrity than SAM and the SAM-based improved RSPrompter under varying terrain and farmland landscape complexity in Sichuan Province. This invention employs the powerful SAM model, leveraging its zero-shot segmentation capability, and combines multi-source remote sensing data with a random forest classifier to construct a field sample set for training the instance segmentation model. This method not only overcomes the limitation of traditional deep learning methods' strong sample dependency but also, due to its advantage of not requiring manual sample delineation, divides large areas into multiple smaller regions for independent model training and segmentation, solving the problem of poor cross-regional transferability in traditional deep learning methods. This invention balances segmentation precision with efficiency, providing a new technical solution for agricultural remote sensing applications such as agricultural resource surveys, crop growth monitoring, and yield estimation in complex terrain areas.

Claims

1. An automated field segmentation method for complex landscapes, characterized in that, Includes the following steps: Step 1: Acquire remote sensing images; Acquire RGB remote sensing images of the target area and crop them into an image set of a preset size; Step 2: Image segmentation; The SAM method was used to segment the images in the image set, and all ground cover masks and their corresponding confidence scores were obtained in the images. Step 3: Construct a field sample set; Step 4: Train the instance segmentation model; The field sample set was divided into a training set and a validation set according to the proportion, and the instance segmentation model was trained. Step 5: Apply the instance segmentation model trained in Step 4 to the field segmentation of the target area.

2. The automated field segmentation method for complex landscapes as described in claim 1, characterized in that, Step 3 includes the following steps: Step 3.1: Acquire multi-band remote sensing images of the target area and calculate the NDVI and NDWI indices: ， ； in, R represents the near-infrared band, and R represents the red band. Indicates the green light band; select the required NIR, NDVI, and NDWI bands, resample to the same resolution as the remote sensing image in step 1, and then crop to a preset image set; Step 3.2: Calculate features and construct a feature set; For each mask obtained in step 2, the following features are calculated: Calculate the mean and standard deviation of the R, G, and B bands of the remote sensing image in step 1; Calculate the mean and standard deviation of the three bands (NIR, NDVI, and NDWI) of the multi-band remote sensing image in step 3.1; Calculate the mask area (area); Calculate the mean and standard deviation of the category homogeneity index; Calculate the aspect ratio; Calculate the boundary complexity; Construct a feature set based on the calculated features; Step 3.3: Extract the field cover; First, the existing land cover products are used to assign initial categories to the masks. Then, the feature set obtained in step 3.2 is combined with the feature set and input into a random forest classifier for training. Finally, the trained random forest classifier is used to classify all the masks obtained in step 2, and only the masks of the preset categories are retained, thereby constructing a field sample set.

3. The automated field segmentation method for complex landscapes as described in claim 2, characterized in that, In step 1, the RGB remote sensing image is cropped into an image set of 512*512 pixels.

4. An automated field segmentation method for complex landscapes as described in claim 2 or 3, characterized in that, In step S3.2, the category homogeneity index The calculation formula is as follows: ； in, The index represents the class uniformity index of the i-th pixel; N is the number of neighboring pixels within the window excluding the center pixel. and These represent the category values ​​of the i-th pixel and its k-th neighboring pixel, respectively. When the categories are the same... It is 1; otherwise, it is 1. The value is 0; the range of HI is 0 to 1.

5. An automated field segmentation method for complex landscapes as described in claim 2 or 3, characterized in that, In step S3.2, the formula for calculating the aspect ratio is as follows: ； in, Indicates the aspect ratio, where length is the projected length of the mask obtained by the PCA method in the principal direction, and width is the projected width of the mask obtained by the PCA method in the direction perpendicular to the principal direction.

6. An automated field segmentation method for complex landscapes as described in claim 2 or 3, characterized in that, In step S3.2, the formula for calculating boundary complexity is as follows: ； Where, Boundary complexity represents the boundary complexity, perimeter represents the total length of the region boundaries obtained by solving the contour, and ideal perimeter is the length of the ideal circular boundary calculated based on the area of ​​the region, and the calculation formula is: 。

Images