African arid region cultivated land parcel extraction method based on MAA-BCNet

Abstract

The invention discloses an African arid region cultivated land parcel extraction method based on MAA-BCNet. The method comprises the following steps: preprocessing a satellite optical image, making a sample, carrying out feature screening on spectral features by using an RF + TOP-K method, and carrying out data dimension reduction by using a principal component analysis method; building an MAA-BCNet model and training the MAA-BCNet model, selecting a training area and a test area, cutting images and labels to make a data set, and performing positive and negative sample equalization and data enhancement on the data set; designing a double-path encoder and a double loss function, wherein an AdamW optimizer is used in a model training process; designing axial attention and edge detection to carry out multi-scale feature fusion; up-sampling is carried out by adopting jump connection; and predicting the cultivated land distribution of the African arid region and drawing. According to the method, the grain safety of the African arid region is guaranteed while the labor and time cost is saved, the problem that the arid region cultivated land block edge is fuzzy is solved, and important technical support is provided for cultivated land management, agricultural irrigation and social sustainable development.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of satellite remote sensing image recognition, and particularly relates to an African arid land cultivated land plot extraction method based on MAA-BCNet. BACKGROUND

[0002] Under the severe challenges of global climate change, water resource shortage and population growth, the accurate monitoring of cultivated land in the African arid land is crucial to food security. Satellite remote sensing technology has become a core means for cultivated land monitoring because it can quickly obtain large-scale surface information. However, when applied in the African arid land, it faces significant technical bottlenecks. In the high-contrast desert background, the edges of cultivated land plots are prone to be blurred and broken in the medium-resolution remote sensing image. In addition, the complex terrain and irregular distribution of cultivated land lead to insufficient extraction accuracy of traditional methods. In the prior art, early methods such as unsupervised classification and edge detection have low accuracy due to the limitations of resolution and noise interference. Object-oriented analysis and machine learning methods (such as SVM and random forest) are easily affected by salt and pepper noise or have insufficient edge fitting capability when dealing with heterogeneous agricultural landscapes. Although the encoder-decoder model based on deep learning has been improved, a single encoder cannot balance multi-scale features, and the fusion encoder has information loss. Moreover, the model has limited spectral differentiation ability for cultivated land and similar land covers (such as grassland and forest land) in the desert background. In addition, the existing model lacks explicit constraints on the boundaries of cultivated land, resulting in misjudgment or missed detection of irregular plot edges. In view of the above problems, there is an urgent need for a remote sensing cultivated land extraction technology that can effectively fuse multi-scale features and enhance edge constraints to adapt to the complex geographical environment and cultivated land form in the African arid land and improve the pixel-level cultivated land plot extraction accuracy, thereby providing reliable technical support for large-scale cultivated land monitoring. SUMMARY

[0003] The application aims to provide an African arid land cultivated land plot extraction method based on MAA-BCNet, which can fuse multi-scale axial attention and multi-level boundary constraints to realize accurate extraction of cultivated land plots in the African arid land.

[0004] To achieve the above functions, the application designs an African arid land cultivated land plot extraction method based on MAA-BCNet, which performs the following steps S1-S5 to extract cultivated land plots in the target area:

[0005] Step S1: Based on the GEE platform, obtain the L2A level optical satellite image of Sentinel-2 of the region to be identified;

[0006] Step S2: Based on the band and spectral index threshold, the cultivated land range in the Sentinel-2 optical satellite image is preliminarily extracted, and based on the extracted region, visual interpretation is performed, binary classification labeling of cultivated land and non-cultivated land is performed, and sample points are generated;

[0007] Step S3: spectral feature extraction is performed on the Sentinel-2 optical satellite image obtained in step S2, the importance of the spectral features is sorted and then randomly combined, the best feature combination is evaluated, principal component analysis is performed, three-band images of the Sentinel-2 optical satellite image after dimension reduction are obtained as input data;

[0008] Step S4: the MAA-BCNet model is built based on the double-branch encoder structure, the Sentinel-2 optical satellite image is divided into a training area and a test area, the three-band images obtained after dimension reduction and the samples corresponding to the range are taken as the input of the MAA-BCNet model, the MAA-BCNet model is trained, and the trained MAA-BCNet model is obtained;

[0009] Step S5: for the target area, the trained MAA-BCNet model is applied to complete the extraction of the cultivated land plot in the target area, and a cultivated land plot distribution map of the target area is drawn.

[0010] As a preferred technical solution of the present application: the specific steps of step S2 are as follows:

[0011] Step S2.1: for the Sentinel-2 optical satellite image, the red, green and blue bands B2, B3 and B4, the near-infrared bands B11 and B12, and the red edge bands B5, B6 and B7 are taken as the bands of the preliminary extracted cultivated land range, the spatial resolution is resampled to 10m, the related vegetation index is calculated, the bands are synthesized, and the cultivated land range is preliminarily extracted based on the vegetation spectrum threshold;

[0012] Step S2.2: the PIE and Arcgis software are used to visually interpret the Sentinel-2 optical satellite image, binary classification label data of cultivated land and non-cultivated land are made, and sample points are automatically generated from the label data by using Arcgis.

[0013] As a preferred technical solution of the present application: the specific method of step S3 is as follows:

[0014] The out-of-bag data error of the RF algorithm is used to sort and randomly combine the importance of the spectral features of the Sentinel-2 optical satellite image, each random combination of spectral features is taken as a subset, the TOP-K method is introduced to cross-validate and evaluate each subset of spectral features to obtain the best feature combination, the principal component analysis method is used to perform feature dimension reduction on the best feature combination, three principal components with a cumulative contribution rate of more than 98% are retained, and a feature map is obtained.

[0015] As a preferred technical solution of the present application: the MAA-BCNet model built in step S4 is based on a double-branch encoder structure, including two branches, one of which is a backbone network encoder composed of a convolutional neural network VGG16, which extracts local features of different scales in the feature map; the other branch is an RMT encoder based on Manhattan self-attention combined with VisionTransformer, which extracts global features of different scales in the feature map;

[0016] The local features extracted by the backbone network encoder and the global features extracted by the RMT encoder are input into the feature fusion module for feature fusion, and the feature map after feature fusion is input into the decoder to upsample the multi-scale feature map through the jump connection.

[0017] As a preferred technical solution of the present application: the feature fusion module sequentially includes a convolutional layer, an axial attention module, an edge detection module, a pooling layer, a convolutional layer and a ReLU activation function;

[0018] The axial attention module uses horizontal axis convolution and vertical axis convolution to capture long-distance dependencies in the spatial dimension, and uses residual connection for output; the edge detection module uses four different Sobel filters to supplement the edge details enhanced by the axial attention features.

[0019] As a preferred technical solution of the present application: the training process of the MAA-BCNet model in step S4 is as follows:

[0020] Step S4.1: Select the training area and test area on the feature map obtained in step S3, and cut the image and label into image blocks with a size of 256x256 without overlapping;

[0021] Step S4.2: Perform positive and negative sample balancing to make the positive and negative sample ratio about 1:1, and use geometric transformation for data enhancement, then randomly divide into a training set and a validation set with a ratio of 8:2;

[0022] Step S4.3: Use the Dice Loss+Focal Loss dual loss function to train the MAA-BCNet model, and the Dice Loss+Focal Loss dual loss function is as follows:

[0023]

[0024] Focal Loss=-α t (1-p t ) γ log(p t )

[0025] Wherein, α t is a balance factor, pt For prediction probability, p for positive class, 1-p for negative class, gamma is the adjustment factor, p i For the probability of the i-th pixel predicted by the model, g i For the true label of the i-th pixel, taking 0 or 1, epsilon is a smoothing term to prevent the denominator from being zero.

[0026] Step S4.4: The MAA-BCNet model training process uses AdamW as the optimizer, the initial learning rate is set to 0.0001, the learning rate is reduced using the cosine annealing method, the batch size is 8, and the training iteration number is 100.

[0027] Advantages: Compared with the prior art, the advantages of the present application include:

[0028] 1. The present application uses Sentinel-2 image to optimize features and reduce dimensionality, proposes a dual-branch encoder structure and introduces multi-scale axial attention and multi-level boundary constraint for cultivated land extraction method, which reduces the calculation amount while improving the extraction accuracy.

[0029] 2. The present application uses Sentinel-2 image data to realize rapid acquisition of multi-scale cultivated land distribution in arid regions, reduces the consumption of manpower and material resources, and greatly saves the monitoring cost. DETAILED DESCRIPTION

[0030] Figure 1 is a flowchart of the MAA-BCNet-based cultivated land plot extraction method in the African arid region according to an embodiment of the present application;

[0031] Figure 2 is a schematic diagram of the MAA-BCNet model according to an embodiment of the present application. DETAILED DESCRIPTION

[0032] The present application will be further described below in conjunction with the drawings. The following examples are only used to more clearly illustrate the technical solutions of the present application, and cannot be used to limit the protection scope of the present application.

[0033] The MAA-BCNet-based cultivated land plot extraction method in the African arid region provided by the embodiment of the present application, with reference to Figure 1 , executes the following steps S1-S5 to complete the extraction of cultivated land plots in the target area:

[0034] Step S1: Based on the GEE (Google Earth Engine) platform, acquire the Sentinel-2 L2A level optical satellite image of the region to be identified;

[0035] Step S2: Based on the waveband and spectral index threshold, the cultivated land range in the Sentinel-2 optical satellite image is preliminarily extracted, based on the extracted region visual interpretation, the cultivated land and non-cultivated land are classified and labeled and sample points are generated;

[0036] The specific steps of step S2 are as follows:

[0037] Step S2.1: For the Sentinel-2 optical satellite image, the red, green and blue wavebands B2, B3 and B4, the near-infrared wavebands B11 and B12, and the red edge wavebands B5, B6 and B7 are used as the wavebands for preliminary extraction of the cultivated land range, the spatial resolution is resampled to 10m, the related vegetation index is calculated, the wavebands are synthesized, and the cultivated land range is preliminarily extracted based on the vegetation spectral threshold;

[0038] Step S2.2: The PIE and Arcgis software are used to visually interpret the Sentinel-2 optical satellite image, and the binary classification label data of cultivated land and non-cultivated land are made, and the sample points are automatically generated from the label data by using Arcgis.

[0039] Step S3: The spectral features of the Sentinel-2 optical satellite image obtained in step S2 are extracted, the importance of the spectral features is sorted, and the best feature combination is evaluated, and principal component analysis is performed to obtain three-band images of the Sentinel-2 optical satellite image after dimensionality reduction, which are used as input data to maximize the use of spectral features of the satellite image while reducing the calculation cost;

[0040] The specific method is as follows:

[0041] The out-of-bag data error of the RF (Random Forest) algorithm is used to sort and randomly combine the spectral features of the Sentinel-2 optical satellite image, each random combination of spectral features is used as a subset, the TOP-K method is introduced to cross-validate and evaluate each subset of spectral features to obtain the best feature combination, and principal component analysis (PCA) is used to reduce the dimensionality of the best feature combination, and three principal components with a cumulative contribution rate of more than 98% are retained to obtain a feature map.

[0042] Step S4: based on the dual-branch encoder structure, a MAA-BCNet (Multi-scale Axial Attention and Boundary-Constrained U-Net with Dual Encoders) model is built, Sentinel-2 optical satellite images are divided into a training area and a test area, three-band images obtained after dimension reduction and samples corresponding to the range are taken as inputs of the MAA-BCNet model, the MAA-BCNet model is trained, and a trained MAA-BCNet model is obtained;

[0043] With reference to Figure 2 The MAA-BCNet model is built by using a PyTorch deep learning library, the model is based on a U-Net++, a dual-branch encoder structure is used for farmland feature extraction, farmland plot segmentation results are obtained through a feature fusion module and a multi-scale decoder, the dual-branch encoder structure includes two branches, one branch is a backbone network encoder composed of a convolutional neural network VGG16, which serves as a local convolutional encoder to provide local features of different scales for subsequent feature fusion, and the other branch is an RMT (Retentive Networks Meet Vision Transformers) encoder based on Manhattan self-attention combined with a Vision Transformer, which works in parallel with the backbone network to extract global features of different scales from feature maps.

[0044] The local features extracted by the backbone network encoder and the global features extracted by the RMT encoder are input into the feature fusion module for feature fusion to generate an attention map with more prominent farmland features; the feature map after feature fusion is input into the decoder to obtain multi-scale feature maps through up-sampling by using a skip connection.

[0045] The feature fusion module sequentially includes a convolutional layer, axial attention modules CBAMs (Convolutional Block Axial Attention Module), an edge detection module ED (Edge Detect), a pooling layer, a convolutional layer and a ReLU activation function.

[0046] The axial attention module CBAMs replaces the spatial attention of the traditional CBAM attention mechanism with axial attention, captures long-distance dependencies in the spatial dimension using horizontal axis convolution and vertical axis convolution, strengthens the expression of arable land spectrum and edge in the corresponding direction of the drought background, replaces the traditional Sigmoid function output with a residual connection, and in scenes where the contrast between arable land and the background is low, this design can retain the original features and prevent the model from losing important semantic information due to excessive attention to edges. The edge detection module ED applies four different Sobel filters to supplement the edge details that the axial attention in the drought background fails to capture, forming a cascading optimization process from semantic enhancement to spatial positioning to edge refinement. The feature-enhanced feature map is fused and input into the decoder, and the multi-scale feature map is upsampled through a jump connection.

[0047] The training process of the MAA-BCNet model is as follows:

[0048] Step S4.1: Select the training area and test area on the feature map obtained in step S3, and cut the image and label into image blocks with a size of 256x256 without overlapping;

[0049] Step S4.2: Perform positive and negative sample balancing to make the positive and negative sample ratio about 1:1, use geometric transformation for data augmentation, and then randomly divide into a training set and a validation set in a ratio of 8:2;

[0050] Step S4.3: Train the MAA-BCNet model using the Dice Loss+Focal Loss dual loss function, which complements the Dice Loss and Focal Loss loss functions to solve classification bias and segmentation accuracy problems and improve the segmentation performance of small-scale plots and edge regions; the Dice Loss+Focal Loss dual loss function is as follows:

[0051]

[0052] Focal Loss = -α t (1-p t ) γ log(p t )

[0053] Where α t is a balance factor, p t is the prediction probability, p for the positive class and 1-p for the negative class, γ is a tuning factor, p i is the probability of the i-th pixel predicted by the model, g i is the true label of the i-th pixel, taking values of 0 or 1, and ε is a smoothing term to prevent the denominator from being zero;

[0054] Step S4.4: The MAA-BCNet model training process uses AdamW as the optimizer, the initial learning rate is set to 0.0001, the learning rate is decreased using the cosine annealing method, the batch size is 8, and the number of training iterations is 100.

[0055] Step S5: For the target area, apply the trained MAA-BCNet model to complete the extraction of cultivated land plots in the target area and draw the cultivated land plot distribution map of the target area.

[0056] The embodiments of the present invention are described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Various changes can be made within the scope of knowledge possessed by ordinary technicians in this field without departing from the spirit of the present invention.

Claims

1. The method for extracting cultivated land plots in arid areas of Africa based on MAA-BCNet is characterized by: Execute the following steps S1 to S5 to complete the extraction of cultivated land plots in the target area: Step S1: Acquire Sentinel-2 L2A optical satellite images of the area to be identified based on the GEE platform; Step S2: Based on the band and spectral index thresholds, the cultivated land area in the Sentinel-2 optical satellite image is preliminarily extracted. Based on the visual interpretation of the extracted area, the cultivated land and non-cultivated land are classified and labeled and sample points are generated; Step S3: Spectral features are extracted from the Sentinel-2 optical satellite image obtained in step S2. The spectral features are ranked by importance and randomly combined. The best feature combination is evaluated and principal component analysis is performed to obtain a three-band image of the Sentinel-2 optical satellite image after dimensionality reduction as the input image. Step S4: Build the MAA-BCNet model based on the dual-branch encoder structure, divide the Sentinel-2 optical satellite image into a training area and a test area, use the three-band image obtained after dimensionality reduction and the samples in the corresponding range as the input of the MAA-BCNet model, train the MAA-BCNet model, and obtain the trained MAA-BCNet model; Step S5: For the target area, apply the trained MAA-BCNet model to complete the extraction of cultivated land plots in the target area and draw the cultivated land plot distribution map of the target area.

2. The method for extracting cultivated land plots in arid areas of Africa based on MAA-BCNet according to claim 1, characterized in that: The specific steps of step S2 are as follows: Step S2.1: For Sentinel-2 optical satellite images, use the red, green, and blue bands B2, B3, and B4, the near-infrared bands B11 and B12, and the red-edge bands B5, B6, and B7 as the bands for preliminary extraction of cultivated land areas. Resample to a uniform spatial resolution of 10 m, calculate relevant vegetation indices, and then perform band synthesis to preliminarily extract cultivated land areas based on vegetation spectral thresholds. Step S2.2: Use PIE and Arcgis software to visually interpret Sentinel-2 optical satellite images to produce binary labeled data for cultivated land and non-cultivated land. Use Arcgis to automatically generate sample points from the labeled data.

3. The method for extracting cultivated land plots in arid areas of Africa based on MAA-BCNet according to claim 1, characterized in that: The specific method of step S3 is as follows: The out-of-bag data error of the RF algorithm is used to sort the importance of the spectral features of the Sentinel-2 optical satellite imagery and then randomly combine them. Each random combination of spectral features is used as a subset. The TOP-K method is introduced to perform cross-validation evaluation on each subset of spectral features to obtain the optimal feature combination. The principal component analysis method is used to reduce the feature dimension of the optimal feature combination, and the three principal components with a cumulative contribution rate of more than 98% are retained to obtain the input data.

4. The method for extracting cultivated land plots in arid areas of Africa based on MAA-BCNet according to claim 1, characterized in that: The MAA-BCNet model built in step S4 is based on a dual-branch encoder structure, consisting of two branches. One branch is a backbone network encoder composed of a convolutional neural network VGG16, which extracts local features of different scales in the feature map; the other branch is an RMT encoder composed of Manhattan self-attention combined with the Vision Transformer, which extracts global features of different scales in the feature map. The local features extracted by the backbone network encoder and the global features extracted by the RMT encoder are input into the feature fusion module for feature fusion. The feature map after feature fusion is input into the decoder, and the multi-scale feature map is upsampled through jump connections.

5. The method for extracting cultivated land plots in arid areas of Africa based on MAA-BCNet according to claim 4, characterized in that: The feature fusion module includes convolution layer, axial attention module, edge detection module, pooling layer, convolution layer and ReLU activation function in sequence; The axial attention module uses horizontal axis convolution and vertical axis convolution to capture long-distance dependencies in the spatial dimension and uses residual connection output; the edge detection module uses four different Sobel filters to supplement the edge details after the axial attention feature enhancement.

6. The method for extracting cultivated land plots in arid areas of Africa based on MAA-BCNet according to claim 1, characterized in that: The training process of the MAA-BCNet model in step S4 is as follows: Step S4.1: Select the training area and the test area on the feature map obtained in step S3, and crop the image and label into image blocks of size 256×256 without overlap; Step S4.2: Balance the positive and negative samples so that the positive and negative sample ratio is approximately 1:

1. After data augmentation using geometric transformation, randomly divide the data into training and validation sets with a ratio of 8:

2. Step S4.3: Use the Dice Loss + Focal Loss dual loss function to train the MAA-BCNet model. The Dice Loss + Focal Loss dual loss function is as follows: Focal Loss=-α t (1-p t ) γ log(p t ) Among them, α t is the balance factor, p t is the predicted probability, which is p for the positive class and 1-p for the negative class, and γ is the adjustment factor, p i is the probability of the i-th pixel predicted by the model, g i is the true label of the i-th pixel, which takes a value of 0 or 1, and ε is a smoothing term to prevent the denominator from being zero; Step S4.4: The MAA-BCNet model training process uses AdamW as the optimizer, the initial learning rate is set to 0.0001, the learning rate is decreased using the cosine annealing method, the batch size is 8, and the number of training iterations is 100.

Images