Method for accurately identifying cultivated land parcels in hilly and mountainous areas based on DeepLabV < 3 + >
By combining the DeepLabV3+ deep learning network model with the ResNet backbone network, the problems of complex terrain and fragmented land parcels in hilly and mountainous areas were solved, and the accurate identification of arable land parcels was achieved, especially when the parcel area was large, the identification accuracy was significantly improved.
Patent Information
- Application Number
- CN202511026234.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-24
- Publication Date
- 2025-11-07
AI Technical Summary
The complex terrain and fragmented land parcels in hilly and mountainous areas make it difficult for existing algorithms to accurately identify farmland. In particular, under complex terrain and fragmented land parcels with large intra-class differences and small inter-class differences, existing technologies are unable to achieve high-precision farmland identification.
We adopted the DeepLabV3+ deep learning network model, combined with the ResNet backbone network and multi-scale prediction strategy. We used Gaofen-6 satellite imagery data for data preprocessing to construct training and testing datasets. We also incorporated dilated convolution technology to identify farmland information at the plot scale.
It achieves accurate identification of cultivated land plots in hilly and mountainous areas, with a high degree of consistency between plot boundary identification and actual values, an accuracy rate of 93%, and a recall rate of 86%. It effectively copes with complex terrain and land use types, and the identification accuracy is significantly improved when the cultivated land area exceeds 100 square meters.
Smart Images
Figure CN120912965A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the field of intelligent agriculture, and particularly relates to a hilly and mountainous area cultivated land plot accurate identification method based on DeepLabV3+. BACKGROUND
[0002] Agricultural mechanization and agricultural equipment are important foundations for transforming the mode of agricultural development and improving rural productivity. Without agricultural mechanization, there is no modernization of agriculture and rural areas. Since 2004, China's agricultural mechanization has made great progress. The comprehensive mechanization rate of crop cultivation and harvesting has increased from 34% in 2004 to 72% in 2024, which has played an important role in promoting agricultural efficiency. However, the regional imbalance of agricultural mechanization is still prominent, and hilly and mountainous areas are the main short board of agricultural mechanization development. In order to accelerate the transformation and upgrading of agricultural mechanization and agricultural equipment industry, the state has proposed to focus on supporting the transformation of farmland "mechanization" in hilly and mountainous areas, expanding the use of large and medium-sized agricultural machinery, and accelerating the improvement of the weak foundation of agricultural mechanization in hilly and mountainous areas. It is necessary to scientifically investigate the farmland mechanization operation conditions and scientifically evaluate the farmland mechanization level in different regions, and based on the evaluation results, the farmland mechanization transformation project is accurately configured. The existing technology of farmland mechanization evaluation faces the following difficulties:
[0003] 1. Challenge of complex terrain and plot fragmentation:
[0004] · Terrain interference: The terrain in hilly and mountainous areas is large, which causes serious shadows and occlusions in optical images (especially the difference between sunny slopes and shady slopes), and the radar image is affected by terrain geometric distortion, increasing the difficulty of extracting farmland features.
[0005] · Plot fragmentation: The distribution of cultivated land is scattered, the plot area is small, the shape is irregular, and the boundary is fuzzy (often interlaced with forest land and grassland), and the existing algorithms (such as CNN, FCN) have insufficient recognition accuracy for small and dense objects with unclear boundaries.
[0006] 2. Recognition dilemma of "large intra-class difference and small inter-class difference":
[0007] (1) Complex planting patterns: There are many crop types and complex planting structures (such as intercropping and relay intercropping), which cause large differences in spectral and backscattering characteristics within the cultivated land (large intra-class difference).
[0008] (2) Confusion with background: Non-cultivated land (such as fallow land, bare land, and specific grassland) has spectral and texture characteristics similar to cultivated land at certain time phases (small inter-class difference), which is easy to misjudge. SUMMARY
[0009] The application obtains remote sensing images of a research area through remote sensing image data of Gao Fen 6, and constructs a training and test data set, carries out research on a plot scale cultivated land information recognition method based on a DeepLab V3+ deep learning network model, adopts ResNet as a backbone network, and fuses a multi-scale prediction strategy and a hollow convolution technology, and the applicability of the plot recognition method is verified through comparison tests of various image conditions.
[0010] Technical scheme:
[0011] The application discloses a hilly and mountainous cultivated land plot precision recognition method based on DeepLabV3+.
[0012] S1, obtaining remote sensing images of Gao Fen 6 satellite with a spatial resolution of 2 m, and performing data preprocessing;
[0013] S2, sample data set making: taking color and texture differences between ridges and plots as features; adopting artificial visual interpretation technology to delineate cultivated land boundaries as label data of the data set; and dividing training sets and test sets;
[0014] S3, constructing and training a DeepLabV3 deep learning network model based on a PyTorch framework to obtain an optimal model;
[0015] S4, inputting the remote sensing images of the hilly and mountainous area after data preprocessing, and using the optimal model to perform cultivated land plot precision recognition.
[0016] Preferably, in S1, the data preprocessing includes radiation calibration, atmospheric correction and orthorectification.
[0017] Preferably, in S2, the sample data is subjected to data enhancement processing, including horizontal flipping, vertical flipping, diagonal line mirroring and counterclockwise and clockwise rotation by 90 degrees.
[0018] Preferably, in S3, the backbone network adopted is ResNet, and model training is performed for 150 epochs, and the number of training samples in each batch is set to 2.
[0019] Preferably, in S3, the backbone network adopted is ResNet34.
[0020] Preferably, in S3, the network structure of the DeepLabV3 deep learning network model selects and uses convolution with different hollow rates in parallel, simulates receptive fields with different spatial resolutions under the premise of not losing resolution, and effectively captures context information.
[0021] Preferably, in S3, the network structure of the DeepLabV3 deep learning network model splices and fuses high-level features rich in semantics and low-level features rich in details, so that the final prediction result is both semantically accurate and boundary clear.
[0022] Preferably, in S3, the network structure of the DeepLabV3 deep learning network model gradually restores the spatial resolution through upsampling operation, and finally obtains a segmentation map matching the size of the input image.
[0023] Preferably, in S4, the input plot area in the hilly and mountainous area is > 100 square meters.
[0024] Advantages of the present application
[0025] (1) The DeepLabV3+ model combined with the ResNet34 backbone network performs well in the task of extracting cultivated land in hilly areas, and the recognition degree of plot boundaries is high. The accuracy rate reaches 93%, and the recall rate reaches 86%. The DeepLabV3+ model has significant advantages in dealing with complex terrain and land use types in hilly areas, and can effectively deal with the problem of fragmented distribution of cultivated land.
[0026] (2) The plot area has a significant impact on the recognition accuracy of the model. When the cultivated land area exceeds 100 square meters, the recognition accuracy is improved to more than 80%. BRIEF DESCRIPTION OF DRAWINGS
[0027] Figure 1 It is a comparison schematic diagram of image data before and after preprocessing in the embodiment.
[0028] Figure 2 It is a network structure diagram of DeepLabv3+ in the embodiment.
[0029] Figure 3 It is a cultivated land recognition effect schematic diagram in the embodiment. DETAILED DESCRIPTION
[0030] The present application will be further described below in conjunction with the embodiments, but the protection scope of the present application is not limited thereto:
[0031] 1.1 Research area overview
[0032] Guanghan City in Sichuan Province (104°28'E, 30°96'N) was chosen as the study area. The study area is located in the northeast of Chengdu Plain, the west of Longquan Mountain, with a total area of 548.69 square kilometers. The region belongs to the subtropical humid climate zone, with a mild climate, distinct seasons, an annual precipitation of 890 mm, and an average annual temperature of 16.3°C. The test area has a good agricultural production foundation, with widely distributed cultivated land within the county, with a total of 212,000 mu of cultivated land, of which more than 50 mu of scale grain area exceeds 160,000 mu. The study area has diverse land features, with fragmented and scattered cultivated land, and land use with typical hilly region characteristics, which is conducive to testing the accuracy of deep learning models in remote sensing feature recognition.
[0033] 1.2 Data preprocessing
[0034] The image data comes from GF-6 satellite with a spatial resolution of 2m, and the image download time is May 2022. The ENVI 5.3 software data processing tool is used for image data preprocessing such as radiation calibration, atmospheric correction, orthorectification, etc. The comparison before and after processing is shown in FIG. 1. Figure 1
[0035] 1.3 Sample data set making
[0036] According to the geographical environment, topographic features and land use types of the study area, artificial visual interpretation technology is used to obtain the ground truth (GT) sample label data required for model construction, which covers multiple types of cultivated land. The label data overlap rate is set to 10%, and the sliding window cropping operation is performed on the remote sensing image and the label data, with the size of the cropped block being 256x256 pixels and the step being 128x128 pixels. Finally, 628 images are obtained. To prevent overfitting during training and enhance the robustness and generalization ability of the model, the original sample data is subjected to data augmentation processing through five ways of horizontal flipping, vertical flipping, diagonal mirroring, and counterclockwise and clockwise rotation by 90 degrees, a total of 3140 images are generated. The sample data is divided into training set and test set by random sampling method, of which the training set accounts for 90% to ensure that the model can learn the diversified features of cultivated land, and the test set accounts for 10% to evaluate the generalization ability of the model on independent data sets that do not participate in training, and verify the effectiveness and robustness of the model.
[0037] 2.1 Hardware environment and training parameter setting
[0038] 1) Hardware configuration. The computing device used in this study has a memory capacity of 8 GB, equipped with a central processing unit of Intel(R) Core(TM) i5-8250U CPU @ 1.60 GHz, and a 4 GB video memory NVIDIA GeForce GTX 1050 graphics card to realize model training.
[0039] 2) Training parameter configuration. In this experiment, the DeepLab deep learning network model was built and trained based on the PyTorch framework, and the parameters were adjusted. The backbone network used is ResNet, and the model training is performed for a total of 150 epochs, with 2 training samples per batch.
[0040] 2.2 Cultivated land recognition network model
[0041] The DeepLabv3+ semantic segmentation network model used integrates the dilated convolution technology, which can capture context information at different scales through the atrous spatial pyramid pooling module, thereby more accurately identifying the size, shape and direction of objects, improving the ability to identify details, and thus improving the accuracy of segmentation. In addition, this model uses an encoder-decoder structure, integrates features at different levels, and gradually reconstructs spatial information to obtain more accurate segmentation object masks. The encoder part is mainly composed of a deep convolutional neural network (DCNN) with dilated convolution. The spatial pyramid pooling module (ASPP) also uses dilated convolution, which can effectively integrate multi-scale information. The model is set up with a Decoder module to further integrate bottom and high-level features to improve the accuracy of edge segmentation. The ResNet series backbone model exhibits different functional characteristics and application scenarios depending on the depth and width differences in its model architecture. Among them, the lightweight ResNet34 model is particularly suitable for resource-constrained devices or application environments that require fast inference. Combined with the model network structure shown in the figure, the use process specifically includes: Figure 2
[0042] 1. Input image:
[0043] An original image is input into the network.
[0044] 2. Basic feature extraction:
[0045] The image is first subjected to a powerful convolutional neural network backbone for feature extraction.
[0046] This backbone network generates feature maps at different stages (different spatial resolutions, different semantic levels).
[0047] The "one-stage convolution", "two-stage convolution",... "five-stage convolution" marked in the figure refer to the feature maps output by the backbone network at different levels. Usually:
[0048] “One stage”: high resolution, low-level semantics (edges, textures).
[0049] “Five stage”: low resolution, high-level semantics (object categories).
[0050] Intermediate stages: resolution gradually decreases, semantic level gradually increases.
[0051] 3. Multi-scale context information extraction (ASPP-Atrous Spatial Pyramid Pooling):
[0052] The 1x1 convolution, 3x3 convolution with dilation rate 6, 3x3 convolution with dilation rate 12, and 3x3 convolution with dilation rate 18 parts in parallel in the flowchart constitute the ASPP module.
[0053] Input: usually the output of the deepest layer of the backbone network (the most semantic information but the lowest resolution) (e.g., the output of the “five-stage convolution”).
[0054] Operation:
[0055] 1x1 convolution: standard point convolution, used for feature transformation and channel adjustment.
[0056] 3x3 convolution with dilation rate 6: using a larger dilation rate (dilation rate = 6), the receptive field is very large, capturing very distant context information in the image.
[0057] 3x3 convolution with dilation rate 12: using a larger dilation rate (dilation rate = 12), the receptive field is larger, capturing more global context information.
[0058] 3x3 convolution with dilation rate 18: using the largest dilation rate (dilation rate = 18), the receptive field is the largest, capturing the most global context information.
[0059] Purpose: using different dilation rate convolutions in parallel, it is equivalent to using different scales of “windows” to observe the image on the same feature layer, so as to capture the context information of different scales around the object itself (small range of adjacent objects, medium range of scenes, large range of global layout) at the same time. This is crucial for understanding the semantics of pixels in the overall scene (e.g., distinguishing between “cars on the road” and “cars in the window”).
[0060] 4. ASPP feature fusion:
[0061] The first splicing operation marked in the figure.
[0062] Input: The output feature maps of the four parallel paths in the ASPP module (1x1 convolution, dilation rate 6 convolution, dilation rate 12 convolution, dilation rate 18 convolution).
[0063] Operation: Concatenate the four feature maps together along the channel dimension.
[0064] Purpose: To fuse the feature information extracted from different receptive fields (different scale contexts).
[0065] 5. Preliminary feature fusion and dimension reduction:
[0066] The concatenation operation is followed by a 1x1 convolution.
[0067] Input: The feature maps obtained after the previous concatenation (with a large number of channels).
[0068] Operation: 1x1 convolution.
[0069] Purpose:
[0070] Dimension reduction: Reduce the number of channels of the concatenated feature maps to reduce computational complexity.
[0071] Feature fusion: Further fuse the information from different paths of the ASPP.
[0072] 6. Introduce low-level detail information (skip connection):
[0073] The second concatenation operation marked in the figure.
[0074] Input:
[0075] The feature map after the 1x1 convolution in the previous step (rich in multi-scale context information, but low in spatial resolution).
[0076] The feature map output by the earlier stage (usually "stage one" or "stage two") of the backbone network (high spatial resolution, contains rich edge, texture, and other detail information, but low semantic level).
[0077] Operation: Concatenate the two feature maps along the channel dimension.
[0078] Key steps: In order to be able to concatenate, the low-level feature map usually needs to be adjusted in channel number by 1x1 convolution (possibly implied after "stage one convolution" in the figure) to match the channel number requirement of the high-level feature map. At the same time, the high-level feature map needs to be upsampled to the same spatial resolution as the low-level feature map.
[0079] Purpose: Combine the rich semantic information of high-level features with the high-resolution detail information of low-level features, preparing for the final fine segmentation. This is a key improvement of modern segmentation models like DeepLabv3+.
[0080] 7. Decoder preliminary fusion:
[0081] Following the second concatenation operation is a 3x3 convolution (annotated below the concatenation in the figure).
[0082] Input: The feature map obtained from the previous step of concatenation (combining context and details).
[0083] Operation: 3x3 convolution.
[0084] Purpose: Further integration and refinement of the fused features.
[0085] 8. First upsampling:
[0086] 4x upsampling operation.
[0087] Input: The feature map refined by the 3x3 convolution in the previous step.
[0088] Operation: Increase the spatial resolution of the feature map by 4 times (e.g., from 1 / 16 of the original size to 1 / 4 of the original size). Common methods include bilinear interpolation, transpose convolution, etc.
[0089] Purpose: Restore the spatial size, preparing for the final output.
[0090] 9. Introduce more mid-level details (optional skip connection):
[0091] The third concatenation operation annotated in the figure.
[0092] Input:
[0093] The feature map after the previous upsampling step (combining high-level semantics and part of low-level details).
[0094] The feature map output by the intermediate stage (e.g., "three-stage" or "four-stage") of the backbone network (medium resolution, medium semantic level, containing some structural information).
[0095] Operation: Concatenate the two feature maps along the channel dimension. Similarly, the backbone feature map may need channel adjustment, and the upsampling feature map needs resolution matching (usually the resolution is close here).
[0096] Purpose: Further integrate more mid-level information, provide richer structural and contour clues, and improve segmentation accuracy (especially at object boundaries).
[0097] 10. Final decoder fusion and refinement:
[0098] Following the third concatenation operation is another 3x3 convolution (annotated below the concatenation in the figure).
[0099] Input: Feature maps from the previous concatenation (with more information fused in).
[0100] Operation: 3x3 convolution.
[0101] Purpose: To perform the final integration and refinement of the fused features.
[0102] 11. Second upsampling (final resolution recovery):
[0103] 4x upsampling operation.
[0104] Input: Feature maps from the previous 3x3 convolution refinement.
[0105] Operation: Upsample the spatial resolution of the feature maps by another factor of 4 (e.g., from 1 / 4 size to the original input image size).
[0106] Purpose: To restore the feature maps to the resolution of the original input image.
[0107] 12. Final classification layer:
[0108] Final two 1x1 convolutions.
[0109] Input: Feature maps upsampled to the original resolution.
[0110] Operation:
[0111] First 1x1 convolution: Usually used to further adjust the number of channels, possibly with some degree of feature fusion.
[0112] Second 1x1 convolution: This is the output layer of the network. Its number of channels equals the number of classes to be segmented (+1 if background is included).
[0113] Output: A three-dimensional tensor [Height, Width, Num_Classes]. For each spatial location (H, W), this tensor contains Num_Classes values in the channel dimension. Typically, these values are converted to probabilities of each pixel belonging to each class through a Softmax function (or Sigmoid function if it's a multi-label). The final segmentation map is the class with the highest probability for each pixel.
[0114] The network structure proposed in this application has the following characteristics:
[0115] 1. Atrous Spatial Pyramid Pooling (ASPP): Convolution with different atrous rates is used in parallel to simulate receptive fields of different spatial resolutions without losing resolution (without downsampling), effectively capturing contextual information.
[0116] 2. High and low-level feature fusion: The high-level features rich in semantics (after ASPP) are spliced and fused with the low-level features rich in details, so that the final prediction result is both semantically accurate and boundary clear.
[0117] 3. Stepwise upsampling: The spatial resolution is gradually restored through upsampling operations, and finally a segmentation map matching the size of the input image is obtained.
[0118] 2.3 Evaluation method of recognition accuracy
[0119] In order to quantitatively evaluate the accuracy of land parcel recognition, the classification accuracy quantitative evaluation method based on confusion matrix is adopted, and the accuracy of land parcel recognition is comprehensively evaluated by two indicators of precision and recall. Precision and recall are commonly used indicators to evaluate the performance of classification model, and their calculation methods are as follows:
[0120] (1) Precision (Precision) is the proportion of actual positive examples in all samples predicted as positive examples (Positive) by the model, which measures the accuracy of the model in positive example prediction. The calculation formula is as follows:
[0121] Precision = TP / (TP+FP) (1)
[0122] (2) Recall (Recall) is the proportion of samples successfully predicted as positive examples by the model in all actual positive examples. It measures the recognition ability of the model to positive examples, and the calculation formula is as follows:
[0123] Recall = TP / (TP+FN) (2)
[0124] In formula 1 and formula 2, TP (True Positive) represents the number of positive samples predicted as positive, i.e. the number of samples whose recognition result is cultivated land and actual cultivated land; FP (False Positive) represents the number of negative samples predicted as positive, i.e. the number of samples whose recognition result is cultivated land but actual non-cultivated land; FN (False Negative) represents the number of positive samples predicted as negative, i.e. the number of samples whose recognition result is non-cultivated land but actual cultivated land.
[0125] 2.4 Comparative test
[0126] (1) Comparison of backbone network
[0127] In the field of target detection, DeepLabV3+ has achieved significant technological breakthroughs, one of which is the adoption of ResNet as its backbone network architecture. Under the framework of DeepLabV3+, ResNet is responsible for performing image feature extraction tasks. This network combines multi-scale prediction strategies and the technology of atrous convolution, significantly improving the efficiency of target detection. ResNet series backbone models exhibit different functional characteristics and application scenarios depending on the depth and width of their network architecture. ResNet34, ResNet50, and ResNet101 will be used to train models and test their accuracy in cultivated land boundary recognition.
[0128] (2) Cultivated land coverage state test
[0129] Given the complexity of agricultural planting systems in hilly and mountainous areas, within the same geographical region and during the same period, there may be heterogeneity between planting and non-planting states, which challenges the consistency of cultivated land recognition accuracy. Therefore, based on the characteristics of sample images, three types of images, including full crop coverage, no crop coverage, and partial coverage, were randomly selected. Two sheets of each type of image were randomly selected, and the sheet size was completely consistent. By comparing the experimental results of different types of images, the influence of coverage type on recognition accuracy was quantitatively analyzed.
[0130] (3) Fragmentation degree of land distribution test
[0131] The distribution of cultivated land in hilly and mountainous areas is characterized by significant fragmentation, with a small number of contiguous distribution and a large number of scattered distribution. To explore the applicability of deep learning technology in the case of uneven distribution of cultivated land, samples were selected according to the proportion of cultivated land in the entire sheet area of 100%, 80%, 70%, 60%, and 40%. The trained model was used for testing to verify the influence of fragmentation on accuracy.
[0132] (4) Land area sensitivity test
[0133] In hilly and mountainous areas, land plots are generally small and have large differences in size, ranging from large farmland of several thousand square meters to scattered cultivated land of only a few tens of square meters. To comprehensively evaluate the performance of the DeepLabV3+ deep learning model in extracting cultivated land of different sizes, and to study the correlation between cultivated land size and experimental results, the scale was set to 1:7000 in ArcgisPro maps, the resolution was set to 128x128 pixels, and the sheet area was set to 4600m2. The average cultivated land area between 32m2 and 182m2 was selected for comparative experimental research to verify the influence of cultivated land area on model recognition accuracy.
[0134] 3.1 Land identification results
[0135] The cultivated land recognition network model based on DeepLabV3+ is constructed, ResNet is adopted as the backbone model, and the data set established by the method in item 1.3 is used to train and test the model respectively to obtain the recognition result of the sample cultivated land in the study area, and part of the sample recognition effect is shown in Figure 3 The deep learning network can effectively identify the range of planar cultivated land. Under the condition of limiting the range of cultivated land, the extracted block-scale cultivated land boundary is located inside the cultivated land, avoiding the misidentification of the boundary of the land outside the cultivated land, and significantly reducing the interference of similar land boundary noise such as road boundary or building boundary in the extraction process. However, the extraction result of block-scale cultivated land boundary is affected by the distribution condition of cultivated land. When processing the boundary of the land block which is concentrated, continuous and has very similar spectral and texture characteristics, the recognition effect is poor.
[0136] In order to explore the stability of the model, comparative tests of sample quantity from 100 to 2000 were carried out, and the experimental results are shown in Table 1.
[0137] Table 1 Sample quantity and experimental accuracy
[0138]
[0139] The results show that when the sample quantity reaches 1300, the model recognition accuracy is significantly improved, reaching 90.4%; when the sample quantity is 1500, the model recognition accuracy reaches 93.1%; and when the sample quantity increases to 2000, the model recognition accuracy is 93%, which does not further improve compared with the sample quantity of 1500, and the model accuracy tends to be stable.
[0140] 3.2 Lightweight backbone model has good recognition accuracy
[0141] According to the test method described in section 2.4, three kinds of backbone models ResNet34, ResNet50 and ResNet101 were selected for comparative experiment, and the evaluation results are shown in Table 2.
[0142] Table 2 Evaluation results of cultivated land area accuracy of different backbone models
[0143]
[0144] The model using ResNet34 as the backbone network showed higher accuracy and better results in the experiment. Its precision was 0.93%, and its recall was 0.86%, both of which exceeded the ResNet50 and ResNet101 backbone models. The efficiency and accuracy of the ResNet34 model make it perform well in processing remote sensing image data of complex topography in hilly areas, especially in resource-constrained and fast inference scenarios. In addition, the stability and robustness of the ResNet34 model provide a strong guarantee for subsequent experimental results, ensuring the accuracy and reliability of the plot extraction.
[0145] 3.3 The larger the plot area, the higher the recognition accuracy
[0146] According to the test method described in 2.4, sample images with an average plot area between 32m 2 and 182m 2 were selected for comparative experiments, and the test results are shown in Table 3.
[0147] Table 3 Extraction accuracy of different cultivated land areas
[0148]
[0149] As shown in Table 3, the recognition accuracy of the plot decreases as the plot area decreases. When the cultivated land area is less than 40 square meters, the recognition accuracy is less than 44%, although the existence of cultivated land can be identified, but it cannot achieve accurate segmentation of the plot boundary. When the cultivated land area exceeds 100 square meters, the recognition accuracy is improved to more than 80%, at which time not only can the cultivated land be accurately identified, but also the plot boundary can be effectively segmented. When the cultivated land area reaches 167 square meters, the recognition accuracy reaches a peak, achieving accurate recognition of cultivated land and fine segmentation of different cultivated lands. It is not difficult to find that when the plot area reaches or exceeds 100m 2 , the model recognition effect is ideal, which can accurately identify the plot boundary line and achieve effective segmentation. However, when the cultivated land unit area decreases to 40m 2 or below, the edge segmentation effect decreases significantly, although the cultivated land area can still be accurately identified, but the adhesion between cultivated lands is significantly enhanced. Therefore, when using the DeepLabV3+ deep learning model to extract cultivated land, the approximate cultivated land area should be evaluated first to avoid poor recognition effect due to too small cultivated land area.
[0150] The specific embodiments described herein are merely illustrative of the present application. Those skilled in the art can make various modifications or supplements to the described specific embodiments or use similar ways to replace them without departing from the spirit of the present application or exceeding the scope defined by the appended claims.
Claims
1. A DeepLabV3+-based hilly and mountainous farmland plot accurate identification method, characterized by It comprises the following steps: S1, obtaining a remote sensing image of Gao Fen 6 satellite with a spatial resolution of 2m, and performing data preprocessing; S2, sample data set making: taking the color and texture difference between the ridge and the field block as the feature; adopting artificial visual interpretation technology to delineate the cultivated land boundary as the label data of the data set; and dividing the training set and the test set; S3, constructing and training a DeepLabV3 deep learning network model based on the PyTorch framework to obtain an optimal model; S4, inputting the remote sensing image of the hilly and mountainous area after data preprocessing, and using the optimal one to accurately identify the cultivated land block.
2. The method of claim 1, wherein In S1, the data preprocessing includes radiation calibration, atmospheric correction, and orthorectification.
3. The method of claim 1, wherein In S2, the sample data is subjected to data enhancement processing, including horizontal flip, vertical flip, diagonal mirror image, and 90-degree counterclockwise and clockwise rotation.
4. The method of claim 1, wherein In S3, the backbone network used is ResNet, and the model training is performed for a total of 150 epochs, and the number of training samples per batch is set to 2.
5. The method of claim 4, wherein In S3, the backbone network used is ResNet34.
6. The method of claim 5, wherein In S3, the network structure of the DeepLabV3 deep learning network model uses convolution with different hole rates in parallel, which simulates the receptive field of different spatial resolutions without losing resolution, effectively capturing context information.
7. The method of claim 5, wherein In S3, the network structure of the DeepLabV3 deep learning network model fuses the high-level features rich in semantics with the low-level features rich in details, so that the final prediction result is both semantically accurate and boundary clear.
8. The method of claim 5, wherein In S3, the network structure of the DeepLabV3 deep learning network model gradually restores the spatial resolution through upsampling operation, and finally obtains a segmentation map matching the size of the input image.
9. The method of claim 1, wherein In S4, the area of the field block in the hilly and mountainous area is >100㎡.