Farmland mulching film remote sensing extraction method based on multi-mode semantic learning

Abstract

The invention discloses a farmland mulching film remote sensing extraction method based on multi-mode semantic learning. The multi-modal semantic learning model is composed of a coarse granularity feature extraction module, a medium granularity feature extraction module, a fine granularity feature extraction module, a multi-level granularity feature fusion module of a multispectral remote sensing image, a multi-level granularity feature fusion module of a synthetic aperture radar remote sensing image and a feature fusion module of a multi-modal remote sensing image. A farmland mulching film identification module; setting a multi-modal remote sensing image input size of the multi-modal semantic learning model, wherein the multi-modal semantic learning model adopts a multi-spectral remote sensing image sample set A and a synthetic aperture radar remote sensing image sample set B as multi-spectral remote sensing image input and synthetic aperture radar remote sensing image input; according to different image texture characteristics of the farmland mulching film on the farmland scale and the field scale, multi-mode remote sensing image feature information of different granularity levels is extracted and fused, and the process of extracting the farmland mulching film from the multi-mode remote sensing image is achieved.

Description

Technical Field

[0001] The present invention relates to the field of deep learning technology, and in particular to a method for extracting farmland mulch remote sensing based on multimodal semantic learning. Background Art

[0002] Laying plastic mulch on farmland is an important step in crop planting in arid and semi-arid areas. Since plastic mulch has the function of heat preservation and moisture retention, it can greatly reduce the evaporation loss of soil moisture during the sowing and germination stages of crops and improve the irrigation efficiency of the land. However, since the material of the plastic mulch is a plastic product, it is difficult to be naturally degraded by the soil, so it will cause pollution to the soil and the natural environment. In recent years, as my country has vigorously promoted the protection of black soil and the development of green agriculture, exploring large-scale and efficient monitoring methods for farmland mulch has become one of the key scientific issues that need to be solved urgently.

[0003] Traditional monitoring of farmland mulch film is mainly based on field investigations, which mainly relies on investigators to visit and measure specific farmland areas. Although the specific status of the mulch film can be recorded on the spot, the monitoring efficiency is low and it is difficult to meet the current large-scale farmland mulch film monitoring needs.

[0004] In recent years, with the development of space science and technology, space remote sensing technology has provided a new way to carry out large-scale surface object monitoring. Remote sensing satellite images can be divided into two different modalities of data: active remote sensing and passive images. Active remote sensing imaging has the advantages of strong penetration and is not easily disturbed by clouds and fog, but its spatial resolution is low, such as synthetic aperture radar remote sensing images. Passive remote sensing data has the characteristics of diverse spatial resolution and high spectral resolution, but it is easily disturbed by clouds and fog. Passive remote sensing data mainly includes multispectral remote sensing images. If remote sensing images of different modalities can be jointly analyzed, that is, multimodal remote sensing image analysis can be carried out, it will be expected to improve the effectiveness of remote sensing image content interpretation.

[0005] In addition, in terms of remote sensing extraction of farmland mulch, the current mainstream farmland mulch extraction mainly relies on manual methods. On the one hand, a lot of preliminary training is required for personnel engaged in remote sensing manual object annotation. On the other hand, personnel engaged in remote sensing manual object annotation have problems such as mislabeling and missing labels due to work fatigue. Therefore, remote sensing extraction of farmland mulch based on manual methods consumes a lot of manpower and material costs, which is difficult to meet the current urgent need for large-scale and efficient monitoring of farmland mulch.

[0006] At present, artificial intelligence technology represented by deep learning is developing rapidly. In particular, semantic learning models for target extraction have been applied to traditional agricultural information statistics, industrial process control, commercial data analysis, image text recognition and other fields. Compared with common types of objects such as buildings and roads in remote sensing objects, farmland mulch has a more complex image texture due to its high transmittance and narrow width. It is necessary to carry out targeted design of the network structure of the semantic learning model to adapt to the characteristics of multimodal remote sensing images with different spectral resolutions and spatial resolutions, including multispectral remote sensing images and synthetic aperture radar remote sensing images, so as to achieve remote sensing accurate extraction of farmland mulch.

[0007] In view of the problems of high labor cost and low efficiency in the existing manual annotation methods for farmland remote sensing extraction, as well as the complex texture of farmland mulch images in remote sensing images, the present invention proposes a farmland mulch remote sensing extraction method based on multimodal semantic learning. From the perspective of extracting feature information of multimodal remote sensing images at different granularity levels, the present invention designs and constructs a multimodal semantic learning model to realize the automated farmland mulch remote sensing extraction process and improve the efficiency of farmland mulch monitoring. Summary of the invention

[0008] The technical problem to be solved by the present invention is the low efficiency of the existing manual annotation method for remote sensing extraction of farmland mulch films. Therefore, a remote sensing extraction method for farmland mulch films based on multimodal semantic learning is proposed to realize the automated remote sensing extraction process of farmland mulch films and improve the monitoring efficiency of farmland mulch films.

[0009] The technical solution adopted by the present invention to solve the above technical problems is: a method for extracting farmland mulch remote sensing images based on multimodal semantic learning, in particular, the proposed multimodal semantic learning model can extract and fuse multimodal remote sensing image feature information of different granularity levels according to the different image texture characteristics of farmland mulch at the farmland scale and field scale, thereby realizing the process of extracting farmland mulch from multimodal remote sensing images. The technical solution process of the present invention is as follows Figure 1 The method of the present invention comprises the following steps:

[0010] 1) Multimodal remote sensing image reading:

[0011] The multimodal remote sensing image sample set includes multispectral remote sensing image sample set A and synthetic aperture radar remote sensing image sample set B, a total of two modal remote sensing image sample data. The sample true value of the farmland mulch area in the multimodal remote sensing image sample set is annotated to provide image samples and annotated true values ​​for the next step of model training.

[0012] 2) Multimodal semantic learning model training:

[0013] Figure 2The schematic diagram of the structure of the multimodal semantic learning model is shown in Figure 2. The multimodal semantic learning model consists of Figure 3 The coarse-grained feature extraction module shown, Figure 4 The medium-granularity feature extraction module shown, Figure 5 The fine-grained feature extraction module shown, Figure 6 The multi-level granularity feature fusion module of the multispectral remote sensing image shown, Figure 7 The multi-level granularity feature fusion module of the synthetic aperture radar remote sensing image shown in the figure, Figure 8 The feature fusion module of multimodal remote sensing images shown in Fig. 9 The farmland mulch recognition module shown in the figure consists of 7 types of modules. Each module consists of a convolution layer, a residual convolution layer, a pooling layer, a merging layer, and an upsampling layer. The kernel function size of the convolution layer includes 1×1 and 3×3, the kernel function size of the residual convolution layer includes 1×1 and 3×3, the pooling window size of the pooling layer is 2×2, and the sampling multiple of the upsampling layer is 2.

[0014] First, the multimodal remote sensing image input size of the multimodal semantic learning model is set: the multispectral remote sensing image is 128 pixels × 128 pixels × 9 channels, and the synthetic aperture radar remote sensing image is 128 pixels × 128 pixels × 2 channels.

[0015] based on Figure 2 The multimodal semantic learning model shown uses a multispectral remote sensing image sample set A and a synthetic aperture radar remote sensing image sample set B as the multispectral remote sensing image input and synthetic aperture radar remote sensing image input of the model.

[0016] The multispectral remote sensing images are respectively input into the coarse granularity feature extraction module, the medium granularity feature extraction module, and the fine granularity extraction module to extract features of different granularity levels, and then the multi-level granularity feature extraction results are input into the multi-level granularity feature fusion module of the multispectral remote sensing image for feature fusion to obtain the multi-level granularity fusion features of the multispectral remote sensing image. The synthetic aperture radar remote sensing images are respectively input into the coarse granularity feature extraction module and the fine granularity extraction module to extract features of different granularity levels, and then the multi-level granularity feature extraction results are input into the multi-level granularity feature fusion module of the synthetic aperture radar remote sensing image for feature fusion to obtain the multi-level granularity fusion features of the synthetic aperture radar remote sensing image. Furthermore, the multi-level granularity fusion features of the multispectral remote sensing images and the multi-level granularity fusion features of the synthetic aperture radar remote sensing images are input into the feature fusion module of the multimodal remote sensing images for feature fusion to obtain the fusion features of the multimodal remote sensing images. Finally, the fusion features of the multimodal remote sensing images are input into the recognition module of the farmland mulch film to obtain the recognition results of the farmland mulch film.

[0017] In the process of designing the network layer structure of each module, the present invention fully considers the differences in spectral resolution and spatial resolution of multimodal remote sensing images including multispectral remote sensing images and synthetic aperture radar remote sensing images, as well as the characteristics of the two remote sensing images in terms of cloud and fog penetration and spectral band range, and then specifically designs the convolution kernel size setting of the convolution layer, the alternating arrangement of the residual convolution layer and the convolution layer, and the progressive layout of the convolution layer and the pooling layer, so as to achieve better farmland mulch film recognition effect. Specifically, compared with synthetic aperture radar remote sensing images, multispectral remote sensing images have higher spatial resolution and spectral resolution, so in Figure 2 In the multimodal semantic learning model shown in Figure 1, multispectral remote sensing images are input into feature extraction modules with three levels of granularity, namely coarse, medium and fine, to extract feature information at different granularity levels. In contrast, synthetic aperture radar remote sensing images have relatively low spatial resolution and spectral resolution, so only two levels of granularity feature extraction modules are designed, namely coarse and fine. In addition, in the feature extraction modules at each granularity level, the convolution layer adopts a serial layout, and the convolution kernel size is 1×1 and 3×3. The purpose is to use a small-scale convolution kernel of 1×1 to obtain local feature information and a large-scale convolution kernel of 3×3 to obtain global feature information. At the same time, the residual convolution layer is used to improve the robustness of network parameter transfer and improve the efficiency of model training.

[0018] In the multi-level granularity feature fusion module of multispectral remote sensing images and the multi-level granularity feature fusion module of synthetic aperture radar remote sensing images, according to the differences in the diversity of granularity levels of feature information between the two, granularity feature merging and fusion processes including three upsampling processes and two upsampling processes are adopted respectively, aiming to better adapt to the respective characteristics of multispectral remote sensing images and synthetic aperture radar remote sensing images.

[0019] In the feature fusion module of multimodal remote sensing images, a targeted design is also carried out based on the characteristics of multimodal remote sensing images, which have high spatial resolution of multispectral remote sensing images and low spatial resolution of synthetic aperture radar remote sensing images. That is, the multi-level granularity fusion features of multispectral remote sensing images are subjected to a convolution layer containing two consecutive layers of 3×3 convolution kernels for more detailed feature extraction, while the multi-level granularity fusion features of synthetic aperture radar remote sensing images only use a single layer of 3×3 convolution kernel convolution layer and residual convolution layer to cooperate. This design can not only adapt to multimodal remote sensing images with different resolution characteristics, but also optimize the model structure, reduce the amount of model calculation, and realize the rapid and accurate remote sensing extraction of farmland mulch based on the proposed model.

[0020] In the farmland film recognition module, convolution layers and residual convolution layers composed of convolution kernels of two sizes, 3×3 and 1×1, are used. The purpose is to perform matrix operations on the fusion feature input of multimodal remote sensing images from coarse to fine, based on the different characteristics of farmland film at the farmland scale and field scale, to obtain the final farmland film recognition result.

[0021] During the model training process, the loss function is binary cross entropy, the optimizer is Adam, the number of iterations is 80, the batch size is 20, and the learning rate is 0.001. Through the above process, a trained multimodal semantic learning model is obtained.

[0022] 3) Remote sensing extraction of farmland mulch film based on multimodal semantic learning model:

[0023] The multimodal remote sensing image to be analyzed is input into the trained multimodal semantic learning model, and the extraction result of farmland mulch is obtained as output.

[0024] Compared with traditional farmland mulch monitoring and remote sensing extraction methods of farmland mulch based on manual identification, the advantages of the present invention are: traditional farmland mulch monitoring methods rely on manual field investigation and monitoring, with a small monitoring range and a long monitoring time. The remote sensing extraction method of farmland mulch based on manual identification requires preliminary training for the staff of object annotation, and there are problems such as unstable farmland mulch extraction accuracy due to staff fatigue and subjective bias in the manual identification and annotation process. In contrast, the remote sensing extraction method of farmland mulch based on multimodal semantic learning proposed in the present invention can target the different image texture characteristics of farmland mulch at the farmland scale and the field scale, and extract and fuse the multimodal remote sensing image feature information of different granularity levels, thereby realizing accurate and rapid extraction of farmland mulch from multimodal remote sensing images. The present invention effectively improves the remote sensing extraction efficiency of farmland mulch, and provides a new idea and feasible method for carrying out large-scale farmland mulch monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 is a flow chart of the method of the present invention;

[0026] Figure 2 It is a schematic diagram of the structure of the multimodal semantic learning model;

[0027] Figure 3 It is a coarse-grained feature extraction module;

[0028] Figure 4 It is a medium-granularity feature extraction module;

[0029] Figure 5 It is a fine-grained feature extraction module;

[0030] Figure 6It is a multi-level granularity feature fusion module for multispectral remote sensing images;

[0031] Figure 7 It is a multi-level granularity feature fusion module for synthetic aperture radar remote sensing images;

[0032] Figure 8 It is a feature fusion module for multimodal remote sensing images;

[0033] Fig. 9 It is the identification module of farmland mulch film; DETAILED DESCRIPTION

[0034] The present invention is further described in detail below with reference to the accompanying drawings.

[0035] The present invention designs a remote sensing extraction method for farmland mulch based on multimodal semantic learning for farmland mulch in multimodal remote sensing images, and realizes an automated remote sensing extraction process for farmland mulch. In particular, the areas where farmland mulch is located are mostly arid and complex terrain areas with complex and diverse climates, and remote sensing observations are often affected by foggy weather. From the perspective of all-weather monitoring of farmland mulch, the present invention proposes a multimodal semantic learning model network structure for remote sensing extraction of farmland mulch, effectively integrating the advantages of high spectral resolution of multispectral remote sensing images and strong cloud and fog penetration ability and resistance to cloud and fog interference of synthetic aperture radar remote sensing images, realizing the complementary advantages between multi-source remote sensing images from two different sensor sources, and innovatively realizing fast and accurate remote sensing extraction of farmland mulch.

[0036] In actual use, the computer will call the program based on the method flow of the present invention to complete the specific work of remote sensing extraction of farmland mulch film based on multimodal semantic learning. Figure 1 The flowchart of the method of the present invention is as follows:

[0037] Step 1: Read in multimodal remote sensing images:

[0038] The multimodal remote sensing image sample set includes multispectral remote sensing image sample set A and synthetic aperture radar remote sensing image sample set B, a total of two modal remote sensing image sample data. The sample true value of the farmland mulch area in the multimodal remote sensing image sample set is annotated to provide image samples and annotated true values ​​for the next step of model training.

[0039] Step 2: Multimodal semantic learning model training:

[0040] Figure 2 The schematic diagram of the structure of the multimodal semantic learning model is shown in Figure 2. The multimodal semantic learning model consists of Figure 3 The coarse-grained feature extraction module shown, Figure 4 The medium-granularity feature extraction module shown, Figure 5The fine-grained feature extraction module shown, Figure 6 The multi-level granularity feature fusion module of the multispectral remote sensing image shown, Figure 7 The multi-level granularity feature fusion module of the synthetic aperture radar remote sensing image shown in the figure, Figure 8 The feature fusion module of multimodal remote sensing images shown in Fig. 9 The farmland mulch recognition module shown in the figure consists of 7 types of modules. Each module consists of a convolution layer, a residual convolution layer, a pooling layer, a merging layer, and an upsampling layer. The kernel function size of the convolution layer includes 1×1 and 3×3, the kernel function size of the residual convolution layer includes 1×1 and 3×3, the pooling window size of the pooling layer is 2×2, and the sampling multiple of the upsampling layer is 2.

[0041] First, the multimodal remote sensing image input size of the multimodal semantic learning model is set: the multispectral remote sensing image is 128 pixels × 128 pixels × 9 channels, and the synthetic aperture radar remote sensing image is 128 pixels × 128 pixels × 2 channels.

[0042] based on Figure 2 The multimodal semantic learning model shown uses a multispectral remote sensing image sample set A and a synthetic aperture radar remote sensing image sample set B as the multispectral remote sensing image input and synthetic aperture radar remote sensing image input of the model.

[0043] The multispectral remote sensing images are respectively input into the coarse granularity feature extraction module, the medium granularity feature extraction module, and the fine granularity extraction module to extract features of different granularity levels, and then the multi-level granularity feature extraction results are input into the multi-level granularity feature fusion module of the multispectral remote sensing image for feature fusion to obtain the multi-level granularity fusion features of the multispectral remote sensing image. The synthetic aperture radar remote sensing images are respectively input into the coarse granularity feature extraction module and the fine granularity extraction module to extract features of different granularity levels, and then the multi-level granularity feature extraction results are input into the multi-level granularity feature fusion module of the synthetic aperture radar remote sensing image for feature fusion to obtain the multi-level granularity fusion features of the synthetic aperture radar remote sensing image. Furthermore, the multi-level granularity fusion features of the multispectral remote sensing images and the multi-level granularity fusion features of the synthetic aperture radar remote sensing images are input into the feature fusion module of the multimodal remote sensing images for feature fusion to obtain the fusion features of the multimodal remote sensing images. Finally, the fusion features of the multimodal remote sensing images are input into the recognition module of the farmland mulch film to obtain the recognition results of the farmland mulch film.

[0044] In the process of designing the network layer structure of each module, the present invention fully considers the differences in spectral resolution and spatial resolution of multimodal remote sensing images including multispectral remote sensing images and synthetic aperture radar remote sensing images, as well as the characteristics of the two remote sensing images in terms of cloud and fog penetration and spectral band range, and then specifically designs the convolution kernel size setting of the convolution layer, the alternating arrangement of the residual convolution layer and the convolution layer, and the progressive layout of the convolution layer and the pooling layer, so as to achieve better farmland mulch film recognition effect. Specifically, compared with synthetic aperture radar remote sensing images, multispectral remote sensing images have higher spatial resolution and spectral resolution, so in Figure 2 In the multimodal semantic learning model shown in Figure 1, multispectral remote sensing images are input into feature extraction modules with three levels of granularity, namely coarse, medium and fine, to extract feature information at different granularity levels. In contrast, synthetic aperture radar remote sensing images have relatively low spatial resolution and spectral resolution, so only two levels of granularity feature extraction modules are designed, namely coarse and fine. In addition, in the feature extraction modules at each granularity level, the convolution layer adopts a serial layout, and the convolution kernel size is 1×1 and 3×3. The purpose is to use a small-scale convolution kernel of 1×1 to obtain local feature information and a large-scale convolution kernel of 3×3 to obtain global feature information. At the same time, the residual convolution layer is used to improve the robustness of network parameter transfer and improve the efficiency of model training.

[0045] In the multi-level granularity feature fusion module of multispectral remote sensing images and the multi-level granularity feature fusion module of synthetic aperture radar remote sensing images, according to the differences in the diversity of granularity levels of feature information between the two, granularity feature merging and fusion processes including three upsampling processes and two upsampling processes are adopted respectively, aiming to better adapt to the respective characteristics of multispectral remote sensing images and synthetic aperture radar remote sensing images.

[0046] In the feature fusion module of multimodal remote sensing images, a targeted design is also carried out based on the characteristics of multimodal remote sensing images, which have high spatial resolution of multispectral remote sensing images and low spatial resolution of synthetic aperture radar remote sensing images. That is, the multi-level granularity fusion features of multispectral remote sensing images are subjected to a convolution layer containing two consecutive layers of 3×3 convolution kernels for more detailed feature extraction, while the multi-level granularity fusion features of synthetic aperture radar remote sensing images only use a single layer of 3×3 convolution kernel convolution layer and residual convolution layer to cooperate. This design can not only adapt to multimodal remote sensing images with different resolution characteristics, but also optimize the model structure, reduce the amount of model calculation, and realize the rapid and accurate remote sensing extraction of farmland mulch based on the proposed model.

[0047] In the farmland film recognition module, convolution layers and residual convolution layers composed of convolution kernels of two sizes, 3×3 and 1×1, are used. The purpose is to perform matrix operations on the fusion feature input of multimodal remote sensing images from coarse to fine, based on the different characteristics of farmland film at the farmland scale and field scale, to obtain the final farmland film recognition result.

[0048] During the model training process, the loss function is binary cross entropy, the optimizer is Adam, the number of iterations is 80, the batch size is 20, and the learning rate is 0.001. Through the above process, a trained multimodal semantic learning model is obtained.

[0049] Step 3: Remote sensing extraction of farmland mulch film based on multimodal semantic learning model:

[0050] The multimodal remote sensing image to be analyzed is input into the trained multimodal semantic learning model, and the extraction result of farmland mulch is obtained as output.

Claims

1. A remote sensing extraction method for farmland mulch film based on multimodal semantic learning, characterized in that: The following steps are involved: Step 1) Read in multimodal remote sensing images: The multimodal remote sensing image sample set includes multispectral remote sensing image sample set A and synthetic aperture radar remote sensing image sample set B, a total of two modal remote sensing image sample data; the sample true value of the farmland mulch area in the multimodal remote sensing image sample set is annotated to provide image samples and annotated true values ​​for the next step of model training; Step 2) Multimodal semantic learning model training: The multimodal semantic learning model consists of seven types of modules: coarse-grained feature extraction module, medium-grained feature extraction module, fine-grained feature extraction module, multi-level granularity feature fusion module of multispectral remote sensing images, multi-level granularity feature fusion module of synthetic aperture radar remote sensing images, feature fusion module of multimodal remote sensing images, and recognition module of farmland mulch film. First, the multimodal remote sensing image input size of the multimodal semantic learning model is set: the multispectral remote sensing image is 128 pixels × 128 pixels × 9 channels, and the synthetic aperture radar remote sensing image is 128 pixels × 128 pixels × 2 channels; The multispectral remote sensing images are respectively input into the coarse granularity feature extraction module, the medium granularity feature extraction module, and the fine granularity extraction module to extract features of different granularity levels, and then the multi-level granularity feature extraction results are input into the multi-level granularity feature fusion module of the multispectral remote sensing images for feature fusion to obtain the multi-level granularity fusion features of the multispectral remote sensing images; The synthetic aperture radar remote sensing image is respectively input into the coarse granularity feature extraction module and the fine granularity extraction module to extract features of different granularity levels, and then the multi-level granularity feature extraction result is input into the multi-level granularity feature fusion module of the synthetic aperture radar remote sensing image for feature fusion to obtain the multi-level granularity fusion feature of the synthetic aperture radar remote sensing image; further, the multi-level granularity fusion feature of the multispectral remote sensing image and the multi-level granularity fusion feature of the synthetic aperture radar remote sensing image are input into the feature fusion module of the multimodal remote sensing image for feature fusion to obtain the fusion feature of the multimodal remote sensing image; finally, the fusion feature of the multimodal remote sensing image is input into the recognition module of the farmland mulch film to obtain the recognition result of the farmland mulch film; The multimodal semantic learning model uses a multispectral remote sensing image sample set A and a synthetic aperture radar remote sensing image sample set B as multispectral remote sensing image input and synthetic aperture radar remote sensing image input to perform network training; the multimodal semantic learning model training is completed; Step 3) Remote sensing extraction of farmland mulch film based on multimodal semantic learning model: The multimodal remote sensing image to be analyzed is input into the trained multimodal semantic learning model, and the extraction result of farmland mulch is output.

2. According to claim 1, a remote sensing extraction method for farmland mulch film based on multimodal semantic learning is characterized in that: The multimodal semantic learning model consists of seven types of modules, including coarse-grained feature extraction module, medium-grained feature extraction module, fine-grained feature extraction module, multi-level granularity feature fusion module of multispectral remote sensing images, multi-level granularity feature fusion module of synthetic aperture radar remote sensing images, feature fusion module of multimodal remote sensing images, and farmland mulch film recognition module. Each module consists of a convolutional layer, a residual convolutional layer, a pooling layer, a merging layer, and an upsampling layer. Among them, the kernel function size of the convolutional layer includes 1×1 and 3×3, the kernel function size of the residual convolution layer includes 1×1 and 3×3, the pooling window size of the pooling layer is 2×2, and the sampling multiple of the upsampling layer is 2.

3. The method for extracting farmland mulch remote sensing data based on multimodal semantic learning according to claim 1, characterized in that: In the network training of the multimodal semantic learning model, the loss function is binary cross entropy, the optimizer is Adam, the number of iterations is 80, the batch number is 20, and the learning rate is 0.001; through the above process, a trained multimodal semantic learning model is obtained.

4. The method for extracting farmland mulch remote sensing data based on multimodal semantic learning according to claim 1, characterized in that: In the steps 2) and 3), a multimodal semantic learning model is used for model training and remote sensing extraction of farmland mulch film.

Images