Protective farming identification method based on unmanned aerial vehicle and satellite cross-scale sensing

Abstract

The invention belongs to the field of ground surface parameter optical remote sensing inversion, and relates to a protective farming identification method based on unmanned aerial vehicle and satellite cross-scale sensing, which comprises the following steps: collecting a ground surface image, training and predicting an improved U-Net semantic segmentation model to obtain a segmented image of straw, matching the straw coverage with a ground coordinate to obtain a registered land surface straw coverage spatial data set, calculating a remote sensing spectral index of a coordinate point of the cultivated land straw coverage by using a multispectral remote sensing image, and taking the land surface straw coverage spatial data set as a target variable; taking a remote sensing spectrum index of a coordinate point corresponding to the land surface straw coverage and wave band information of a multispectral remote sensing image as characteristic variables, inputting a target variable and the characteristic variables into an XGBoost model for training, and performing spatial prediction on the cultivated land straw coverage. According to the method, large-scale straw coverage calculation and protective farming identification can be rapidly and accurately carried out.

Description

Technical Field

[0001] The present invention belongs to the field of optical remote sensing inversion of surface parameters, and in particular relates to a conservation tillage identification method based on cross-scale sensing of unmanned aerial vehicles and satellites. Background Art

[0002] Traditional agricultural farming methods are mainly aimed at improving land productivity and meeting human food needs. However, long-term high-intensity traditional farming has a large disturbance on the soil through frequent tillage, which aggravates soil wind erosion and water erosion, leading to soil structural instability, loss of biodiversity, and decreased soil fertility. These problems seriously threaten the country's food and ecological security under the new situation. Conservation tillage is a modern farming technology system with crop straw mulching and no-till (minimum tillage) sowing as its main content. It achieves the purpose of preventing soil erosion, increasing fertilizer, conserving soil moisture, fixing carbon, and improving crop production capacity by increasing organic materials and reducing soil disturbance. At present, conservation tillage has become a key measure to ensure national food security and achieve sustainable agricultural development. In view of this, it is very important to carry out large-scale and accurate identification of conservation tillage and clarify the temporal variation characteristics of the scope of conservation tillage, especially in the main grain producing areas in Northeast China where the area of ​​conservation tillage has increased year by year under the black soil protection strategy but there is a lack of effective monitoring system.

[0003] Traditional conservation tillage identification methods often rely on field surveys, using rope-pulling methods to collect on-site data on surface straw coverage, and generally consider cultivated land with a straw coverage of not less than 30% as conservation tillage land. This type of traditional method has a complicated, time-consuming and labor-intensive survey process, making it difficult to conduct large-scale continuous monitoring. The lack of an efficient method for measuring ground straw coverage has greatly limited the identification of the scope of conservation tillage and its spatiotemporal evolution characteristics. Summary of the invention

[0004] Aiming at the technical bottleneck of lack of high-precision ground straw coverage measurement method and difficulty in identifying large-scale conservation tillage, the present invention provides a conservation tillage identification method based on UAV and satellite cross-scale sensing. The present invention is achieved in this way. A conservation tillage identification method based on UAV and satellite cross-scale sensing includes the following steps: S1 used drones to collect surface images of farmland with different straw coverage conditions at high altitudes before sowing crops; S2 preprocesses the collected images, creates training labels, improves the U-Net semantic segmentation model, and inputs the preprocessed images and training labels to train the improved U-Net semantic segmentation model; S3 Input all collected images into the trained improved U-Net semantic segmentation model for prediction, obtain the straw segmentation image, mosaic the straw segmentation image and count the ratio of the number of pixels of the straw part to the total number of pixels to calculate the straw coverage, match the straw coverage with the ground coordinates to obtain the registered surface straw coverage spatial dataset; S4 Use multispectral remote sensing images to calculate the remote sensing spectral index of the coordinate points of cultivated land straw coverage, take the surface straw coverage spatial dataset as the target variable, and take the remote sensing spectral index of the coordinate points corresponding to the surface straw coverage and the band information of the multispectral remote sensing image as the feature variables. Input the target variable and the feature variable into the XGBoost model for training to establish an XGBoost model of remote sensing spectral index, band information of multispectral remote sensing image and straw coverage. Input the band information of multispectral remote sensing image and the calculated remote sensing spectral index into the trained XGBoost model for spatial prediction of cultivated land straw coverage, and judge the plots with cultivated land straw coverage greater than or equal to 30% as protective tillage plots.

[0005] Further, the improved U-Net semantic segmentation model in S2 comprises: an encoder and a decoder, wherein the encoder sequentially passes through a first convolutional layer, a first pooling layer, a first Bottleneck residual block, a second Bottleneck residual block, a third Bottleneck residual block, a fourth Bottleneck residual block, a fifth Bottleneck residual block, a sixth Bottleneck residual block, a seventh Bottleneck residual block, an eighth Bottleneck residual block, a ninth Bottleneck residual block, a tenth Bottleneck residual block, an eleventh Bottleneck residual block, a twelfth Bottleneck residual block, a thirteenth Bottleneck residual block, a fourteenth Bottleneck residual block, a fifteenth Bottleneck residual block and a sixteenth Bottleneck residual block from the input data; The decoder includes, in order from data input to output: a first upsampling layer, a second convolutional layer, a third convolutional layer, a second upsampling layer, a fourth convolutional layer, a fifth convolutional layer, a third upsampling layer, a sixth convolutional layer, a seventh convolutional layer, a fourth upsampling layer, an eighth convolutional layer, a ninth convolutional layer and a fifth upsampling layer; The output of the first convolutional layer is jump connected to the output of the fourth upsampling layer; the output of the third Bottleneck residual block is jump connected to the output of the third upsampling layer; the output of the seventh Bottleneck residual block is jump connected to the output of the second upsampling layer; and the output of the thirteenth Bottleneck residual block is jump connected to the output of the first upsampling layer.

[0006] Furthermore, the structures of the first Bottleneck residual block, the fourth Bottleneck residual block, the eighth Bottleneck residual block and the fourteenth Bottleneck residual block include: starting from the input, they are divided into two paths, the first path includes the tenth convolutional layer, the eleventh convolutional layer and the twelfth convolutional layer, the second path includes the thirteenth convolutional layer and the RELU activation function layer, the data of the first path passes through the tenth convolutional layer, the eleventh convolutional layer and the twelfth convolutional layer in sequence, the data of the second path passes through the thirteenth convolutional layer and the first RELU activation function layer in sequence, and the twelfth convolutional layer is output to the first RELU activation function layer.

[0007] Furthermore, the structure of the Bottleneck residual block other than the first Bottleneck residual block, the fourth Bottleneck residual block, the eighth Bottleneck residual block and the fourteenth Bottleneck residual block includes: starting from the input, it is divided into two paths, the data of the first path passes through the fourteenth convolution layer, the fifteenth convolution layer and the sixteenth convolution layer in sequence, the data of the second path includes a RELU activation function layer, and the sixteenth convolution layer outputs to the second RELU activation function layer.

[0008] Furthermore, the remote sensing spectral index includes: normalized difference vegetation index, normalized difference tillage index, normalized difference straw index, shortwave infrared angle index, first normalized difference index, second normalized difference index, normalized difference senescent vegetation index, soil tillage index and dry fuel index.

[0009] Furthermore, the normalized difference vegetation index is: ; Normalized Difference Tillage Index: ; Normalized difference straw index: ; Short-wave infrared angle index: ; First normalized difference index: ; Second normalized difference index: ; Normalized difference senescent vegetation index: ; Soil tillage index: ; Dry Fuel Index: ; , , , and All are collected from Sentinel-2 multispectral data. It indicates a band with a central wavelength of 0.665 μm and a spatial resolution of 10 m; It indicates a band with a central wavelength of 0.842 μm and a spatial resolution of 10 m; It indicates a band with a central wavelength of 0.865 μm and a spatial resolution of 20 m; It indicates a band with a central wavelength of 1.610 μm and a spatial resolution of 20 m; Indicates a band with a central wavelength of 2.190μm and a spatial resolution of 20m.

[0010] Compared with the prior art, the present invention has the following beneficial effects: Compared with the traditional manual rope method to measure ground straw coverage, the use of drones combined with deep learning can quickly and cheaply obtain large-scale, high-quality, and high-precision ground straw coverage data. The organic combination of high-precision ground straw coverage data and high-resolution multispectral remote sensing images, supplemented by advanced machine learning algorithms, can quickly and accurately perform large-scale straw coverage measurement and conservation tillage identification, which is helpful for analyzing the spatiotemporal evolution characteristics of conservation tillage and accurately implementing sustainable cultivated land utilization and protection measures. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 An overall flow chart of the conservation tillage identification method using cross-scale sensing of drones and satellites provided in an embodiment of the present invention; Figure 2 An image annotation result diagram provided by an embodiment of the present invention; Figure 3 A structural diagram of an improved U-Net semantic segmentation model provided by an embodiment of the present invention; Figure 4 A schematic diagram of the structures of a first Bottleneck residual block, a fourth Bottleneck residual block, an eighth Bottleneck residual block and a fourteenth Bottleneck residual block provided in an embodiment of the present invention; Figure 5 A schematic diagram of the structures of the remaining Bottleneck residual blocks except the first Bottleneck residual block, the fourth Bottleneck residual block, the eighth Bottleneck residual block and the fourteenth Bottleneck residual block provided in the embodiment of the present invention; Figure 6 Different input images and prediction results in the ground straw segmentation result diagram provided by the embodiment of the present invention; Figure 7 In the ground straw segmentation result diagram provided in the embodiment of the present invention, the predicted image is mosaicked to obtain the predicted result of the original image taken, (a) the original image, (b) the predicted result; Figure 8 Performance evaluation of the straw coverage remote sensing inversion prediction model provided by the embodiment of the present invention; Fig. 9The conservation tillage identification result of the present invention is mapped, (a) is the final conservation tillage mapping result, (b) is the local enlarged view and the corresponding remote sensing image. DETAILED DESCRIPTION

[0012] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0013] like Figure 1 As shown in the figure, the conservation tillage identification method based on UAV and satellite cross-scale sensing includes the following steps: In order to accurately identify the tillage method used in cultivated land, S1 selected 89 representative cultivated lands in the central and western regions of Jilin Province before spring sowing, and used a Dji Air3 drone to fly to a height of 10 meters to collect surface images of cultivated land with different straw coverage conditions; S2 preprocesses the collected farmland images with different straw coverage conditions, creates training labels, improves the U-Net semantic segmentation model, and inputs the preprocessed images and training labels to train the improved U-Net semantic segmentation model; The specific steps of S2 are as follows: The original image resolution is 8064*6048, which is too large for deep learning methods, so it is cropped. A total of 60 photos with different lighting conditions, different soil moisture conditions, different soil colors and types, and different straw coverage levels were selected. Five 1024*1024 size pictures were randomly cropped from each selected photo to obtain 300 images for annotation and training. The labelme annotation tool was used to annotate the images, where the straw part was labeled as 1 and the rest were labeled as _background_. The annotation results are shown in the figure below. Figure 2 As shown; Due to the complexity of the surface conditions, in order to achieve higher-precision ground straw segmentation, it is necessary to extract deeper features during feature extraction. The present invention improves the U-Net semantic segmentation model and introduces the Bottleneck residual block in the deep residual network in the encoder part. The Bottleneck residual block is more efficient than the ordinary convolution layer. It first reduces the number of channels through 1x1 convolution, then performs 3x3 convolution feature extraction, and finally restores the number of channels with 1x1 convolution. For details, see Figure 3As shown, the improved U-Net semantic segmentation model includes: an encoder and a decoder, wherein the encoder starts from the input data and sequentially passes through the first convolution layer (7x7 convolution), the first pooling layer (3x3 pooling), the first Bottleneck residual block, the second Bottleneck residual block, the third Bottleneck residual block, the fourth Bottleneck residual block, the fifth Bottleneck residual block, the sixth Bottleneck residual block, the seventh Bottleneck residual block, the eighth Bottleneck residual block, the ninth Bottleneck residual block, the tenth Bottleneck residual block, the eleventh Bottleneck residual block, the twelfth Bottleneck residual block, the thirteenth Bottleneck residual block, the fourteenth Bottleneck residual block, the fifteenth Bottleneck residual block and the sixteenth Bottleneck residual block; The decoder includes, in order from data input to output: a first upsampling layer (2x2 convolution), a second convolution layer (3x3 convolution), a third convolution layer (3x3 convolution), a second upsampling layer (2x2 convolution), a fourth convolution layer (3x3 convolution), a fifth convolution layer (3x3 convolution), a third upsampling layer (2x2 convolution), a sixth convolution layer (3x3 convolution), a seventh convolution layer (3x3 convolution), a fourth upsampling layer (2x2 convolution), an eighth convolution layer (3x3 convolution), a ninth convolution layer (3x3 convolution) and a fifth upsampling layer (1x1 convolution); The output of the first convolutional layer is jump connected to the output of the fourth upsampling layer; the output of the third Bottleneck residual block is jump connected to the output of the third upsampling layer; the output of the seventh Bottleneck residual block is jump connected to the output of the second upsampling layer; and the output of the thirteenth Bottleneck residual block is jump connected to the output of the first upsampling layer.

[0014] See also Figure 4 As shown in the figure, the first Bottleneck residual block, the fourth Bottleneck residual block, the eighth Bottleneck residual block and the fourteenth Bottleneck residual block are used when the number of input and output channels is different. The structure includes: starting from the input, it is divided into two paths, the first path includes the tenth convolution layer, the eleventh convolution layer and the twelfth convolution layer, the second path includes the thirteenth convolution layer and the RELU activation function layer, the data of the first path passes through the tenth convolution layer, the eleventh convolution layer and the twelfth convolution layer in sequence, the data of the second path passes through the thirteenth convolution layer and the first RELU activation function layer in sequence, and the twelfth convolution layer is output to the first RELU activation function layer.

[0015] See also Figure 5 As shown in the figure, when the number of input and output channels is the same, the structure of the Bottleneck residual block is used to divide the input into two paths. The data of the first path passes through the 14th convolution layer, the 15th convolution layer and the 16th convolution layer in sequence, and the data of the second path includes the RELU activation function layer, and the 16th convolution layer outputs to the second RELU activation function layer. That is, the other Bottleneck residual blocks except the first Bottleneck residual block, the fourth Bottleneck residual block, the eighth Bottleneck residual block and the fourteenth Bottleneck residual block.

[0016] In the decoding stage of the present invention, after each upsampling, the feature map in the decoder is spliced ​​with the feature layer output by the improved encoder. Under this structure, the shallow feature layer retains the detailed information of the image, while the deep feature layer provides high-level semantic information. The improved U-Net semantic segmentation model structure enhances the feature extraction capability, makes the network deeper but still can be efficiently trained and maintains the multi-scale feature fusion capability of the jump connection, thereby improving the precision of image segmentation. The prepared data set is divided into a training set and a validation set according to a ratio of 8:2, wherein the training set is enhanced by rotating, flipping, adding noise, etc. to obtain 1200 images and input into the improved U-Net semantic segmentation model for training. The Total Loss loss function is established by combining the DiceLoss loss function and the BCE Loss loss function to guide the training process of the model. The advantage is that the combination of two different loss functions can comprehensively consider the classification accuracy at the pixel level (through the BCE Loss loss function) and the similarity of the segmentation results (through the Dice Loss loss function). The BCE Loss loss function has good stability in the early stage of training, which helps to accelerate the convergence speed of the model. The Dice Loss function focuses more on the similarity at the pixel level, which can prompt the model to generate smoother and more continuous segmentation results. The F1-Score, mean intersection over union (mIoU), mean pixel accuracy (mPA), and accuracy are used to evaluate the model performance. The calculation method of the Total Loss function is: , The Dice Loss loss function is the Dice similarity coefficient loss function, and the BCE Loss loss function is the binary cross entropy loss function. The calculation formulas are as follows: , in represents the pixel labels of the true segmented image, represents the pixel category of the predicted segmented image, express and The number of intersection elements between and Respectively , The number of elements in in, It is The predicted value of a sample ranges from (0, 1); It is The actual label of each sample, which takes the value of 0 or 1; surface.

[0017] The best trained model has an F1 score of 0.924, an average intersection over union (mIoU) of 85.91%, an average pixel accuracy (mPA) of 92.25%, and an accuracy of 92.97%.

[0018] S3 inputs all collected images into the trained improved U-Net semantic segmentation model for prediction, and obtains the segmented images of straw. Different input images and prediction results are shown in the figure below. Figure 6 As shown in the figure above, the upper figure shows different input images, and the lower figure shows the corresponding prediction results of different input images. The predicted images are mosaicked to obtain the prediction results of the original image taken. Figure 7 As shown, Figure 7 (a) is the original image. Figure 7 (b) is the prediction result. The straw coverage is calculated by counting the proportion of the number of pixels in the straw part to the total number of pixels, and then matched with the ground coordinates to obtain the straw coverage spatial data set. S4 uses multispectral remote sensing images to calculate the remote sensing spectral index of the coordinate points of cultivated land straw coverage, takes the spatial dataset of surface straw coverage as the target variable, and takes the remote sensing spectral index of the coordinate points corresponding to the surface straw coverage and the band information of the multispectral remote sensing image as the feature variables. The target variable and the feature variable are input into the XGBoost model for training to establish an XGBoost model of remote sensing spectral index, band information of multispectral remote sensing images and straw coverage. The band information of the multispectral remote sensing image and the calculated remote sensing spectral index are input into the trained XGBoost model for spatial prediction of cultivated land straw coverage, and plots with cultivated land straw coverage greater than or equal to 30% are judged as protective farming plots.

[0019] In S4: The multispectral remote sensing images are the L2A-level Sentinel-2 remote sensing images closest to the image acquisition date downloaded from the Copernicus Data Space Ecosystem. Since the resolution of each band of Sentinel-2 is not completely consistent and there are some interference factors in the remote sensing images, the remote sensing images are resampled to 10m resolution using SNAP software and interference factors such as clouds and cloud shadows are removed through mask processing; The characteristics of cultivated land straw coverage are manifested as mixed spectral characteristics of soil and residue in optical remote sensing images. The characteristics of cultivated land straw coverage can be extracted by constructing remote sensing indexes through mathematical combinations of various bands or spectral conversions of two or more bands. The present invention selects the following remote sensing spectral indices: normalized difference vegetation index, normalized difference tillage index, normalized difference straw index, shortwave infrared angle index, first normalized difference index, second normalized difference index, normalized difference senescent vegetation index, soil tillage index and dry fuel index. The remote sensing spectral index is calculated using Sentinel-2 multispectral data, and the calculation method of the remote sensing spectral index is: Normalized Difference Vegetation Index: ; Normalized Difference Tillage Index: ; Normalized difference straw index: ; Short-wave infrared angle index: ; First normalized difference index: ; Second normalized difference index: ; Normalized difference senescent vegetation index: ; Soil tillage index: ; Dry Fuel Index: ; , , , and All are collected from Sentinel-2 multispectral data. It indicates a band with a central wavelength of 0.665 μm and a spatial resolution of 10 m; It indicates a band with a central wavelength of 0.842 μm and a spatial resolution of 10 m; It indicates a band with a central wavelength of 0.865 μm and a spatial resolution of 20 m; It indicates a band with a central wavelength of 1.610 μm and a spatial resolution of 20 m; Indicates a band with a central wavelength of 2.190μm and a spatial resolution of 20m.

[0020] The prepared data is input into the XGBoost model, and the five-fold cross-validation method is used for training and evaluation. The evaluation indicators include the determination coefficient R 2 , RMSE, MAE. Figure 8The determination coefficient R is calculated by using the measured straw coverage and the predicted straw coverage. 2 The XGBoost model performance reached a good level. The trained XGBoost model was used to predict the straw coverage of all cultivated land. With 30% as the boundary, the straw coverage of less than 30% was classified as traditional tillage, and the straw coverage of 30% or more was classified as conservation tillage. A confusion matrix was established, and the accuracy and Kappa coefficient were calculated to evaluate the conservation tillage identification results. The accuracy reached 0.96 and the Kappa coefficient was 0.90. The final conservation tillage mapping results are shown in Figure 2. Fig. 9 As shown in (a), traditional tillage and conservation tillage are distinguished by color. The local enlarged image and the corresponding remote sensing image are shown in Fig. 9 As shown in (b), it shows that the present invention can identify the farming method at the field scale.

[0021] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A conservation tillage identification method based on UAV and satellite cross-scale sensing, characterized in that: The following steps are involved: S1 used drones to collect surface images of farmland with different straw coverage conditions at high altitudes before sowing crops; S2 preprocesses the collected images, creates training labels, improves the U-Net semantic segmentation model, and inputs the preprocessed images and training labels to train the improved U-Net semantic segmentation model; S3 Input all collected images into the trained improved U-Net semantic segmentation model for prediction, obtain the straw segmentation image, mosaic the straw segmentation image and count the ratio of the number of pixels of the straw part to the total number of pixels to calculate the straw coverage, match the straw coverage with the ground coordinates to obtain the registered surface straw coverage spatial dataset; S4 Use multispectral remote sensing images to calculate the remote sensing spectral index of the coordinate points of cultivated land straw coverage, take the surface straw coverage spatial dataset as the target variable, and take the remote sensing spectral index of the coordinate points corresponding to the surface straw coverage and the band information of the multispectral remote sensing image as the feature variables. Input the target variable and the feature variable into the XGBoost model for training to establish the XGBoost model of remote sensing spectral index, band information of multispectral remote sensing image and straw coverage. Input the band information of multispectral remote sensing image and the calculated remote sensing spectral index into the XGBoost model for spatial prediction of cultivated land straw coverage, and judge the plots with cultivated land straw coverage greater than or equal to 30% as protective farming plots.

2. The conservation tillage identification method based on UAV and satellite cross-scale sensing according to claim 1 is characterized in that: The improved U-Net semantic segmentation model in S2 comprises: an encoder and a decoder, wherein the encoder sequentially passes through a first convolutional layer, a first pooling layer, a first Bottleneck residual block, a second Bottleneck residual block, a third Bottleneck residual block, a fourth Bottleneck residual block, a fifth Bottleneck residual block, a sixth Bottleneck residual block, a seventh Bottleneck residual block, an eighth Bottleneck residual block, a ninth Bottleneck residual block, a tenth Bottleneck residual block, an eleventh Bottleneck residual block, a twelfth Bottleneck residual block, a thirteenth Bottleneck residual block, a fourteenth Bottleneck residual block, a fifteenth Bottleneck residual block and a sixteenth Bottleneck residual block from the input data; The decoder includes, in order from data input to output: a first upsampling layer, a second convolutional layer, a third convolutional layer, a second upsampling layer, a fourth convolutional layer, a fifth convolutional layer, a third upsampling layer, a sixth convolutional layer, a seventh convolutional layer, a fourth upsampling layer, an eighth convolutional layer, a ninth convolutional layer and a fifth upsampling layer; The output of the first convolutional layer is jump connected to the output of the fourth upsampling layer; the output of the third Bottleneck residual block is jump connected to the output of the third upsampling layer; the output of the seventh Bottleneck residual block is jump connected to the output of the second upsampling layer; and the output of the thirteenth Bottleneck residual block is jump connected to the output of the first upsampling layer.

3. The conservation tillage identification method based on UAV and satellite cross-scale sensing according to claim 2 is characterized in that: The structures of the first Bottleneck residual block, the fourth Bottleneck residual block, the eighth Bottleneck residual block and the fourteenth Bottleneck residual block include: starting from the input, they are divided into two paths, the first path includes the tenth convolution layer, the eleventh convolution layer and the twelfth convolution layer, the second path includes the thirteenth convolution layer and the RELU activation function layer, the data of the first path passes through the tenth convolution layer, the eleventh convolution layer and the twelfth convolution layer in sequence, the data of the second path passes through the thirteenth convolution layer and the first RELU activation function layer in sequence, and the twelfth convolution layer outputs to the first RELU activation function layer.

4. The conservation tillage identification method based on UAV and satellite cross-scale sensing according to claim 3 is characterized in that: The structure of the Bottleneck residual block other than the first Bottleneck residual block, the fourth Bottleneck residual block, the eighth Bottleneck residual block and the fourteenth Bottleneck residual block includes: starting from the input, it is divided into two paths, the data of the first path passes through the fourteenth convolution layer, the fifteenth convolution layer and the sixteenth convolution layer in sequence, the data of the second path includes a RELU activation function layer, and the sixteenth convolution layer outputs to the second RELU activation function layer.

5. The conservation tillage identification method based on UAV and satellite cross-scale sensing according to claim 1 is characterized in that: The remote sensing spectral index includes: normalized difference vegetation index, normalized difference tillage index, normalized difference straw index, shortwave infrared angle index, first normalized difference index, second normalized difference index, normalized difference senescent vegetation index, soil tillage index and dry fuel index.

6. The conservation tillage identification method based on UAV and satellite cross-scale sensing according to claim 5 is characterized in that: Normalized Difference Vegetation Index: ; Normalized Difference Tillage Index: ; Normalized difference straw index: ; Shortwave infrared angle index: ; First Normalized Difference Index: ; Second Normalized Difference Index: ; Normalized difference senescent vegetation index: ; Soil tillage index: ; Dry Fuel Index: ; in, , , , and All are collected from Sentinel-2 multispectral data. It indicates a band with a central wavelength of 0.665 μm and a spatial resolution of 10 m; It indicates a band with a central wavelength of 0.842 μm and a spatial resolution of 10 m; It indicates a band with a central wavelength of 0.865 μm and a spatial resolution of 20 m; It indicates a band with a central wavelength of 1.610 μm and a spatial resolution of 20 m; Indicates a band with a central wavelength of 2.190μm and a spatial resolution of 20m.

Images