A soil moisture remote sensing detection method based on deep learning

Abstract

The present invention relates to the field of soil detection technology, and specifically to a soil moisture remote sensing detection method based on deep learning; the method comprises: obtaining multiple data sets of a target area; wherein each data set comprises any remote sensing feature in remote sensing feature data obtained in advance and a label corresponding to the remote sensing feature; the label comprises a real water content label and a soft label; multiple prediction models constructed are trained using multiple data sets to obtain each trained prediction model; during training, one data set corresponds to one prediction model; remote sensing features in remote sensing feature data of a to-be-detected area obtained in real time are respectively input into the corresponding trained prediction model, and the corresponding soil moisture content is output; all soil moisture contents are weighted and summed to obtain the weighted soil moisture content. That is, the scheme of the present invention can accurately monitor soil moisture.

Description

Technical Field

[0001] The present invention relates to the field of soil detection technology. More specifically, the present invention relates to a soil moisture remote sensing detection method based on deep learning. Background Art

[0002] Water conservancy is the lifeline of agriculture, and a good harvest of grain cannot be achieved without irrigation. Changes in soil moisture (water content) affect evaporation, infiltration, runoff, erosion and other processes at the land-atmosphere interface. It is a key parameter in the fields of agriculture, hydrology, meteorology, ecology, and is an important basis for determining whether to irrigate an irrigation area and the amount of irrigation water.

[0003] Traditional methods of soil moisture monitoring mainly include contact site monitoring and non-contact remote sensing observation data inversion.

[0004] Among them, the site monitoring methods include drying method, negative pressure meter method, soil moisture meter method, seed moisture meter detection method, etc., which are not affected by the atmosphere and vegetation and have high data accuracy. However, the number of sites is limited and the sampling frequency is low, and it is impossible to form a continuous and large-scale spatial data set. At the same time, since soil moisture is affected by multiple factors such as soil physical properties, terrain, vegetation type, and climatic conditions, there is a large spatial non-uniformity.

[0005] Non-contact remote sensing technology can quickly obtain spectral information of large areas of ground objects in time and space, and can not only monitor the balance of surface water revenue and expenditure at a macro level, but also reflect the changes in ground object spectrum and surface evapotranspiration caused by water surplus and deficit at a micro level. At present, various methods and models have been proposed in the field of satellite remote sensing soil moisture monitoring, mainly including the apparent thermal inertia method and the vegetation index method.

[0006] Among them, thermal inertia is a thermal property of soil. Thermal inertia plays a decisive role in the temperature change of ground objects. The value of soil thermal inertia is closely related to soil moisture content. Soil moisture can be inverted by estimating soil thermal inertia, but it is difficult to directly obtain parameters using remote sensing methods. In practical applications, the apparent thermal inertia method is mainly used to replace thermal inertia. Apparent thermal inertia only involves two parameters: surface temperature difference and surface reflectivity. Since this method relies on soil thermal characteristics to estimate soil moisture, when the vegetation coverage is high, the accuracy will be affected by mixed pixels and reduced. Therefore, this method is mainly suitable for bare soil and areas with low vegetation coverage.

[0007] The vegetation index method is a method that uses remote sensing technology to evaluate vegetation coverage and its growth vitality. It is based on the spectral characteristics of vegetation and forms various vegetation indices by combining satellite visible light and near-infrared bands. Using vegetation indices to distinguish vegetation types and vegetation coverage and compare growth conditions in different periods is an important way to monitor drought conditions by remote sensing. This is because vegetation growth is closely related to soil moisture. If the water supply is sufficient, vegetation grows well; otherwise, growth deteriorates. Therefore, some scholars have constructed a variety of vegetation indices, including the normalized difference vegetation index (NDVI), the simple vegetation index (SVI), the ratio vegetation index (RVI), etc., based on the spectral characteristics of vegetation with strong absorption in the red light band and strong reflection in the near-infrared band, through the band combination method.

[0008] In the related technologies, such as the patent application document with publication number CN112287294A and titled "A spatiotemporal bidirectional soil moisture interpolation method based on deep learning", it fully considers the spatiotemporal three-dimensional characteristics of the data (meteorological elements, vegetation parameters, soil information, topographic conditions and remote sensing soil moisture content, etc.), integrates the seasonal variation law of soil water in time and its geographical characteristics in space, and uses the deep learning model to effectively establish the intrinsic connection between meteorology, influencing factors and soil moisture content, thereby improving the interpolation accuracy of soil moisture content.

[0009] In the above scheme, the label during network model training is the water content of each location in the collected remote sensing image. However, since the collected remote sensing images cover a wide range of geographical locations, and the acquisition equipment can only collect the water content of some locations in the remote sensing images, it cannot correspond one-to-one with the pixels of all locations in the remote sensing images. Therefore, some locations do not have real water content, which makes the accuracy of the trained network model not high, thereby affecting the accuracy of soil moisture monitoring. Summary of the invention

[0010] The purpose of the present invention is to propose a soil moisture remote sensing detection method based on deep learning to solve the problem that the prior art cannot accurately monitor soil moisture; to this end, the present invention provides a solution in the following aspect.

[0011] The present invention provides a soil moisture remote sensing detection method based on deep learning, comprising:

[0012] Acquire multiple data sets of the target area; wherein each data set includes any remote sensing feature in the pre-acquired remote sensing feature data and a label corresponding to the remote sensing feature;

[0013] The label includes a real water content label and a soft label; the real water content label is determined by the water content collected by the collection device, and the soft label is: a condition is set that the remote sensing feature of the location without real water content is the same as the remote sensing feature corresponding to the real water content label. If the condition is met, the probability density of each real water content corresponding to the remote sensing feature is used as the soft label of the remote sensing feature of the location; if the condition is not met, a weighted probability value is used as the soft label of the location without real water content, and the weighted probability value is positively correlated with the probability density of the first L real water contents of the nearest neighbors of the location without real water content, where L≥2;

[0014] Using multiple data sets to train multiple prediction models, each trained prediction model is obtained; during training, one data set corresponds to only one prediction model;

[0015] Input the remote sensing features in the remote sensing feature data of the area to be detected acquired in real time into the corresponding trained prediction model, and output the corresponding soil moisture content;

[0016] The weighted soil moisture content is obtained by taking the weighted sum of all soil moisture contents.

[0017] In the above scheme, by obtaining soft labels of different remote sensing features, a data set containing different remote sensing features and their corresponding labels is obtained, and used to train the network model, and then multiple network models are obtained. By obtaining soft labels, the label data in the data set can be enriched, the accuracy of model training can be improved, and the accuracy of the prediction results of each prediction model can be improved; at the same time, multiple prediction results of multiple remote sensing features are integrated to further improve the accuracy of soil moisture monitoring.

[0018] Optionally, the weighted probability value is:

[0019] ;

[0020] in, The value of the kth remote sensing feature at position a is The weighted probability value of water content at time and They represent the values ​​of the kth remote sensing features of the two nearest neighbors of position a, b and c. The value of the kth remote sensing feature at position b is The probability density function of water content is The value of the kth remote sensing feature at position c is The probability density function of the water content at .

[0021] In the above scheme, when calculating the weighted probability value of the position without true water content, the weight is set to the difference between the remote sensing features, which can accurately determine the accuracy of the true water content of the two nearest neighbor positions.

[0022] Optionally, the remote sensing features include normalized difference vegetation index, normalized vegetation wetness index, ratio vegetation index, normalized difference infrared index, water stress index, water band index, fused vegetation index, local incidence angle and polarized soil backscatter coefficient.

[0023] The above scheme integrates multi-dimensional remote sensing features, which can improve the accuracy of subsequent soil moisture monitoring.

[0024] Optionally, the loss function of each prediction model is:

[0025] ;

[0026] in, represents the loss function of the prediction model corresponding to the k-th remote sensing feature, represents the number of true water content labels in the kth remote sensing feature, represents the number of soft labels in the k-th remote sensing feature, represents the label value of the i-th true water content label in the k-th remote sensing feature, represents the predicted value corresponding to the i-th true water content label in the k-th remote sensing feature, represents the sth label value of the jth soft label in the kth remote sensing feature, represents the predicted value corresponding to the jth soft label in the kth remote sensing feature, It represents the probability density of the occurrence of the sth label value of the jth soft label, and S is the total number of label values ​​corresponding to the soft label.

[0027] The above scheme can improve the accuracy of the prediction model by calculating the loss.

[0028] Optionally, the multiple prediction models all adopt a convolutional neural network model.

[0029] Optionally, the weighted sum of all soil moisture contents is performed to obtain the weighted soil moisture content as follows:

[0030] Calculate the model evaluation index of each prediction model;

[0031] The weighted soil moisture content was obtained by performing weighted summation on the soil moisture content according to the evaluation indexes of each model.

[0032] Optionally, the weighted soil moisture content is:

[0033] ;

[0034] in, represents the weighted soil moisture content, represents the soil moisture content of the prediction model corresponding to the kth remote sensing feature, represents the mean square error of the prediction model corresponding to the kth remote sensing feature, It represents the sum of the mean square errors of all prediction models, and K is the total number of remote sensing features.

[0035] In the above scheme, by introducing the model evaluation index of different trained prediction models to adjust the soil moisture content of the corresponding prediction model, a more accurate and reasonable final soil moisture content can be obtained.

[0036] Optionally, the pre-acquired remote sensing feature data is obtained by preprocessing the collected remote sensing images.

[0037] Optionally, the probability density of the true water content is:

[0038] ;

[0039] in, represents the probability that the true water content is x when the value of the f-th remote sensing feature is y, It indicates the number of times the actual water content is x when the value of the fth remote sensing feature is y. It indicates the number of all true water contents corresponding to the value of the f-th remote sensing feature being y.

[0040] The beneficial effects of the present invention are:

[0041] The solution of the present invention collects remote sensing image data of the target area and obtains remote sensing features of different dimensions, then determines corresponding labels for the remote sensing features of different dimensions, and then trains corresponding network models, thereby obtaining multiple trained network models for predicting the moisture content under remote sensing features of different dimensions, and integrates the prediction results from multiple angles, and finally integrates the soil moisture content of the surface, thereby improving the accuracy of the prediction results. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The above and other objects, features and advantages of the exemplary embodiments of the present invention will become readily understood by reading the detailed description below with reference to the accompanying drawings. In the accompanying drawings, several embodiments of the present invention are shown in an exemplary and non-limiting manner, and the same or corresponding reference numerals represent the same or corresponding parts, wherein:

[0043] Figure 1 A flowchart of a soil moisture remote sensing detection method based on deep learning in this embodiment is schematically shown;

[0044] Figure 2 The structural flow chart of a soil moisture remote sensing detection method based on deep learning in this embodiment is schematically shown. DETAILED DESCRIPTION

[0045] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0046] Taking a certain target area as an example, a soil moisture remote sensing detection method based on deep learning of the present invention is introduced.

[0047] Specifically, Figure 1 and Figure 2 As shown, a soil moisture remote sensing detection method based on deep learning in this embodiment includes the following steps:

[0048] Step S1, obtaining multiple data sets of the target area. Specifically, each data set in this embodiment includes any remote sensing feature in the remote sensing feature data and a label corresponding to the remote sensing feature.

[0049] Specifically, the remote sensing feature data is obtained based on the remote sensing image data of the target area, wherein the remote sensing image data includes Sentinel-1 SAR images and Sentinel-2 optical images.

[0050] In this embodiment, the Sentinel-1 SAR image and the Sentinel-2 optical image are preprocessed to obtain remote sensing feature data.

[0051] Remote sensing characteristic data include optical remote sensing parameters and microwave remote sensing parameters. The microwave remote sensing parameters are obtained by preprocessing the Sentinel-1 SAR images through image filtering, radiation calibration, geocoding, decibel conversion, etc. The optical remote sensing parameters are obtained by preprocessing the Sentinel-2 optical images through radiation calibration, atmospheric correction, resampling, etc., thereby obtaining the remote sensing characteristics of the optical remote sensing images.

[0052] Among them, optical remote sensing parameters include Normalized Difference Vegetation Index (NDVI), Normalized Difference Moisture Index (NDMI), Ratio Vegetation Index (RVI), Normalized Difference Infrared Index (NDII), Water Stress Index (WSI), Water Band Index (WBI), Fusion Vegetation Index (FVI); microwave remote sensing parameters include local incidence angle , remote sensing characteristics of polarization soil backscatter coefficient, etc.

[0053] For example, taking the normalized difference vegetation index as an example, the normalized difference vegetation index is based on the characteristics that the chlorophyll in the vegetation has a low reflectivity in the red band of the visible light band and a relatively high reflectivity in the green band, and the reflectivity gradually decreases in the range from the near-infrared band to the mid-infrared band. The formula is as follows:

[0054] NDVI = (NIR-RED) / (NIR+RED); where NIR is the reflectivity of the near-infrared band and RED is the reflectivity of the red band.

[0055] Among them, since the acquisition of remote sensing optical parameters and microwave remote sensing parameters are existing technologies, they will not be introduced in detail here.

[0056] The labels in this embodiment include real moisture content labels and soft labels, wherein the real moisture content labels use professional equipment (such as soil moisture monitors) to collect the real moisture content of the soil at a pre-calibrated position, wherein the real moisture content refers to the proportion of water in the soil, usually expressed as a percentage or volumetric moisture content (m³ / m³), illustratively, the real moisture content can be 20%, 40%, 50%, 60%, etc.

[0057] The soft label is a label assigned to a location without real water content, which is used to determine the probability of the location belonging to different real water contents. In this embodiment, the soft label is set because the geographical location of the remote sensing images collected in practical applications is large and extensive, and the water content collected by professional equipment cannot include every pixel in the remote sensing image. Therefore, it is necessary to give an approximate water content to the location without real water content in the remote sensing image, so that the subsequent training model is more accurate.

[0058] In addition, the soft labels in this embodiment correspond to the remote sensing features, that is, each remote sensing feature corresponds to a soft label, and different remote sensing features have different soft labels.

[0059] Specifically, the process of obtaining the soft label is as follows:

[0060] First, determine the number of different true water contents under the same remote sensing feature and the probability density of each true water content, specifically: ;in, represents the probability that the true water content is x when the value of the f-th remote sensing feature is y, It indicates the number of times the actual water content is x when the value of the fth remote sensing feature is y. It indicates the number of all true water contents corresponding to the value of the f-th remote sensing feature being y.

[0061] Exemplarily, when the remote sensing feature is the normalized difference vegetation index, and the normalized vegetation index is y, the corresponding true water content has four cases, namely A, B, C, and D. If the true water content is A, B, C, and D and the number of occurrences is 5, 2, 2, and 1 respectively, then the number of all true water contents is 10. At this time, when the normalized vegetation index is y, the probability density of the true water content is A, B, C, and D is 0.5, 0.2, 0.2, and 0.1 respectively. Then, there are 4 soft labels for positions without true water content, and the probability density of the first label value A of the soft label is 0.5, the probability density of the second label value B of the soft label is 0.2, the probability density of the third label value C of the soft label is 0.2, and the probability density of the fourth label value D of the soft label is 0.1.

[0062] The "first", "second", "third" and "fourth" in the above description are only used to distinguish the soft tags, and are not limited to the above order.

[0063] Secondly, according to the set conditions, the soft labels are assigned to the locations without real water content, where the condition is that the remote sensing features of the locations without real water content are the same size as the remote sensing features corresponding to the real water content labels; the details are as follows:

[0064] (1) When the above conditions are met, the probability density of each true water content corresponding to the remote sensing feature is used as the soft label of the remote sensing feature at that location. There may be multiple or one soft label for a location without true water content.

[0065] (2) When the above conditions are not met, calculate the weighted probability value of the location without true moisture content.

[0066] The weighted probability value is positively correlated with the probability density of the two nearest real water contents of the position without real water content. The two nearest real water contents are determined according to the position. Specifically, the distance between the position of the real water content and the current position without real water content can be counted, and then the distances are sorted from small to large, and then the real water contents of the first L nearest neighbors are selected, where L is greater than or equal to 2.

[0067] In one embodiment, when L=2, the weighted probability value is:

[0068] ;

[0069] in, The value of the kth remote sensing feature at position a is The weighted probability value of water content at time and They represent the values ​​of the kth remote sensing features of the two nearest neighbors of position a, b and c. The value of the kth remote sensing feature at position b is The probability density function of water content is The value of the kth remote sensing feature at position c is The probability density function of the water content at .

[0070] When L=3, when calculating the weighted probability value, the weight of each water content can also be the proportion of the numerical values ​​of each remote sensing feature.

[0071] In another embodiment, the weighted probability value can also be directly assigned weights to the real water contents of the two nearest neighbors according to the size of the distance (the smaller the distance, the greater the weight assigned, and vice versa), and the probability density of the two real water contents is weighted using the assigned weights. Exemplarily, when the distance between the positions of the two real water contents and the position without real water content is the same, the two weights are 0.5 and 0.5 respectively. Among them, when L=3, the weight can also be set according to the size of the distance.

[0072] It should be noted that, in this embodiment, the soft label of the water content at the location without the real water content is determined according to the real water content probability density corresponding to the remote sensing feature at the location without the real water content. Then, the soft label of each location without the real water content in other remote sensing features can be calculated. Thus, the soft labels of the locations without the real water content labels of all remote sensing features can be determined.

[0073] The soft labels are obtained by calculating the relationship between the remote sensing image features and the water content at the locations with real water content labels, thereby determining the soft labels at the locations without real water content, and then constructing a data set of different remote sensing features and their corresponding labels.

[0074] In this embodiment, by subdividing the soft label corresponding to each remote sensing feature, different remote sensing features and their corresponding labels can be constructed, data sets of different dimensions are obtained, and the accuracy of predicting water content by each remote sensing feature is improved.

[0075] Step S2, using multiple data sets to train the constructed multiple prediction models to obtain each trained prediction model.

[0076] The multiple prediction models constructed in this embodiment can use the same network model, such as all using convolutional neural network models, such as CNN network models; of course, different network models can also be used, such as some using convolutional neural network models and some using support vector machines or other network models.

[0077] It should be noted that when multiple prediction models all use convolutional neural network models, during the training process, one data set is only used to train one prediction model, and multiple data sets are used to train multiple prediction models respectively, thereby obtaining multiple trained prediction models. However, due to the different data sets, the trained prediction models are also different.

[0078] Among them, the loss function of each prediction model in this embodiment is:

[0079] ;

[0080] in, represents the loss function of the prediction model corresponding to the k-th remote sensing feature, represents the number of true water content labels in the kth remote sensing feature, represents the number of soft labels in the k-th remote sensing feature, represents the label value of the i-th true water content label in the k-th remote sensing feature, represents the predicted value corresponding to the i-th true water content label in the k-th remote sensing feature, represents the sth label value of the jth soft label in the kth remote sensing feature, represents the predicted value corresponding to the jth soft label in the kth remote sensing feature, It represents the probability density of the occurrence of the sth label value of the jth soft label, and S is the total number of label values ​​corresponding to the soft label.

[0081] The sth label value in the above is the size of the corresponding true water content. Since a soft label without a corresponding position of true water content has multiple probability densities, each probability density corresponds to a true water content, therefore, there may be multiple label values ​​for the soft label without a corresponding position of true water content.

[0082] For example, when the data set is the normalized difference vegetation index and its corresponding labels, and the network model is a CNN network model, the specific training process is as follows:

[0083] The normalized difference vegetation index and its corresponding label are input into the CNN network model. During the training process, the loss between the output prediction value and the label is calculated.

[0084] Use gradient descent to adjust neural network model parameters to minimize prediction error;

[0085] Iteratively adjust the parameters of the CNN network model until the loss is less than a certain value or the set number of training times is reached, and finally a trained neural network model is obtained.

[0086] Step S3, inputting the remote sensing features in the remote sensing feature data of the area to be detected acquired in real time into the corresponding trained prediction model, and outputting the corresponding soil moisture content.

[0087] In this embodiment, remote sensing image data of the area to be detected is collected and preprocessed to obtain remote sensing feature data, and remote sensing features in the remote sensing feature data are respectively input into the corresponding trained prediction model to output the soil moisture content corresponding to the remote sensing features. The preprocessing method is the same as the preprocessing method in step S1, and will not be repeated here.

[0088] Among them, each remote sensing feature corresponds to a soil moisture content.

[0089] Step S4: performing weighted summation on all soil moisture contents to obtain weighted soil moisture contents.

[0090] In one embodiment, all soil moisture contents are weighted and summed to obtain the weighted soil moisture content. The process of obtaining the weighted soil moisture content is as follows: calculating the model evaluation index of each prediction model; and weightedly summing each soil moisture content according to each model evaluation index to obtain the weighted soil moisture content.

[0091] The model evaluation indicators mentioned above may be accuracy, precision, F1 score, mean square error, mean absolute error, cross-validation score, etc. Since the calculation of model evaluation indicators is an existing technology, it will not be described in detail here.

[0092] Specifically, when the model evaluation index is the mean square error, the weighted soil moisture content is: ;in, represents the weighted soil moisture content, represents the soil moisture content of the prediction model corresponding to the kth remote sensing feature, represents the mean square error of the prediction model corresponding to the kth remote sensing feature, It represents the sum of the mean square errors of all prediction models, and K is the total number of remote sensing features.

[0093] It should be noted that the total number of remote sensing features is equal to the total number of prediction models.

[0094] In another embodiment, the weight of each soil moisture content when performing the weighted summation of all soil moisture contents in the above-mentioned method can also be determined according to the importance of the remote sensing features corresponding to each soil moisture content. For example, the weight assigned to the corresponding soil moisture content of a more important remote sensing feature is larger, and vice versa.

[0095] In the above embodiment, by performing model evaluation on different trained prediction models to determine the accuracy of different prediction models, and using the different accuracy conditions as weights to adjust the soil moisture content output by the corresponding prediction model, a more accurate and reasonable final soil moisture content can be obtained.

[0096] The solution of the present invention not only takes into account the remote sensing features of microwave remote sensing, but also introduces the remote sensing features of optical images through the constructed deep learning model, so as to obtain soft labels of remote sensing features in multiple dimensions, and then train the corresponding deep learning model, thereby improving the accuracy of the trained model. When determining the soil moisture content of the surface, the predicted results of the moisture content corresponding to the remote sensing features in multiple dimensions are combined, thereby eliminating the influence of vegetation and realizing accurate monitoring of soil moisture conditions.

[0097] In the description of this specification, “plurality” means at least two, such as two, three or more, etc., unless otherwise clearly and specifically defined.

[0098] Although this specification has shown and described a number of embodiments of the present invention, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Those skilled in the art will conceive of many modifications, changes and alternatives without departing from the ideas and spirit of the present invention. It should be understood that in the practice of the present invention, various alternatives to the embodiments of the present invention described herein may be employed.

Claims

1. A soil moisture remote sensing detection method based on deep learning, characterized in that: include: Acquire multiple data sets of the target area; wherein each data set includes any remote sensing feature in the pre-acquired remote sensing feature data and a label corresponding to the remote sensing feature; The label includes a real water content label and a soft label; the real water content label is determined by the water content collected by the collection device, and the soft label is: a condition is set that the remote sensing feature of the location without real water content is the same as the remote sensing feature corresponding to the real water content label. If the condition is met, the probability density of each real water content corresponding to the remote sensing feature is used as the soft label of the remote sensing feature of the location; if the condition is not met, a weighted probability value is used as the soft label of the location without real water content, and the weighted probability value is positively correlated with the probability density of the first L real water contents of the nearest neighbors of the location without real water content, where L≥2; Using multiple data sets to train multiple prediction models, each trained prediction model is obtained; during training, one data set corresponds to only one prediction model; Input the remote sensing features in the remote sensing feature data of the area to be detected acquired in real time into the corresponding trained prediction model, and output the corresponding soil moisture content; Perform weighted summation on all soil moisture contents to obtain weighted soil moisture content; When L=2, the weighted probability value is: ； in, The value of the kth remote sensing feature at position a is The weighted probability value of water content at time and They represent the values ​​of the kth remote sensing features of the two nearest neighbors of position a, b and c. The value of the kth remote sensing feature at position b is The probability density function of water content is The value of the kth remote sensing feature at position c is The probability density function of the water content at .

2. The soil moisture remote sensing detection method based on deep learning according to claim 1, characterized in that: The remote sensing features include normalized difference vegetation index, normalized vegetation wetness index, ratio vegetation index, normalized difference infrared index, water stress index, water band index, fused vegetation index, local incidence angle and polarized soil backscatter coefficient.

3. The soil moisture remote sensing detection method based on deep learning according to claim 1, characterized in that: The loss function of each prediction model is: ； in, represents the loss function of the prediction model corresponding to the k-th remote sensing feature, represents the number of true water content labels in the kth remote sensing feature, represents the number of soft labels in the k-th remote sensing feature, represents the label value of the i-th true water content label in the k-th remote sensing feature, represents the predicted value corresponding to the i-th true water content label in the k-th remote sensing feature, represents the sth label value of the jth soft label in the kth remote sensing feature, represents the predicted value corresponding to the jth soft label in the kth remote sensing feature, It represents the probability density of the occurrence of the sth label value of the jth soft label, and S is the total number of label values ​​corresponding to the soft label.

4. The soil moisture remote sensing detection method based on deep learning according to claim 3 is characterized in that: The multiple prediction models all adopt convolutional neural network models.

5. The soil moisture remote sensing detection method based on deep learning according to claim 1, characterized in that: The process of obtaining the weighted soil moisture content by weighted summing of all soil moisture contents is as follows: Calculate the model evaluation index of each prediction model; The weighted soil moisture content was obtained by performing weighted summation on the soil moisture content according to the evaluation indexes of each model.

6. The soil moisture remote sensing detection method based on deep learning according to claim 5, characterized in that: The weighted soil moisture content is: ； in, represents the weighted soil moisture content, represents the soil moisture content of the prediction model corresponding to the kth remote sensing feature, represents the mean square error of the prediction model corresponding to the kth remote sensing feature, It represents the sum of the mean square errors of all prediction models, and K is the total number of remote sensing features.

7. The soil moisture remote sensing detection method based on deep learning according to claim 1, characterized in that: The pre-acquired remote sensing feature data is obtained by pre-processing the collected remote sensing images.

8. The soil moisture remote sensing detection method based on deep learning according to claim 6, characterized in that: The probability density of the true water content is: ； in, represents the probability that the true water content is x when the value of the f-th remote sensing feature is y, It indicates the number of times the actual water content is x when the value of the fth remote sensing feature is y. It indicates the number of all true water contents corresponding to the value of the f-th remote sensing feature being y.

Images