Soil salinity inversion method, device and electronic equipment

Abstract

The present application provides a soil salinity inversion method, device and electronic equipment, which relate to the field of soil data processing. The soil salinity inversion method includes: extracting remote sensing image features, temporal geoscience features and spatial geoscience features from multi-source remote sensing data; calculating the soil salinity band sensitivity weight and soil salinity time sensitivity weight of multi-source remote sensing data based on remote sensing image features and temporal geoscience features; using the soil salinity band sensitivity weight and soil salinity time sensitivity weight for remote sensing image features and temporal geoscience features to obtain a one-dimensional attention map; obtaining a two-dimensional attention map based on the one-dimensional attention map and spatial geoscience features; and obtaining soil salinity inversion results based on the two-dimensional attention map. The soil salinity inversion method provided in the embodiment of the present application effectively combines multi-source remote sensing image data and natural environmental elements, fully considers the spatiotemporal dynamic change characteristics of soil salinity, and significantly improves the accuracy and reliability of soil salinity inversion.

Description

Technical Field

[0001] The present application relates to a field, and in particular, to a soil salinity inversion method, device and electronic equipment. Background Art

[0002] Saline-alkali land, also known as saline soil, refers to a type of land where excessive salt and alkali components accumulate on the soil surface, including salt land, alkaline land, salinized land and alkaline land. Soil salinization refers to the phenomenon and process in which the salt content in the soil surface continues to increase and exceeds a certain limit. Soil salinization is a serious land degradation problem and an important manifestation of soil environmental disasters. The causes of soil salinization include natural and human factors such as seawater erosion, high groundwater levels and unreasonable human reclamation. Accurate monitoring of the salt content of cultivated soil is a prerequisite for the scientific development and utilization of saline-alkali land, increasing crop yields and ensuring national food security.

[0003] Traditional soil salinity surveys are mainly based on field measurements, which is labor-intensive and has limited accuracy in traditional interpolation results. Soil salinity inversion methods based on machine learning and deep learning have bottlenecks in model optimization and room for improvement in inversion accuracy. Summary of the invention

[0004] The purpose of the embodiments of the present application is to provide a soil salinity inversion method, device and electronic device, which integrates multi-source remote sensing image data, natural environmental factors, etc., to construct a rich soil salinity inversion data set, and uses multi-band and multi-time series remote sensing data to establish a deep linear neural network through deep learning technology. In addition, an attention mechanism is introduced to couple natural environmental factors into the soil salinity inversion model, and to screen and emphasize key time and environmental factors that have a greater impact on soil salinity.

[0005] In a first aspect, an embodiment of the present application provides a soil salinity inversion method, the method comprising: extracting remote sensing image features, temporal geological features and spatial geological features from multi-source remote sensing data; calculating the soil salinity band sensitivity weights and the soil salinity time sensitivity weights of the multi-source remote sensing data based on the remote sensing image features and the temporal geological features; applying the soil salinity band sensitivity weights and the soil salinity time sensitivity weights to the remote sensing image features and the temporal geological features to obtain a one-dimensional attention map; obtaining a two-dimensional attention map based on the one-dimensional attention map and the spatial geological features; and obtaining a soil salinity inversion result based on the two-dimensional attention map.

[0006] In the above implementation process, the soil salinity inversion method provided by this application integrates multi-source remote sensing data and natural environmental elements, and uses deep learning technology and attention mechanism to greatly improve the accuracy and stability of soil salinity inversion. This soil salinity inversion method can dynamically monitor the seasonal changes and long-term trends of soil salinity, provide a soil salinity distribution map with high spatial resolution, and is suitable for soil salinity monitoring in different regions and environmental conditions. It enhances the reliability and applicability of the model and provides strong technical support for the scientific development and utilization of saline-alkali land, increasing crop yields, and ensuring national food security.

[0007] Optionally, in an embodiment of the present application, the multi-source remote sensing image data includes: multi-source remote sensing image data; extracting remote sensing image features, temporal geological features and spatial geological features from the multi-source remote sensing data, including: extracting remote sensing image features including spectral features and texture features from the multi-source remote sensing image data.

[0008] Optionally, in an embodiment of the present application, the multi-source remote sensing image data includes: geological time series data and geological spatial data; remote sensing image features, time series geological features and spatial geological features are extracted from the multi-source remote sensing data, including: extracting time series features including precipitation and temperature from the geological time series data; extracting spatial geological features including soil type, crop planting information and terrain features from the geological spatial data.

[0009] In the above implementation process, the application can extract remote sensing image features including spectral features and texture features from multi-source remote sensing image data, and combine geoscientific time series features and spatial geoscientific features to form a rich soil salinity inversion data set. It effectively integrates multi-source, multi-time series, and multi-scale data, improves the accuracy and stability of the soil salinity inversion model, can dynamically monitor changes in soil salinity, and provide a soil salinity distribution map with high spatial resolution.

[0010] Optionally, in an embodiment of the present application, based on remote sensing image features and time series geological features, soil salinity band sensitivity weights and soil salinity time sensitivity weights of multi-source remote sensing data are calculated, including: performing average pooling and global maximum pooling on the remote sensing image features and time series features, and performing feature fusion of the average pooling results and the global maximum pooling results to obtain pooling fusion features; inputting the pooling fusion features into the fully connected layer and activation function to obtain soil salinity band sensitivity weights and soil salinity time sensitivity weights.

[0011] In the above implementation process, through a variety of feature extraction and pooling methods, the important information about soil salinity in remote sensing data is effectively captured, and the expressiveness and robustness of the features are improved. And by using the fully connected layer and activation function, the accurate modeling of the band and time sensitivity of soil salinity is achieved, so that the model can effectively adjust the degree of attention to different data sources according to the weight.

[0012] Optionally, in an embodiment of the present application, the soil salt band sensitivity weight and the soil salt time sensitivity weight are used for remote sensing image features and time series geological features to obtain a one-dimensional attention map, including: obtaining soil salt band sensitive features and soil salt time sensitive features based on the soil salt band sensitivity weight and the soil salt time sensitive weight, respectively; fusing the soil salt band sensitive features and the soil salt time sensitive features to obtain a soil salt sensitive fusion feature; inputting the soil salt sensitive fusion feature into the fully connected layer and the activation function to obtain a one-dimensional attention map.

[0013] In the above implementation process, the soil salinity inversion method provided in the embodiment of the present application, through the application of deep learning technology and attention mechanism, enables the inversion model to utilize multi-source remote sensing data more comprehensively and accurately, solves the problem of rich data but low information utilization in traditional methods, and enhances the model's perception of spatial and temporal changes in soil salinity.

[0014] Optionally, in an embodiment of the present application, a two-dimensional attention map is obtained based on the one-dimensional attention map and the spatial geological features, including: fusing the one-dimensional attention map and the spatial geological features to obtain spatiotemporal fusion features; and using the spatial attention mechanism on the spatiotemporal fusion features to obtain a two-dimensional attention map.

[0015] In the above implementation process, the soil salinity inversion method provided by the embodiment of the present application effectively combines time series information and spatial distribution characteristics by fusing one-dimensional attention maps and spatial geological features, thereby improving the model's perception of soil salinity changes. And further introducing the spatial attention mechanism, the model can more accurately distinguish the soil salinity changes in different regions and time points, thereby improving the spatial resolution and accuracy of the inversion results.

[0016] Optionally, in an embodiment of the present application, a soil salinity inversion result is obtained based on a two-dimensional attention map, including: decoding a one-dimensional attention map and a two-dimensional attention map, and at each decoding layer, using a jump connection to fuse the corresponding encoding layer feature map with the current decoding layer feature map; at the last layer of the decoding network, convolving the result of fusing the corresponding encoding layer feature map with the current decoding layer feature map to obtain a soil salinity inversion result.

[0017] In the above implementation process, the soil salinity inversion method provided by the present application effectively combines the feature information of the encoding layer and the decoding layer through the skip connection mechanism, making full use of high-level semantic features and detail information, and improving the prediction accuracy of the soil salinity inversion model. By applying the convolution operation in the last layer of the decoding network, the feature integration and processing capabilities are further enhanced, ensuring that the final generated soil salinity inversion result has high resolution and accuracy.

[0018] In the second aspect, an embodiment of the present application provides a soil salinity inversion device, which includes: a feature extraction module, a one-dimensional attention module, a two-dimensional attention module and an inversion module; the feature extraction module is used to extract remote sensing image features, temporal geoscientific features and spatial geoscientific features from multi-source remote sensing data; the one-dimensional attention module is used to calculate the soil salinity band sensitivity weight and the soil salinity time sensitivity weight of the multi-source remote sensing data based on the remote sensing image features and the temporal geoscientific features; the soil salinity band sensitivity weight and the soil salinity time sensitivity weight are used for the remote sensing image features and the temporal geoscientific features to obtain a one-dimensional attention map; the two-dimensional attention module is used to obtain a two-dimensional attention map based on the one-dimensional attention map and the spatial geoscientific features; the inversion module is used to obtain the soil salinity inversion result based on the two-dimensional attention map.

[0019] In a third aspect, an embodiment of the present application provides an electronic device, comprising a memory and a processor, wherein the memory stores program instructions, and when the processor reads and runs the program instructions, it executes the steps in any implementation of the above-mentioned soil salinity inversion method.

[0020] In a fourth aspect, an embodiment of the present application further provides a computer-readable storage medium, wherein the computer program instructions are stored in the computer-readable storage medium. When the computer program instructions are read and executed by a processor, the steps in any implementation of the above-mentioned soil salinity inversion method are executed. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments of the present application will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0022] Figure 1 A flow chart of a soil salinity inversion method provided in an embodiment of the present application;

[0023] Figure 2 A schematic diagram of an inversion model provided in an embodiment of the present application;

[0024] Figure 3 A flowchart of weight generation provided in an embodiment of the present application;

[0025] Figure 4 A flowchart for generating a one-dimensional attention map provided in an embodiment of the present application;

[0026] Figure 5 A flowchart of generating a two-dimensional attention map provided in an embodiment of the present application;

[0027] Figure 6 An inversion flow chart provided for an embodiment of the present application;

[0028] Figure 7 A schematic diagram of a module of a soil salinity inversion device provided in an embodiment of the present application;

[0029] Figure 8 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0030] The technical solutions in the embodiments of the present application will be described below in conjunction with the accompanying drawings in the embodiments of the present application. For example, the flowcharts and block diagrams in the accompanying drawings show the possible architectures, functions and operations of the systems, methods and computer program products according to multiple embodiments of the present invention. In this regard, each box in the flowchart or block diagram may represent a module, a program segment or a part of a code, and the module, a program segment or a part of a code contains one or more executable instructions for implementing the specified logical functions. It should also be noted that in some alternative implementations, the functions marked in the box may also occur in an order different from that marked in the accompanying drawings. For example, two consecutive boxes can actually be executed substantially in parallel, and they may sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or the flowchart, and the combination of boxes in the block diagram and / or the flowchart, can be implemented by a dedicated hardware-based system that performs a specified function or action, or can be implemented by a combination of dedicated hardware and computer instructions. In addition, the functional modules in the various embodiments of the present invention can be integrated together to form an independent part, or each module can exist separately, or two or more modules can be integrated to form an independent part.

[0031] Salt-alkali land, also known as saline soil, refers to land types with excessive accumulation of salt and alkali components on the soil surface, including salt land, alkaline land, salinized land and alkaline land. Soil salinization refers to the phenomenon and process in which the salt content of the soil surface continues to increase and exceeds a certain limit. About 1 billion hectares of farmland in the world are affected by salinization to varying degrees, resulting in a loss of up to 50% in grain production. In China, soil salinization mainly occurs in arid and semi-arid areas, and saline-alkali land accounts for about 10% of the total cultivated land area in the country, and the proportion is even higher in the north. Soil salinization is one of the main environmental problems causing land degradation worldwide, seriously affecting land health and agricultural production. Accurate monitoring of the salt content of cultivated soil is a prerequisite for the scientific development and utilization of saline-alkali land, increasing crop yields, and ensuring national food security.

[0032] Traditional field survey and soil testing methods select representative observation points for field sampling, and then obtain soil salinity distribution maps through interpolation. With the enrichment of data sources, a quantitative model of remote sensing reflectance and soil salinity is established, and the model is extended to the entire study area in combination with remote sensing images and multi-type highly correlated natural environmental characteristics to obtain the spatial distribution of soil salinity. Traditional methods rely on dense ground measurement points, which makes it difficult to achieve large-scale and long-term soil salinization monitoring. Therefore, the traditional soil salinity survey method has a large workload, and the accuracy of traditional interpolation results is limited.

[0033] With the increasing variety of remote sensing data, many studies have begun to use multi-source remote sensing data to obtain physical characteristics that are more sensitive to soil salinity, including hyperspectral remote sensing, multispectral remote sensing and radar. Remote sensing monitoring has become a mainstream method.

[0034] Hyperspectral remote sensing is currently an ideal means of accurate quantitative detection of soil salinization. It has the characteristics of dense spectral segments and rich information. It can detect the tiny interactions between soil salt and spectral information and is used for physical information analysis of salt. However, due to this, there is obvious collinearity between different bands, which increases the model complexity of data analysis and reduces the operational stability of the algorithm.

[0035] Synthetic Aperture Radar (SAR) has the characteristics of all-day and all-weather detection, but its band number is limited and the spatial resolution is low, which is not enough to fully understand the subtle interaction between soil salinity and spectral characteristics through radar bands alone. Therefore, SAR data is often used as an auxiliary source combined with other types of remote sensing data for soil salinity detection.

[0036] Although traditional optical remote sensing has the characteristics of high spatial resolution, multi-spectral, multi-temporal and multi-phase, it is easily affected by clouds and fog, making it difficult to obtain high-quality images; therefore, the accuracy of soil salinity inversion based on spectral bands or single parameters such as vegetation index and salt index is limited.

[0037] In terms of inversion methods, the existing indexes are significantly affected by different surface covers and have poor stability; from simple linear models to machine learning methods, the fitting quality is limited and it is difficult to handle the complex nonlinear mechanism of temporal and spatial changes in soil salinity.

[0038] The traditional salt inversion method based on remote sensing index is unstable in sensitivity to salt due to differences in geographical environment and vegetation type, and the fitting level of soil salinity is limited.

[0039] At present, in order to improve the accuracy of remote sensing inversion of soil salinity, related research is mainly carried out from the optimization of remote sensing inversion model construction methods. In model construction, more and more studies use machine learning algorithms to establish soil salinity inversion models with their powerful simulation capabilities, and select the optimal model by comparing the performance of different algorithms; and are committed to optimizing the spectral index characteristics that best reflect soil salinity to improve the accuracy of model inversion. However, due to the low correlation between soil salinity and remote sensing spectra, as well as the complexity of natural influences, the deep learning multi-layer perceptron fitting method still faces challenges in dealing with the complexity of soil parameter modeling, and the accuracy of soil salinity inversion still needs to be improved.

[0040] Based on this, this application proposes a soil salinity inversion method, device and electronic device. The soil salinity inversion method constructs a rich soil salinity inversion data set by integrating multi-source remote sensing image data, natural environmental factors, etc., and uses multi-band and multi-time series remote sensing data to establish a deep linear neural network through deep learning technology. In addition, an attention mechanism is introduced to couple natural environmental factors into the soil salinity inversion model, and to screen and emphasize key time and environmental factors that have a greater impact on soil salinity.

[0041] Please see Figure 1 , Figure 1 A flow chart of a soil salinity inversion method provided in an embodiment of the present application; The present application provides a soil salinity inversion method, which can be performed by Figure 8 electronic devices to perform.

[0042] The soil salinity inversion method includes the following steps:

[0043] Step S100: extracting remote sensing image features, temporal geoscience features and spatial geoscience features from multi-source remote sensing data.

[0044] In the above step S100, remote sensing image features, temporal geoscience features and spatial geoscience features are extracted from multi-source remote sensing data.

[0045] Among them, remote sensing image features refer to the spectral information obtained through remote sensing images, such as reflectivity, brightness, etc. Including hyperspectral, multispectral and SAR (synthetic aperture radar) data, multi-source remote sensing data can provide rich spectral and spatial information for analyzing the salt content of soil.

[0046] Temporal geoscience characteristics refer to geoscience data obtained at different time points, such as temperature, precipitation, etc. Temporal geoscience characteristics reflect the temporal variation trend of soil salinity.

[0047] Spatial geospatial characteristics refer to geographic spatial data, such as topography, terrain, soil type, etc. Spatial geospatial characteristics can help understand the spatial distribution of soil salinity.

[0048] Step S200: Calculate the soil salinity band sensitivity weight and soil salinity time sensitivity weight of multi-source remote sensing data based on remote sensing image features and temporal geoscience features.

[0049] In the above step S200, the sensitivity of different spectral bands to soil salinity is calculated to identify which bands are highly correlated with soil salt. Also, the sensitivity of different time series data to soil salinity is calculated to identify which time point features are highly correlated with soil salinity.

[0050] It should be noted that the calculation of soil salinity band sensitivity weight and soil salinity time sensitivity weight can be carried out through deep learning technology, using neural networks to train and optimize multi-source data, and then obtain soil salinity band sensitivity weight and soil salinity time sensitivity weight, which are used to evaluate the characteristics of the time points and which bands contribute most to soil salinity inversion.

[0051] Step S300: The soil salinity band sensitivity weight and the soil salinity time sensitivity weight are used for remote sensing image features and temporal geoscience features to obtain a one-dimensional attention map.

[0052] In the above step S300, by combining the band and time-sensitive weights, the remote sensing image features and the temporal geological features are weighted to generate a one-dimensional attention map.

[0053] The generated one-dimensional attention map can highlight the features that have the greatest impact on soil salinity inversion, reduce the interference of noise and irrelevant information, and improve the accuracy of soil salinity inversion.

[0054] Step S400: Obtain a two-dimensional attention map based on the one-dimensional attention map and spatial geological features.

[0055] In the above step S400, the one-dimensional attention map and the spatial geospatial features are combined to further integrate the spatial information and generate a two-dimensional attention map.

[0056] It should be understood that the two-dimensional attention map can reflect the spatial distribution of soil salinity in a more detailed manner, capture the differences between different plots, and improve the spatial resolution.

[0057] Step S500: Based on the two-dimensional attention map, obtain the soil salinity inversion result.

[0058] In the above step S500, the soil salinity is inverted by using a two-dimensional attention map through a deep linear neural network, and finally a spatial distribution map of soil salinity is generated. The inversion result can accurately reflect the distribution of soil salinity in the study area and provide a scientific basis for land management and crop planting.

[0059] pass Figure 1It can be seen that the soil salinity inversion method provided in this application greatly improves the accuracy and stability of soil salinity inversion by integrating multi-source remote sensing data and natural environmental elements, and using deep learning technology and attention mechanism. This soil salinity inversion method can dynamically monitor the seasonal changes and long-term trends of soil salinity, provide a soil salinity distribution map with high spatial resolution, and is suitable for soil salinity monitoring in different regions and environmental conditions. It enhances the reliability and applicability of the model, and provides strong technical support for the scientific development and utilization of saline-alkali land, increasing crop yields, and ensuring national food security.

[0060] In an optional embodiment, the multi-source remote sensing image data includes: multi-source remote sensing image data. The above step S100 extracts remote sensing image features, temporal geoscience features and spatial geoscience features from the multi-source remote sensing data, which can be achieved by: extracting remote sensing image features including spectral features and texture features from the multi-source remote sensing image data.

[0061] Multi-source remote sensing image data, including multispectral remote sensing data, hyperspectral remote sensing data and synthetic aperture radar (SAR) data. Among them, multispectral remote sensing data: including multispectral image data provided by satellites such as Landsat and Sentinel, covering visible light, near infrared and short-wave infrared bands. Hyperspectral remote sensing data includes hyperspectral image data such as AVIRIS and Hyperion, providing continuous spectral band information. Synthetic aperture radar (SAR) data such as Sentinel-1 and other SAR data provide radar image information all day and all weather.

[0062] In the above implementation process, remote sensing image features including spectral features and texture features are extracted from multi-source remote sensing image data.

[0063] For the extraction of spectral features, the reflectance values ​​of each band are first extracted, such as red light, green light, blue light, near infrared, etc. Then the commonly used spectral indexes, such as the Normalized Difference Vegetation Index (NDVI) and the Normalized Difference Salinity Index (NDSI), are calculated to enhance the sensitivity to specific landforms. Furthermore, PCA (Principal Component Analysis) or other dimensionality reduction methods are used to reduce the dimensionality of high-dimensional spectral data and extract the main spectral features.

[0064] For the extraction of texture features, texture features such as contrast, homogeneity, entropy, etc. are extracted from the image to reflect the roughness and structure of the soil surface. Furthermore, multi-scale texture information can be extracted through wavelet transform to capture the soil surface features at different scales.

[0065] In an optional embodiment, the multi-source remote sensing image data includes: geospatial time series data and geospatial space data. The above step S100 extracts remote sensing image features, temporal geospatial features and spatial geospatial features from the multi-source remote sensing data, which can be achieved by: extracting temporal features including precipitation and temperature from the geospatial time series data; extracting spatial geospatial features including soil type, crop planting information and terrain features from the geospatial space data.

[0066] Geoscience time series data, including meteorological data and historical soil data; meteorological data, such as temperature, precipitation, humidity and other meteorological variables, are usually obtained from meteorological stations or meteorological satellites, reflecting the time series changes of soil salinity. Historical soil data, including the measured soil salinity data accumulated over many years, reflects the long-term trend of soil salinity changes.

[0067] Geospatial data, including terrain data and land use / cover data. Terrain data, including DEM (digital elevation model), slope, aspect and other terrain features, reflects the impact of terrain on soil salt distribution. Land use / cover data, such as forest, grassland, farmland and other land cover types, affects soil salt accumulation and distribution.

[0068] In the above implementation process, remote sensing image features including spectral features and texture features are extracted from multi-source remote sensing image data, that is, temporal features related to soil salinity are extracted from geoscientific time series data, and spatial features related to soil salinity are extracted from geoscientific spatial data.

[0069] For the extraction of temporal geological features, the data of temperature, precipitation, humidity, etc. at different time scales (such as day, month, year) are statistically analyzed to calculate the average value, cumulative value, change rate, etc. Alternatively, the historical soil salinity measured data is used to extract its change trend and periodic characteristics, and prediction and analysis are performed through time series analysis methods.

[0070] For spatial geoscience feature extraction, terrain features can use DEM data to calculate terrain features such as slope, aspect, and elevation, and process them through terrain analysis tools such as ArcGIS and QGIS. Land cover features can analyze land use / cover types, and obtain land cover type maps through remote sensing image classification methods such as supervised classification and unsupervised classification.

[0071] It can be seen that this application can extract remote sensing image features including spectral features and texture features from multi-source remote sensing image data, and combine geoscientific time series features and spatial geoscientific features to form a rich soil salinity inversion data set. It effectively integrates multi-source, multi-time series, and multi-scale data, improves the accuracy and stability of the soil salinity inversion model, can dynamically monitor changes in soil salinity, and provide a soil salinity distribution map with high spatial resolution.

[0072] Please see Figure 2 , Figure 2 This is a schematic diagram of the inversion model provided in the embodiment of the present application; the soil salinity inversion method provided in the present application can be obtained by Figure 2 The inversion model implementation shown in .

[0073] Please Figure 2 See on the basis of Figure 3 , Figure 3 A flowchart of weight generation provided in an embodiment of the present application; in an optional implementation of an embodiment of the present application, the above step S200 calculates the soil salinity band sensitivity weight and soil salinity time sensitivity weight of multi-source remote sensing data based on remote sensing image features and temporal geoscience features, Figure 2 The "one-dimensional channel attention module" in is completed.

[0074] Step S200 in the above implementation process can be implemented by the following steps:

[0075] Step S210: performing average pooling and global maximum pooling on the remote sensing image features and the time series features, and performing feature fusion on the average pooling result and the global maximum pooling result to obtain pooling fusion features.

[0076] In the above step S210, remote sensing image features and time series geological features are extracted from multi-source remote sensing data, including spectral features, texture features, and geological features of time series changes. Average pooling and global maximum pooling operations are performed on the extracted remote sensing image features and time series features, respectively. These two pooling operations can effectively extract the overall information of the features. Average pooling captures the average value of the features, while global maximum pooling highlights the most significant parts of the features. Furthermore, the results of average pooling and global maximum pooling are subjected to feature fusion, so as to comprehensively utilize the information extracted by different pooling methods to obtain more comprehensive and rich pooling fusion features.

[0077] Step S220: Input the pooled fusion features into the fully connected layer and the activation function to obtain the soil salinity band sensitivity weight and the soil salinity time sensitivity weight.

[0078] In the above step S220, the pooled fusion features are input into the fully connected layer, and nonlinear transformation is performed through the activation function to obtain the soil salinity band sensitive weight and the soil salinity time sensitive weight. The fully connected layer can map the high-dimensional pooled fusion features to a lower-dimensional weight space, and the activation function increases the nonlinear expression ability of the model, which helps to capture complex soil salinity change patterns.

[0079] pass Figure 2 and Figure 3It can be seen that through a variety of feature extraction and pooling methods, the important information about soil salinity in remote sensing data is effectively captured, and the expressiveness and robustness of the features are improved. And by using the fully connected layer and activation function, the accurate modeling of the band and time sensitivity of soil salinity is achieved, so that the model can effectively adjust the degree of attention to different data sources according to the weight.

[0080] Please Figure 2 See on the basis of Figure 4 , Figure 4 A flowchart for generating a one-dimensional attention map provided in an embodiment of the present application; In an optional implementation of an embodiment of the present application, the above step S300 uses the soil salinity band sensitivity weight and the soil salinity time sensitivity weight for remote sensing image features and temporal geoscience features to obtain a one-dimensional attention map, and step S300 is Figure 2 The "one-dimensional channel attention module" in is completed, and finally a one-dimensional attention map is output.

[0081] The above step S300 can be implemented by the following steps:

[0082] Step S310: obtaining a soil salinity band sensitivity feature and a soil salinity time sensitivity feature based on the soil salinity band sensitivity weight and the soil salinity time sensitivity weight, respectively.

[0083] In the above step S310, the sensitivity of different spectral bands is weighted to extract band sensitivity features. For example, in the hyperspectral data, a specific band that is sensitive to salt reflection is selected, and the band sensitivity features are obtained by weighted average (based on Figure 2 The multi-temporal attention module in .

[0084] In addition, different time series features are weighted to extract time-sensitive features. For example, in the time series meteorological data, the data at specific time points that have a significant impact on the change in salinity are selected, and the time-sensitive features are obtained by weighted averaging (based on Figure 2 The multi-band attention module in is completed).

[0085] Step S320: Fusing the soil salt band sensitivity feature and the soil salt time sensitivity feature to obtain the soil salt sensitivity fusion feature.

[0086] In the above step S3200, the band sensitive features and the time sensitive features are fused to generate soil salinity sensitive fused features. This can be achieved by feature concatenation or weighted summation. For example, assuming that the band sensitive features are B1, B2, B3 and the time sensitive features are T1, T2, the comprehensive features obtained after fusion are B1, B2, B3, T1, T2.

[0087] Step S330: Input the soil salt sensitivity fusion feature into the fully connected layer and the activation function to obtain a one-dimensional attention map.

[0088] In the above step S330, the fused features are input into the fully connected layer for feature transformation and weight learning, and nonlinear activation functions (such as ReLU and Sigmoid) are used to enhance the expression ability of the model.

[0089] Through the fully connected layer and activation function, a one-dimensional attention map is generated to highlight the features that have the greatest impact on soil salinity inversion.

[0090] pass Figure 2 and Figure 4 It can be seen that the soil salinity inversion method provided in the embodiment of the present application, through the application of deep learning technology and attention mechanism, enables the inversion model to utilize multi-source remote sensing data more comprehensively and accurately, solves the problem of rich data but low information utilization in traditional methods, and enhances the model's perception of spatial and temporal changes in soil salinity.

[0091] Please continue in Figure 2 See on the basis of Figure 5 , Figure 5 A flowchart of generating a two-dimensional attention map provided in an embodiment of the present application; In an optional implementation of an embodiment of the present application, the above step S400 obtains a two-dimensional attention map according to the one-dimensional attention map and the spatial geography feature, which can be obtained by Figure 2 The “two-dimensional spatial attention module” in the embodiment of the present invention is implemented, and step S400 in the above process may include the following steps:

[0092] Step S410: Fusing the one-dimensional attention map with the spatial geospatial features to obtain spatiotemporal fusion features.

[0093] In the above step S410, the one-dimensional attention map and the spatial geography feature are fused. This can be done by simple element-by-element multiplication, concatenation, or using a specific neural network structure. For example, the one-dimensional attention map can be element-by-element multiplied with each pixel of the spatial geography feature to add attention to a specific area of ​​space.

[0094] Step S420: Use the spatial attention mechanism on the spatiotemporal fusion features to obtain a two-dimensional attention map.

[0095] In the above step S420, based on the fused spatiotemporal features, a spatial attention mechanism is introduced to generate a two-dimensional attention map. Exemplarily, a convolutional neural network or an attention mechanism network is used to generate a two-dimensional attention map based on the fused spatiotemporal features. The relationship between adjacent pixels in space and the importance weights of different positions in the one-dimensional attention map can be considered to generate a more accurate two-dimensional attention distribution.

[0096] pass Figure 2 and Figure 5 It can be seen that the soil salinity inversion method provided in the embodiment of the present application effectively combines time series information and spatial distribution characteristics by fusing one-dimensional attention maps and spatial geological features, thereby improving the model's perception of soil salinity changes. And by further introducing a spatial attention mechanism, the model can more accurately distinguish soil salinity changes in different regions and time points, thereby improving the spatial resolution and accuracy of the inversion results.

[0097] Please see Figure 6 , Figure 6 Inversion flow chart provided in the embodiment of the present application; in an optional implementation of the embodiment of the present application, the above step S500 obtains the soil salinity inversion result based on the two-dimensional attention map, which can be achieved by the following steps:

[0098] Step S510: Decode the one-dimensional attention map and the two-dimensional attention map, and at each decoding layer, use a jump connection to fuse the corresponding encoding layer feature map with the current decoding layer feature map.

[0099] In the above step S510, in each layer of the decoding network, the corresponding encoding layer feature map is fused with the current decoding layer feature map. Exemplarily, the fusion can be achieved through a skip connection to ensure that important high-level semantic information and spatial details are retained during the decoding process. It is worth noting that the skip connection allows information to be directly transferred from the encoder to the corresponding layer of the decoder, which can provide richer semantic information and finer spatial resolution, thereby helping to more accurately predict the soil salinity inversion results.

[0100] Step S520: In the last layer of the decoding network, the result of fusing the corresponding encoding layer feature map with the current decoding layer feature map is convolved to obtain the soil salinity inversion result.

[0101] In the above step S520, for the convolution operation of the last layer, the skip connection allows information to be directly transmitted from the encoder to the corresponding layer of the decoder, which can provide richer semantic information and finer spatial resolution, thereby helping to more accurately predict the soil salinity inversion results.

[0102] pass Figure 6 It can be seen that the soil salinity inversion method provided by this application effectively combines the feature information of the encoding layer and the decoding layer through the skip connection mechanism, making full use of high-level semantic features and detail information, and improving the prediction accuracy of the soil salinity inversion model. By applying the convolution operation in the last layer of the decoding network, the feature integration and processing capabilities are further enhanced, ensuring that the final generated soil salinity inversion result has high resolution and accuracy.

[0103] Please see Figure 7 , Figure 7 A module schematic diagram of a soil salinity inversion device provided in an embodiment of the present application; the present application provides a soil salinity inversion device, the soil salinity inversion device 100 includes a feature extraction module 110, a one-dimensional attention module 120, a two-dimensional attention module 130 and an inversion module 140.

[0104] The soil salinity inversion device 100 can be based on DNN as the basic framework, combined with multi-band multi-temporal remote sensing image data sets, agricultural management and crop distribution, and natural environmental factors (including DEM soil type and other information, time series temperature, precipitation, etc.) to form a soil salinity inversion data set. Based on the soil salinity inversion data set, a convolutional attention module (CBAM) is used between convolution blocks, and multi-temporal attention and multi-band attention are combined to take the multi-dimensional feature map F (C×H×W) as input, and the one-dimensional attention module 120 and the two-dimensional attention module 130 are sequentially inferred to obtain a one-dimensional attention map Mc (C×1×1) and a two-dimensional attention map Ms (S×H×W). Among them, s is the number of features used in the two-dimensional attention module 130, and C, H, and W represent different dimensions of the feature map.

[0105] The feature extraction module 110 is used to extract remote sensing image features, temporal geoscience features and spatial geoscience features from multi-source remote sensing data.

[0106] The one-dimensional attention module 120 is used to calculate the soil salinity band sensitivity weight and the soil salinity time sensitivity weight of multi-source remote sensing data based on remote sensing image features and temporal geological features; the soil salinity band sensitivity weight and the soil salinity time sensitivity weight are used for remote sensing image features and temporal geological features to obtain a one-dimensional attention map.

[0107] The two-dimensional attention module 130 is used to obtain a two-dimensional attention map according to the one-dimensional attention map and the spatial geospatial features.

[0108] The inversion module 140 is used to obtain a soil salinity inversion result based on the two-dimensional attention map.

[0109] In an optional embodiment, the multi-source remote sensing image data includes: multi-source remote sensing image data; in the process of extracting remote sensing image features, temporal geological features and spatial geological features from the multi-source remote sensing data, the feature extraction module 110 is used to: extract remote sensing image features including spectral features and texture features from the multi-source remote sensing image data.

[0110] In an optional embodiment, the multi-source remote sensing image data includes: geological time series data and geological spatial data; in the process of extracting remote sensing image features, time series geological features and spatial geological features from the multi-source remote sensing data, the feature extraction module 110 is used to: extract time series features including precipitation and temperature from the geological time series data; extract spatial geological features including soil type, crop planting information and terrain features from the geological spatial data.

[0111] In an optional embodiment, in the process of calculating the soil salinity band sensitivity weights and the soil salinity time sensitivity weights of multi-source remote sensing data based on remote sensing image features and time series geological features, the one-dimensional attention module 120 is used to: perform average pooling and global maximum pooling on the remote sensing image features and time series features, and perform feature fusion of the average pooling results and the global maximum pooling results to obtain pooled fusion features; use the pooled fusion features as input to the fully connected layer and activation function to obtain the soil salinity band sensitivity weights and the soil salinity time sensitivity weights.

[0112] In an optional embodiment, in the process of using soil salt band sensitivity weights and soil salt time sensitivity weights for remote sensing image features and time series geological features to obtain a one-dimensional attention map, the one-dimensional attention module 120 is used to: obtain soil salt band sensitivity features and soil salt time sensitivity features based on soil salt band sensitivity weights and soil salt time sensitivity weights, respectively; fuse soil salt band sensitivity features and soil salt time sensitivity features to obtain soil salt sensitive fusion features; use the soil salt sensitive fusion features as input to a fully connected layer and an activation function to obtain a one-dimensional attention map.

[0113] For the one-dimensional attention module 120: the input Sentinel-1 and Sentinel-2 remote sensing image feature maps F (C × H × W) are subjected to global maximum pooling and global average pooling respectively to obtain two 1 × 1 × C sequence feature maps of multi-bands in the target period of soil salinity inversion, and then respectively input into a two-layer weight sharing neural network (MLP), the number of neurons in the first layer is C / r, r represents the number of repetitions of neuron training, the activation function is ReLU, the number of neurons in the second layer is C, and then the features output by the MLP are subjected to pixel-based addition operations, and then through the Sigmoid activation operation to generate the channel attention map M c Finally, M c A pixel-based multiplication operation is performed on the input feature map F, and the features at different spatial positions are weighted and combined according to the attention weight, so as to focus on important feature areas and ignore the features of unimportant areas; the full-band fusion features of the remote sensing image calculated under the multi-band data calculation are measured through the calculation of the multi-band attention module, and are used as the feature input of the two-dimensional attention module 130.

[0114] For remote sensing soil salinity index data, precipitation, temperature, groundwater and other time-series natural environmental factors, they also enter the multi-time series data encoding stage under the time attention mechanism, and map each time series data (such as precipitation, temperature, groundwater, etc.) into a continuous vector space to obtain the corresponding embedded vector sequence x1, x2,…, x T , and then the embedding vector sequence is input into the encoder (Formula 1, Formula 2). The output of the encoding stage is used as the input of the bidirectional long short-term memory model (LSTM) module, combining the information in the forward and backward directions of the time series. Finally, the output of each moment is combined with the temporal attention mechanism and aggregated with the output of each convolution module respectively; the attention weight of each time step is calculated, and then the spatiotemporal features of all time steps are aggregated into the overall soil salt sensitivity fusion feature; the temporal attention mechanism calculates the correlation score based on the contribution rate of each time step to the inversion performance. Specifically, the parameter θ is used. A A feed-forward neural network, The attention weight α of each time step is obtained by spatial averaging and softmax normalization operation t (Formula 3), and then the features H of all time steps are t Aggregate into time feature C (Formula 4). t ∈[0,1],α t A higher value indicates that time t contains more information.

[0115] Z0=Attn(Z l-1 ,Z l-1 ,Z l-1 )+Z l-1 (1)

[0116] Among them, Attn is the Multi-Head Attention operation, and the specific calculation formula is:

[0117]

[0118] Finally, the temporal attention weight Mc time A pixel-based multiplication operation is performed on the input feature map F, and the full-temporal fusion features of the remote sensing image calculated under multi-temporal data are measured through the calculation of the channel attention module for the feature input of the spatial attention module.

[0119] In an optional embodiment, in the process of obtaining a two-dimensional attention map based on a one-dimensional attention map and spatial geological features, the two-dimensional attention module 130 is used to: fuse the one-dimensional attention map and the spatial geological features to obtain spatiotemporal fusion features; and use a spatial attention mechanism on the spatiotemporal fusion features to obtain a two-dimensional attention map.

[0120] Combined with the terrain, soil texture, and crop planting conditions (spatial geoscience features) of the natural environment, and adding the full-band and full-time fusion feature image layer constructed in the one-dimensional attention module, the CNN spatial convolution method is used as the basic structure, combined with the spatial attention mechanism module in CBAM. First, the above image feature map is input through three different groups of convolution modules and linear transformation layers to obtain three new feature maps, namely the query matrix Q, the key matrix K, and the value matrix V.

[0121] The specific operation is as follows: Query matrix Q: Q = XW Q +b Q , key matrix K: K = XW K +b K , value matrix V: V = XW V +b V ; Further, the similarity between the query matrix Q and the key matrix K is calculated to obtain an attention scoremap A (Formula 5):

[0122] A(x,y)=sim(Q(x,y),K) (5)

[0123] Among them, sim represents a similarity function, such as dot product or scaled dot product. Secondly, the attention score A is normalized by the softmax function to obtain the attention weight map α (Formula 6):

[0124] α(x,y)=softmax(A(x,y)) (6)

[0125] Finally, the attention weight map α is weighted and summed with the value feature map V to obtain the attention output feature map O (Formula 7):

[0126]

[0127] In an optional embodiment, in the process of obtaining the soil salinity inversion result based on the two-dimensional attention map, the inversion module 140 is used to: decode the one-dimensional attention map and the two-dimensional attention map, and at each decoding layer, use a jump connection to fuse the corresponding encoding layer feature map with the current decoding layer feature map; at the last layer of the decoding network, convolve the result of the fusion of the corresponding encoding layer feature map and the current decoding layer feature map to obtain the soil salinity inversion result.

[0128] See also Figure 8 , Figure 8The electronic device 200 provided in the embodiment of the present application includes: a processor 201 and a memory 202, wherein the memory 202 stores machine-readable instructions executable by the processor 201, and when the machine-readable instructions are executed by the processor 201, the above method is executed.

[0129] Based on the same inventive concept, an embodiment of the present application also provides a computer-readable storage medium, in which computer program instructions are stored. When the computer program instructions are read and executed by a processor, the steps in any implementation method of the above-mentioned soil salinity inversion method are executed.

[0130] The computer-readable storage medium may be a random access memory (RAM), a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), or any other medium that can store program code.

[0131] In this article, relational terms such as first and second, etc. are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the statement "include..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.

[0132] The above description is only an embodiment of the present application and is not intended to limit the protection scope of the present application. For those skilled in the art, the present application may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A soil salinity inversion method, characterized in that: The method comprises: Extracting remote sensing image features, temporal geoscience features and spatial geoscience features from multi-source remote sensing data; wherein the multi-source remote sensing data includes geoscience temporal data and geoscience spatial data; Performing average pooling and global maximum pooling on the remote sensing image features and time series features, and performing feature fusion on the average pooling results and the global maximum pooling results to obtain pooling fusion features; inputting the pooling fusion features into the fully connected layer and the activation function to obtain the soil salinity band sensitivity weight and the soil salinity time sensitivity weight; Based on the soil salt band sensitivity weight and the soil salt time sensitivity weight, respectively, a soil salt band sensitivity feature and a soil salt time sensitivity feature are obtained; the soil salt band sensitivity feature and the soil salt time sensitivity feature are fused to obtain a soil salt sensitive fusion feature; the soil salt sensitive fusion feature is input into a fully connected layer and an activation function to obtain a one-dimensional attention map; Obtaining a two-dimensional attention map according to the one-dimensional attention map and the spatial geoscientific features; Based on the two-dimensional attention map, a soil salinity inversion result is obtained; Among them, the extraction of remote sensing image features, time series geological features and spatial geological features from multi-source remote sensing data includes: extracting time series geological features including precipitation and temperature from the geological time series data; extracting spatial geological features including soil type, crop planting information and terrain characteristics from the geological spatial data.

2. The method according to claim 1, characterized in that in, The multi-source remote sensing data includes: multi-source remote sensing image data; The extracting of remote sensing image features, temporal geoscience features and spatial geoscience features from multi-source remote sensing data includes: Remote sensing image features including spectral features and texture features are extracted from the multi-source remote sensing image data.

3. The method according to claim 1, characterized in that The step of obtaining a two-dimensional attention map according to the one-dimensional attention map and the spatial geoscientific features includes: Fusion of the one-dimensional attention map and the spatial geoscience feature to obtain a spatiotemporal fusion feature; A spatial attention mechanism is used on the spatiotemporal fusion features to obtain the two-dimensional attention map.

4. The method according to claim 1, characterized in that: The soil salinity inversion result is obtained based on the two-dimensional attention map, including: Decoding the one-dimensional attention map and the two-dimensional attention map, and at each decoding layer, fusing the corresponding encoding layer feature map with the current decoding layer feature map using a skip connection; In the last layer of the decoding network, the result of fusing the corresponding encoding layer feature map with the current decoding layer feature map is convolved to obtain the soil salinity inversion result.

5. A soil salinity inversion device, characterized in that: The device comprises: a feature extraction module, a one-dimensional attention module, a two-dimensional attention module and an inversion module; The feature extraction module is used to extract remote sensing image features, temporal geoscience features and spatial geoscience features from multi-source remote sensing data; wherein the multi-source remote sensing data includes geoscience temporal data and geoscience spatial data; the extraction of remote sensing image features, temporal geoscience features and spatial geoscience features from multi-source remote sensing data includes: extracting temporal geoscience features including precipitation and temperature from the geoscience temporal data; extracting spatial geoscience features including soil type, crop planting information and terrain features from the geoscience spatial data; The one-dimensional attention module is used to perform average pooling and global maximum pooling on the remote sensing image features and time series features, and perform feature fusion of the average pooling results and the global maximum pooling results to obtain pooling fusion features; the pooling fusion features are input into the fully connected layer and the activation function to obtain the soil salt band sensitive weight and the soil salt time sensitive weight; based on the soil salt band sensitive weight and the soil salt time sensitive weight, the soil salt band sensitive feature and the soil salt time sensitive feature are obtained respectively; the soil salt band sensitive feature and the soil salt time sensitive feature are fused to obtain the soil salt sensitive fusion feature; the soil salt sensitive fusion feature is input into the fully connected layer and the activation function to obtain a one-dimensional attention map; The two-dimensional attention module is used to obtain a two-dimensional attention map according to the one-dimensional attention map and the spatial geoscience feature; The inversion module is used to obtain a soil salinity inversion result based on the two-dimensional attention map.

6. An electronic device, characterized in that: The electronic device comprises a memory and a processor, wherein the memory stores program instructions, and when the processor runs the program instructions, the steps in the method according to any one of claims 1 to 4 are executed.

7. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer program instructions, and when the computer program instructions are executed by a processor, the steps in the method according to any one of claims 1 to 4 are executed.

Images