Unmanned aerial vehicle remote sensing inversion soil moisture content accurate monitoring method and system

Abstract

The invention relates to the technical field of soil moisture content monitoring, and discloses a method and a system for accurately monitoring soil moisture content through remote sensing inversion of an unmanned aerial vehicle. Acquiring original spectral data of different soil types and environmental conditions through a spectral sensor to obtain an initial multispectral image data set; dividing soil types by using a support vector machine classification method according to pixel reflection characteristics to form a soil type grouping set; calculating a multispectral band reflectivity variance for each group, and screening candidate bands to construct a set; processing the candidate wave band set through a random forest regression model, adjusting the weight in combination with an environment variable, and performing iterative screening to form an optimized wave band combination sequence; extracting spectral features of the optimized wave band, and carrying out weighted fusion on the soil variation region to output a processed image; and calculating the soil moisture content by using a linear inversion model based on the pixel reflectivity of the processed image, and outputting a distribution diagram when the deviation reaches the standard. The method is suitable for different soil and environments, the wave band is accurately selected, the precision is improved, and high-quality data is provided for agricultural and environmental governance.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of soil moisture content monitoring, and in particular to a method and system for accurately monitoring soil moisture content by unmanned aerial vehicle remote sensing. BACKGROUND

[0002] Accurate monitoring of soil moisture content is the core basis of agricultural production, environmental management and optimal allocation of water resources, directly affecting the scientificity of crop irrigation decision-making, the effectiveness of ecological protection and the efficiency of water resource utilization, and is of key significance to food security and ecological balance. Unmanned aerial vehicle remote sensing technology has become an important technical means for monitoring soil moisture content due to its high efficiency, flexibility and high resolution, and can quickly cover a large area of land and obtain spectral data, making accurate monitoring possible. However, it still faces many technical bottlenecks in practical application, and needs to be improved in monitoring accuracy and environmental adaptability. Current methods for monitoring soil moisture content by unmanned aerial vehicle remote sensing have obvious limitations. Most techniques rely on a single monitoring mode or fixed data processing flow, making it difficult to adapt to changes in different soil types and complex environmental conditions. For example, traditional remote sensing methods often use general band combinations for data processing, without considering the influence of soil property differences on spectral response, resulting in large fluctuations in monitoring accuracy between different plots, especially in complex terrain or climate change areas, where monitoring data is easily distorted, and cannot meet the demand for real-time monitoring of dynamic changes in soil moisture for precision agriculture, restricting the efficiency of scientific irrigation and field management. The core technical problem is that existing unmanned aerial vehicle remote sensing monitoring techniques cannot efficiently utilize multispectral data to improve the accuracy of soil moisture content inversion - they cannot accurately identify and optimize the combination of near-infrared, short-wave infrared and other bands that are sensitive to soil moisture, and lack a self-adaptive band selection mechanism, making fixed combinations unable to adapt to the spectral reflectance differences of different soil types such as clay and sand, resulting in low data utilization efficiency and large inversion errors, which cannot provide reliable data support for agricultural production and resource management, and there is an urgent need to develop an optimization technique that can dynamically adapt to soil and environmental conditions. SUMMARY

[0003] The application provides a method and system for accurately monitoring soil moisture content by unmanned aerial vehicle remote sensing, which can adapt to different soils and environments, accurately select bands, improve accuracy, and provide high-quality data for agriculture and environmental management.

[0004] In the first aspect, to solve the above technical problems, the application provides a method for accurately monitoring soil moisture content by unmanned aerial vehicle remote sensing, comprising: acquiring initial multispectral image data sets containing near-infrared and short-wave infrared bands by a mobile platform-mounted spectral sensor collecting original spectral data of different physicochemical properties of soil types and different environmental conditions; According to the pixel reflection characteristics of the initial multi-spectral image data set, the soil type of the pixel region of the land block is divided by using a support vector machine classification method, regions with a clay component ratio exceeding a preset ratio threshold are marked as a first group, and the remaining regions are marked as a second group, to form a soil type grouping set; For each group in the soil type grouping set, the reflectivity variance of the multi-spectral band is calculated respectively, and the multi-spectral band with a reflectivity variance exceeding a preset variance threshold is selected as a candidate band to construct a corresponding candidate band set; A preset random forest regression model is used to process the candidate band set, environmental condition variables including illumination and vegetation coverage are input to dynamically adjust the band weight, and the band combination covering the water-sensitive peak value in the top three of the weight order is iteratively screened to form an optimized band combination sequence; The spectral features corresponding to the optimized band combination sequence are extracted from the initial multi-spectral image data set as a data set feature, and the weighted average method is used to fuse the features of the soil variation region of the complex land block to output a processed image with uniform pixel value distribution and enhanced water-sensitive response; Based on the pixel reflectivity of the processed image, a linear inversion model is used to calculate the soil water content value, and if the deviation of the soil water content value from the historical verification data is less than a preset water content threshold, a soil water content distribution map is output; The spectral features include band reflectivity, spectral feature parameters, and band-to-band difference and ratio. The historical verification data includes soil water content measured values measured by field sampling in past monitoring of the target land block, corresponding sampling point coordinates, and synchronous recorded illumination intensity and vegetation coverage.

[0005] In an optional implementation, according to the pixel reflection characteristics of the initial multi-spectral image data set, the soil type of the pixel region of the land block is divided by using a support vector machine classification method, regions with a clay component ratio exceeding a preset ratio threshold are marked as a first group, and the remaining regions are marked as a second group, to form a soil type grouping set, which includes: The pixel reflection characteristics are input into a pre-trained support vector machine classification model to determine the soil component composition ratio of each pixel region through operation analysis; According to the soil component composition ratio, the pixel region with a clay component ratio exceeding a preset ratio threshold is identified as a clay-dominant type region; Regions identified as clay-dominant type are marked as a first group, and the remaining regions not exceeding the preset ratio threshold are marked as a second group; The spatial distribution information of all regions marked as the first group and the second group is integrated to construct a soil type grouping set reflecting the soil type zoning characteristics of the target land block. The support vector machine classification model is trained by sample spectrum data covering clay and other soil types, and has built-in spectrum response discrimination rules for different soil types.

[0006] In an optional implementation, for each group in the soil type grouping set, the reflectance variance of the multi-spectral band is calculated respectively, and the multi-spectral band with reflectance variance exceeding a preset variance threshold is selected as a candidate band to construct a corresponding candidate band set, including: According to each group in the soil type grouping set, the pixel area range covered by each group and the corresponding multi-spectral band are determined; For all pixel areas contained in each group, the reflectance variance of the corresponding group on each multi-spectral band is calculated respectively; The reflectance variance of each multi-spectral band is compared with the preset variance threshold, and the multi-spectral band exceeding the preset variance threshold is selected and included in the candidate range; Based on the candidate range, the candidate range bands are respectively summarized and filtered according to the soil type grouping to construct a candidate band set corresponding to each group; The reflectance variance is used to represent the sensitivity of the band to the moisture change of the soil type.

[0007] In an optional implementation, the preset random forest regression model is used to process the candidate band set, the environmental condition variable including light and vegetation coverage is input to dynamically adjust the band weight, the band combination covering the water sensitive peak value and ranking the top three in the weight order is iteratively selected to form an optimized band combination sequence, including: The band data in the candidate band set is input into the random forest regression model, and the environmental condition variable including light intensity and vegetation coverage ratio is imported as an adjustment parameter; The sensitive contribution of each band to soil moisture under the current environmental condition is analyzed by operation and analysis, and the weight value of each band is dynamically generated and adjusted; According to the weight value, the bands are sorted, a preset proportion of bands are selected as preliminary candidate bands from high to low, and it is checked whether the top three bands in the preliminary candidate bands cover the water sensitive peak value; If yes, the preliminary candidate band is determined as an effective optimization combination; If not, the weight influence factor of the environmental condition variable is readjusted, and the band weight calculation and sorting selection are iteratively performed; The finally determined effective optimization combination is sorted according to the soil type grouping to form the optimized band combination sequence; The random forest regression model is trained by containing samples of band data and water content correlation under different soil types and environmental conditions, and has a built-in weight calculation rule for the sensitivity of bands to water.

[0008] In an optional implementation, the spectral features corresponding to the optimized band combination sequence are extracted from the initial multispectral image dataset as dataset features, and a weighted average method is used for feature fusion of the soil variation region of the complex plot, and a processed image with uniform pixel value distribution and enhanced water-sensitive response is output, including: determining a target band according to the optimized band combination sequence, and extracting spectral features corresponding to the target band from the initial multispectral image dataset to form a spectral feature dataset matched with the optimized band combination; Based on the spatial distribution of the spectral feature dataset, the soil variation region composed of soil type transition zone and micro-topographic change area is identified, and the pixel coordinate range and the corresponding spectral feature subset are marked; For the marked soil variation region and the corresponding spectral feature subset, the spectral feature weight value is set according to the proportion of different position soil types; Based on the spectral feature weight value, a weighted average algorithm is used to perform fusion operation on the spectral feature subset to generate a unified feature value covering all pixels of the soil variation region; Embed the unified feature value into the spectral feature dataset according to the pixel coordinate range to form the fused image data; Perform uniformity verification on the fused image data, calculate the gradient change rate of pixel value distribution, and determine that the verification is passed when the gradient change rate is less than a preset threshold, and output the verified image data as the processed image; The target band is a multispectral band with high sensitivity to soil water content variation after weight screening, and covers the water-sensitive peak values of near-infrared and short-wave infrared, and reflects the soil moisture content difference.

[0009] In an optional implementation, based on the pixel reflectivity of the processed image, a linear inversion model is used to calculate the soil moisture content value, and if the deviation of the soil moisture content value from the historical verification data is less than a preset moisture content threshold, a soil moisture content distribution map is output, including: Extract the reflectivity data of each pixel region in the processed image and input it into the linear inversion model to obtain the corresponding soil moisture content calculation value of each pixel region through calculation; Retrieve the historical verification data of the target plot, compare the soil moisture content calculation value with the historical verification data in the same region, and calculate the deviation between them; When the deviation degree is less than a preset deviation threshold, the effective calculation result is visualized according to a pixel area space distribution to generate a soil water content distribution map reflecting the difference in soil water content in the land plot; The reflectivity data is spectral reflectivity information enhanced by feature fusion and sensitive to water content, the linear inversion model is trained by associating samples of soil reflectivity and measured water content, and the historical verification data includes measured soil water content and corresponding spectral reflectivity features at different positions.

[0010] In an optional implementation, after the soil water content distribution map is output, the following is further included: The regional water variation data of the soil water content distribution map is acquired, a clustering algorithm is used to group the data according to the continuity of the water content gradient, the stable regions and the regions to be optimized are marked, and an extended environment adaptive grouping set is formed; Based on the environment adaptive grouping set, a wave band sensitivity index of spectral reflectivity dynamic change is calculated, a wave band set with enhanced sensitivity is constructed by screening wave bands with a sensitivity higher than a preset threshold, secondary feature extraction is performed on the processed image through the enhanced wave band set, adaptive fusion is performed on the regions to be optimized, an enhanced image with uniform distribution after fusion is combined with the original distribution map, and a final soil water content distribution map is obtained.

[0011] In an optional implementation, the acquisition of the regional water variation data of the soil water content distribution map, the use of a clustering algorithm to group the data according to the continuity of the water content gradient, the marking of the stable regions and the regions to be optimized, and the formation of an extended environment adaptive grouping set include the following steps: The water content values and spatial position information of each pixel in the soil water content distribution map are extracted, the water content difference between adjacent pixels is calculated, and a regional water variation data set containing the absolute value of the water content, the spatial change gradient, and the position correlation is integrated; Based on the regional water variation data set, the distance determination standard and the minimum aggregation unit of the clustering algorithm are set in combination with the soil water conduction characteristics of the land plot, and a density clustering algorithm is used to perform aggregation grouping on adjacent pixels according to the natural continuity of the water content gradient; For each group formed by clustering, the dispersion degree of the water content and the uniformity of the spatial distribution are calculated, and when both meet the preset standard, the group is marked as a stable region, and when either does not meet the preset standard, the group is marked as a region to be optimized; The marked stable regions and the regions to be optimized are associated with corresponding soil type grouping information, and the terrain undulation and the vegetation coverage are supplemented as environmental labels, and an environment adaptive grouping set containing the region boundary, the water content statistical characteristics, and the environmental labels is integrated.

[0012] In an optional embodiment, the soil moisture content final distribution map is obtained by performing secondary feature extraction on the processed image through the enhanced band set and performing adaptive fusion on the region to be optimized, synthesizing the enhanced image with uniform distribution after fusion and the original distribution map, comprising: Based on the optimized enhanced band set, secondary spectral feature extraction is performed on the generated processed image to obtain fine spectral feature data of the region to be optimized under the enhanced band; Based on the fine spectral feature data, an adaptive fusion algorithm is called to dynamically adjust the fusion weight according to the soil type and terrain characteristics of the region to be optimized to complete feature fusion; The distribution uniformity of the fused image data is verified, the soil moisture natural distribution characteristics are matched by analyzing the pixel value gradient change rule, and the image data meeting the distribution rule is marked as an enhanced image; The marked enhanced image is aligned with the initial output soil moisture content distribution map in spatial coordinates, the original data of the stable region and the fused data of the region to be optimized are retained, and the soil moisture content final distribution map is generated.

[0013] In a second aspect, the application also provides an unmanned aerial vehicle remote sensing soil moisture content inversion precision monitoring system, comprising: A data acquisition module: through a spectrum sensor carried by a mobile platform, original spectrum data of different physicochemical properties of soil types and different environmental conditions are collected, and initial multispectral image data sets containing near-infrared bands and short-wave infrared bands are obtained; A soil grouping module: according to the pixel reflection characteristics of the initial multispectral image data set, a support vector machine classification method is used to divide the soil types of the pixel region of the land, and the region with clay composition ratio exceeding a preset ratio threshold is marked as a first grouping, and the remaining region is marked as a second grouping, forming a soil type grouping set; A candidate band screening module: for each grouping in the soil type grouping set, the reflectivity variance of the multispectral band is calculated respectively, and the multispectral band with reflectivity variance exceeding a preset variance threshold is selected as a candidate band to construct a corresponding candidate band set; A band optimization module: a preset random forest regression model is used to process the candidate band set, environmental condition variables including light and vegetation coverage are input to dynamically adjust the band weight, and the band combination covering the water sensitive peak value in the top three of the weight order is iteratively screened to form an optimized band combination sequence; A weighted fusion module: the spectral features corresponding to the optimized band combination sequence are extracted from the initial multispectral image data set as a data set feature, and the weighted average method is used for feature fusion on the soil variation region of the complex land to output a processed image with uniform pixel value distribution and enhanced water sensitive response; The image output module: based on the pixel reflectivity of the processed image, a linear inversion model is used to calculate the soil moisture content value, and if the deviation of the soil moisture content value from the historical verification data is less than a preset moisture content threshold, a soil moisture content distribution map is output. The spectral features include band reflectivity, spectral feature parameters, and inter-band difference and ratio, and the historical verification data include soil moisture content measured values measured by field sampling in past monitoring of the target plot, corresponding sampling point coordinates, and synchronous recorded illumination intensity and vegetation coverage at the time of sampling.

[0014] Compared with the prior art, the present application has the following beneficial effects: (1) The mobile platform carries the spectral sensor to collect original spectral data of multiple soil types and multiple environmental conditions, and obtain initial multispectral image data sets containing near-infrared and short-wave infrared bands. This acquisition method covers the key water-sensitive bands and can fully capture the spectral response differences under different soil physicochemical properties and environmental conditions, laying a data foundation for subsequent accurate inversion and improving the ability to acquire spectral information of complex plots. (2) The support vector machine classification method is used to divide the soil types according to the pixel reflectance characteristics to form a soil type grouping set. By accurately distinguishing the clay content threshold and the remaining area, the soil types are finely grouped, avoiding the processing deviation caused by mixed soil characteristics, providing grouping basis for subsequent targeted band selection, and improving the pertinence of data processing. (3) The multispectral band reflectivity variance is calculated for each soil grouping, the candidate bands are selected, and the set is constructed. By focusing on the bands with reflectivity variance exceeding the threshold, the bands more sensitive to soil moisture response are preliminarily selected, and redundant bands are removed, improving the data utilization efficiency and providing a high-quality candidate basis for subsequent optimization of band combination. (4) The random forest regression model is used to process the candidate band set, and the band weight is dynamically adjusted in combination with environmental variables such as illumination and vegetation coverage to iteratively select and optimize the band combination sequence. This adaptive weight adjustment mechanism can adapt to different environmental conditions and soil types, select the optimal band combination covering the water-sensitive peak, solve the problem of poor adaptability of fixed band combination, and improve the scientificity of band combination. (5) The weighted average method is used to fuse the features of the complex plot soil variation area, and the processed image with uniform pixel value distribution and enhanced water-sensitive response is output. By fusing the features of the variation area, the pixel value mutation is eliminated, the water-sensitive signal is strengthened, and the processed image is more consistent with the real distribution of soil moisture, providing high-quality image data for subsequent inversion. (6) Based on the pixel reflectivity of the processed image, a linear inversion model is used to calculate the soil moisture content, and a deviation threshold is used to control the output distribution map in combination with historical verification data. Through deviation checking, the accuracy of the inversion result is ensured, and the output soil moisture content distribution map can accurately reflect the spatial distribution of water, providing reliable data support for agricultural irrigation decision-making and environmental management, and enhancing the application value of the monitoring result. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 is a flowchart of a method for accurately monitoring soil moisture content by unmanned aerial vehicle remote sensing according to an embodiment of the present application; Figure 2 is a structural schematic diagram of a system for accurately monitoring soil moisture content by unmanned aerial vehicle remote sensing according to an embodiment of the present application. DETAILED DESCRIPTION

[0016] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0017] With reference to Figure 1 , the present application provides a method for accurately monitoring soil moisture content by unmanned aerial vehicle remote sensing, comprising the following steps: S11, acquiring initial multispectral image data sets containing near-infrared and short-wave infrared bands by a mobile platform-mounted spectral sensor to collect original spectral data of different physicochemical properties of soil types and different environmental conditions; S12, according to the pixel reflectance characteristics of the initial multispectral image data set, using a support vector machine classification method to divide the soil types of the pixel area of the land, marking the area where the clay component ratio exceeds the preset ratio threshold as the first group, and marking the remaining area as the second group, to form a soil type grouping set; S13, for each group in the soil type grouping set, respectively calculating the reflectance variance of the multispectral bands, and selecting the multispectral bands with reflectance variance exceeding the preset variance threshold as candidate bands to construct a corresponding candidate band set; S14, using a preset random forest regression model to process the candidate band set, inputting environmental condition variables including illumination and vegetation coverage to dynamically adjust the band weight, and iteratively selecting the band combination covering the top three weight orders and the water-sensitive peak to form an optimized band combination sequence; S15, extracting the spectral features corresponding to the optimized waveband combination sequence from the initial multispectral image dataset as dataset features, and using a weighted average method to perform feature fusion on the soil variation area of the complex plot, and outputting a processed image with uniform pixel value distribution and enhanced water sensitivity response; S16, based on the pixel reflectivity of the processed image, using a linear inversion model to calculate the soil water content value, and if the deviation of the soil water content value from the historical verification data is less than a preset water content threshold, outputting a soil water content distribution map.

[0018] In step S11, the original spectral data of different physicochemical properties of soil types and different environmental conditions are collected by the spectral sensor carried by the mobile platform, and the initial multispectral image dataset containing near-infrared waveband and short-wave infrared waveband is obtained.

[0019] In one embodiment, a certain plain agricultural test field is selected as the monitoring area, which contains two types of soil with different physicochemical properties, i.e. clay (physicochemical properties: clay content 35%, organic matter content 2.2%) and sandy soil (physicochemical properties: clay content 12%, organic matter content 1.8%), and there are two different environmental conditions, i.e. sunny noon (light intensity 85000 lux, no cloud) and cloudy afternoon (light intensity 32000 lux, cloudy). A DJI M300 RTK unmanned aerial vehicle is used as the mobile platform, and a Sentera Double 4K multispectral sensor (spectral resolution 5nm) is carried. The sensor synchronously captures the spectral information of the near-infrared waveband (730nm, 790nm key water-sensitive sub-waveband) and the short-wave infrared waveband (1240nm, 1650nm key water-sensitive sub-waveband) at a flight height of 120m from the ground and a flight speed of 5m / s. A multispectral image is generated every 0.5 seconds, and after removing the invalid images with edge blur and abnormal light, an initial multispectral image dataset containing near-infrared and short-wave infrared waveband and covering the entire monitoring area is obtained. The pixel resolution of the dataset reaches 10cm / pixel, and the spectral reflection difference under different soil types and environmental conditions can be clearly reflected.

[0020] In step S12, according to the pixel reflection characteristics of the initial multispectral image dataset, the soil type of the plot pixel area is divided by using the support vector machine classification method, the area with clay content exceeding a preset proportion threshold is marked as the first group, and the remaining area is marked as the second group, forming a soil type grouping set.

[0021] In an embodiment, the initial multispectral image dataset of the above-mentioned plain agricultural test field is taken, the reflection characteristics (including reflectance values of near-infrared 730 nm, 790 nm and short-wave infrared 1240 nm, 1650 nm bands) of each pixel in the dataset are extracted, and input into a pre-trained support vector machine classification model - the model has been trained by 500 groups of sample spectral data covering test field clay (clay content 30%-40%), sandy loam soil (clay content 10%-15%) and other soil types, and has built-in spectral response discrimination rules of different soil types in target bands, which can infer the proportion of soil composition through reflectance difference. Set the clay component proportion of 25% as the preset proportion threshold, and the model will identify the pixel area with clay component proportion exceeding 25% (concentrated in the northeast of the test field, accounting for about 35% of the total area) as the clay dominant type area and mark it as the first group; the pixel area with clay component proportion ≤25% (mainly sandy loam soil, distributed in the southwest of the test field, accounting for about 65% of the total area) is marked as the second group. Then integrate the spatial distribution information such as pixel coordinate range and region boundary of the two groups to form a soil type grouping set that can clearly reflect the soil type zoning characteristics of the test field. In step S13, for each group in the soil type grouping set, the reflectance variance of the multispectral band is calculated respectively, and the multispectral band with reflectance variance exceeding the preset variance threshold is selected as the candidate band to construct the corresponding candidate band set.

[0022] In one embodiment, the soil type grouping set of the above plain agricultural test field is first determined to determine the coverage of the first type grouping (clay dominant area, northeast of the test field, pixel coordinate range X1-Y1 to X2-Y2) and the second type grouping (sandy loam area, southwest of the test field, pixel coordinate range X3-Y3 to X4-Y4), and the corresponding multispectral band data (including 450-900nm near-infrared, 900-1700nm short-wave infrared, a total of 12 sub-bands) of the two groups are extracted. For the first type grouping, the reflectivity variance of the 12 bands is calculated, and the near-infrared 730nm band variance is 0.028, the 790nm band variance is 0.032, the short-wave infrared 1240nm band variance is 0.025, and the 1650nm band variance is 0.029. The variances of the remaining bands are all less than 0.02. For the second type grouping, the near-infrared 730nm band variance is 0.031, the 790nm band variance is 0.027, the short-wave infrared 1240nm band variance is 0.033, and the 1650nm band variance is 0.026. The variances of the remaining bands are less than 0.02. The preset variance threshold is set to 0.02, and the bands with reflectivity variance greater than 0.02 in the two groups are included in the candidate range. After grouping and summarizing, the first type grouping candidate band set (730nm, 790nm, 1240nm, 1650nm) and the second type grouping candidate band set (730nm, 790nm, 1240nm, 1650nm) are constructed. The higher the reflectivity variance of the band, the stronger the sensitivity to the change of the corresponding soil type water.

[0023] In step S14, the candidate band set is processed by using a preset random forest regression model, environmental condition variables including light and vegetation coverage are input to dynamically adjust the band weight, and the band combination covering the water sensitive peak value in the top three weight order is iteratively screened to form an optimized band combination sequence.

[0024] In one embodiment, the candidate band set of the above-mentioned first type (clay dominant) and second type (sandy loam) grouping (both containing 730nm, 790nm, 1240nm, 1650nm bands) is input into the preset random forest regression model - the model has been trained by 200 groups of different soil types (clay, sandy loam, etc.), environmental conditions (light intensity 30000-90000lux, vegetation coverage ratio 0-40%) and soil moisture content correlation samples, and the model has built-in band sensitivity weight calculation rules. The environmental condition variables during monitoring are simultaneously imported: the light intensity is 85000lux and the vegetation coverage ratio is 15% when collecting in the first grouping area, and the light intensity is 32000lux and the vegetation coverage ratio is 20% when collecting in the second grouping area, as adjustment parameters. After the model operation, the weight values for the first grouping candidate band are generated: 790nm (0.32), 1240nm (0.28), 730nm (0.22), and 1650nm (0.18); and the weight values for the second grouping are generated: 1240nm (0.31), 730nm (0.29), 790nm (0.23), and 1650nm (0.17). The top three bands are selected as the preliminary candidate according to the weight from high to low, and it is found that the top three bands of the two groups all cover 730nm (near-infrared water-sensitive peak), 790nm (near-infrared water-sensitive peak), and 1240nm (short-wave infrared water-sensitive peak), which is determined as an effective optimized combination. Finally, the optimized band combination sequence is formed according to the soil type grouping: the first grouping (790nm, 1240nm, 730nm), and the second grouping (1240nm, 730nm, 790nm).

[0025] In step S15, the spectral features corresponding to the optimized band combination sequence are extracted from the initial multispectral image data set as data set features, and the weighted average method is used for feature fusion of the soil variation area of the complex plot, and the processed image with uniform pixel value distribution and enhanced water-sensitive response is output.

[0026] In one embodiment, the above-mentioned optimized waveband combination sequence is implemented, and for the first type (clay dominant) grouping, the target wavebands are 790 nm (near-infrared water-sensitive peak), 1240 nm (short-wave infrared water-sensitive peak), and 730 nm (near-infrared auxiliary sensitive waveband); for the second type (sandy soil) grouping, the target wavebands are 1240 nm, 730 nm, and 790 nm. The spectral features corresponding to the target wavebands of the two groupings are extracted from the initial multispectral image dataset, specifically including: the original reflectance values of each waveband (for example, the 790 nm waveband clay region reflectance mean value is 0.32, and the sandy soil region mean value is 0.28), the difference between wavebands (for example, the 1240 nm and 730 nm reflectance difference, the clay region mean value is 0.15, and the sandy soil region mean value is 0.12), and the ratio between wavebands (for example, the 790 nm and 1240 nm reflectance ratio, the clay region mean value is 1.12, and the sandy soil region mean value is 1.08), which are integrated to form a spectral feature dataset precisely matched with the optimized waveband combination. Based on the spatial distribution regularity of the spectral feature dataset, by comparing the spectral feature differences of adjacent pixels (when the difference value exceeds 0.08 or the ratio deviates from the mean value by more than 15%, it is determined that the feature is mutated), the soil type transition zone (clay and sandy soil are distributed alternately, and the feature mutation pixel ratio reaches 30%) in the middle of the test field and the micro-topographic fluctuation area (due to the terrain difference, the spectral reflectance fluctuates, and the pixel reflectance standard deviation in the same waveband exceeds 0.05) in the east are identified, which together constitute the soil variation area; through image coordinate positioning, the pixel coordinate range of the area is marked as X5 (longitude 118.23°E)-Y5 (latitude 32.56°N) to X6 (longitude 118.25°E)-Y6 (latitude 32.54°N), and the spectral feature data of all pixels in the area are extracted to form a spectral feature subset containing 2800 pixel points. For the marked soil variation area and the corresponding spectral feature subset, the soil composition of the area is analyzed by field sampling, and it is determined that the clay accounts for 60% and the sandy soil accounts for 40%. Accordingly, the spectral feature weight values are set: the spectral features corresponding to the clay type (such as 790 nm reflectance 0.32 and 1240 nm and 730 nm difference 0.15) are weighted as 0.6, and the spectral features corresponding to the sandy soil type (such as 790 nm reflectance 0.28 and 1240 nm and 730 nm difference 0.12) are weighted as 0.4. The weighted average algorithm is used to perform fusion operation on the spectral feature subset pixel by pixel, for example, the original clay feature reflectance of a certain pixel point is 0.33, the sandy soil feature reflectance is 0.27, and the fused reflectance = 0.33x0.6 + 0.27x0.4 = 0.306. In this way, the unified feature value of all pixels covering the soil variation area is generated, and the feature mutation is eliminated. The unified feature value is embedded in the original spectral feature data set according to the corresponding pixel coordinate range (X5-Y5 to X6-Y6), replacing the discrete feature values of the original variation region, to form a fused complete image data; uniformity verification is performed on the fused image data, a sliding window method (window size 5x5 pixels) is used to calculate the gradient change rate of the pixel values in each window (gradient change rate = difference between the maximum and minimum pixel values in the window / average of the pixel values in the window), and the average gradient change rate is counted after traversing the entire image. The uniformity preset threshold is set to 0.05, the average gradient change rate of the fused image is 0.03 after calculation, which is less than the threshold, and it is determined that the verification is passed, and finally the processed image with continuous pixel value distribution, no obvious mutation and 18% improvement in moisture sensitive band response intensity is output.

[0027] In step S16, based on the pixel reflectivity of the processed image, the soil moisture content value is calculated by using a linear inversion model, and if the deviation of the soil moisture content value from the historical verification data is less than a preset moisture content threshold, a soil moisture content distribution map is output.

[0028] In one embodiment, based on the above-mentioned processed image (continuous pixel value distribution, 18% enhancement of moisture sensitive band response), the reflectivity data of all pixel regions in the image is first extracted - this data is enhanced spectral reflectivity information after feature fusion, specifically including the reflectivity of the first type (clay dominated) grouping target waveband 790nm, 1240nm, 730nm, and the reflectivity of the second type (sandy soil) grouping target waveband 1240nm, 730nm, 790nm. These reflectivity data are input into a pre-trained linear inversion model, which is trained by experimental field soil reflectivity and measured moisture content correlation samples (300 groups in total, covering clay, sandy soil and different moisture content intervals) for nearly 3 years, and the model has built-in reflectivity-moisture content corresponding relationship formulas for different soil types: The formula for the clay region is: where θ1 is the soil moisture content of the clay region (%), R 790 , R 1240 , and R 730 are the reflectivities at 790nm, 1240nm, and 730nm, respectively, and the model parameters a1=-0.25, a2=-0.31, a3=-0.22, and b1=0.45 The formula for the sandy soil region is: where θ2 is the soil moisture content of the sandy soil region (%), and the model parameters c1=-0.28, c2=-0.29, c3=-0.24, and d1=0.42 The soil water content calculation value of each pixel region is obtained by the above formula operation (for example, the calculation value of the partial pixels in the clay region is 18.5%, and the calculation value of the partial pixels in the sandy loam region is 15.2%).

[0029] Subsequently, the historical verification data of the test field is called, which includes the actually measured soil water content of 100 sampling points in the past 5 monitoring times (for example, the actually measured value at coordinates X5-Y5 in March 2024 is 18.2%, and the actually measured value at coordinates X6-Y6 is 15.0%), the coordinates of the corresponding sampling points, and the synchronous recorded light intensity and vegetation coverage. The soil water content calculation value of each pixel is compared with the historical verification data of the same coordinates, and the deviation degree is calculated (for example, the deviation at coordinates X5-Y5 is =|18.5%-18.2%|=0.3%, and the deviation at coordinates X6-Y6 is =|15.2%-15.0%|=0.2%). The preset deviation threshold is set to 1.0%, and the statistical deviation degree of 98% of the pixels is less than the threshold, and it is determined that the inversion result is effective. Finally, the effective calculation result is visualized according to the spatial distribution of the pixel region, and a gradient color spectrum from blue (low soil water content, ≤12%) to red (high soil water content, ≥20%) is used for labeling, and a soil water content distribution map is generated, which can clearly reflect that the soil water content in the clay region of the test field is higher than that in the sandy loam region, and the soil water content gradient in the middle transition zone is gentle.

[0030] It is worth noting that after outputting the soil water content distribution map, the following is also included: Obtain the regional water variation data of the soil water content distribution map, group according to the water gradient continuity using a clustering algorithm, mark the stable region and the region to be optimized, form an extended environment adaptive grouping set, calculate the band sensitivity index of the spectral reflectance dynamic change based on the environment adaptive grouping set, select the bands with sensitivity exceeding the preset threshold to construct an optimized enhanced band set, perform secondary feature extraction on the processed image through the enhanced band set and adaptive fusion on the region to be optimized, synthesize the enhanced image with uniform distribution after fusion and the original distribution map to obtain the final soil water content distribution map.

[0031] In one embodiment, the generated test field soil water content distribution map (clearly indicating the difference in soil water content between the clay and sandy loam regions, and the gentle gradient in the middle transition zone) is obtained, the regional water variation data of the distribution map is obtained, the regions are divided into stable and optimized regions according to the water gradient continuity through a density clustering algorithm, an extended environment adaptive grouping set is formed, the band sensitivity index is calculated based on the set, the bands exceeding the threshold are selected to construct an optimized enhanced band set, the features of the processed image are extracted again and the optimized region is fused, and finally the enhanced image and the original distribution map are synthesized to obtain the final soil water content distribution map. Further explanation: Extract the water content values of all pixels (such as clay area 16.8%-19.2%, sandy loam area 13.5%-16.0%) and spatial coordinates from the water content distribution map, calculate the difference of adjacent pixel water content (the difference of stable area is usually less than or equal to 0.5%, and the difference of local area can reach 1.2%), and integrate to form a regional water variation data set containing water content absolute value, spatial variation gradient and location correlation. Combined with the soil hydraulic conductivity characteristics of the test field (clay conductivity 0.15 cm / h, sandy loam 0.32 cm / h), set the distance judgment standard of density clustering algorithm as water content difference ≤0.8%, and the minimum aggregation unit as 8×8 pixels. According to the continuity of water content gradient, the pixels are aggregated and grouped. Calculate the dispersion degree (standard deviation ≤0.6%) and spatial uniformity (gradient change rate ≤0.04) of each group after clustering, and mark the standard deviation 0.3%-0.5% and the gradient change rate 0.02%-0.03% of the clay core area and the sandy loam core area as stable areas; mark the standard deviation 0.7%-0.9% and the gradient change rate 0.05%-0.06% of the eastern micro-terrain area and the transition zone edge as areas needing optimization. Finally, correlate the corresponding soil type grouping information, supplement the eastern terrain relief (height difference 3-5m) and vegetation coverage (18%-22%) as environmental tags, and form an expanded environmental adaptation grouping set. Further explanation: Based on the above set, calculate the band sensitivity index of spectral reflectance dynamic change, select the 790nm, 1240nm, 1380nm bands with sensitivity greater than 0.85, and construct an optimized enhanced band set. Perform secondary spectral feature extraction on the image processed in the previous section (water-sensitive response enhancement 18%), and obtain fine spectral feature data (such as 1380nm band reflectivity fluctuation reduced by 12% compared with the original processed image) of the optimization area (eastern micro-terrain area, transition zone edge) in these three bands.

[0032] Call the adaptive fusion algorithm, dynamically adjust the fusion weight according to the soil type (eastern area contains a small amount of clay and sandy loam mixture) and terrain characteristics (height difference 3-5m) of the area needing optimization, and use the formula: Where W i is the fusion weight of the i-th pixel, S i is the soil type proportion coefficient (clay proportion 55%×0.6 + sandy loam proportion 45%×0.4=0.51), T i is the terrain correction coefficient (0.95 due to height difference 3-5m), V i is the vegetation influence coefficient (1.0 for 20% vegetation coverage), weight coefficients c1=0.6, c2=0.2, c3=0.2, and W i= 0.6 * 0.51 + 0.2 * 0.95 + 0.2 * 1.0 = 0.306 + 0.19 + 0.2 = 0.696. Through this dynamic weight, the spectral feature mutation is eliminated, the fusion effect of the clay feature weight 0.55, the sandy loam feature weight 0.45, and the additional 0.05 smoothing weight of the terrain undulation is realized.

[0033] The uniformity check is performed on the fused image data, the pixel value gradient change rule is analyzed (the gradient change rate should be less than or equal to 0.04, which meets the characteristics of the soil moisture changing gently with the terrain), and the enhanced image is marked after the check is passed. The enhanced image and the initial water content distribution map are aligned in space coordinates, and the image synthesis algorithm is used to integrate the information: the original water content data of the stable area (clay and sandy loam core area) is retained, and the area to be optimized is replaced by the fused data, and finally the final soil water content distribution map is generated - the water content gradient of the area to be optimized in the map is smoother (the difference in the eastern micro-terrain area is reduced to within 0.6%), the overall data accuracy is improved by 15% compared with the original distribution map, and the real distribution of the soil moisture in the test field can be more accurately reflected.

[0034] In summary, the unmanned aerial vehicle remote sensing soil moisture inversion method of the present application forms a logical closed loop with the logic of "data collection - soil grouping - band optimization - feature fusion - accuracy verification": first, the technical process is progressive, near-infrared and short-wave infrared water-sensitive data are collected first, then the soil is grouped by support vector machine, the bands are optimized by variance screening and random forest, the optimized image of the variation area is fused, and finally the accuracy is verified to form a complete technical chain. Second, accurately solve the pain points, use variance screening and dynamic weight to improve the efficiency of multispectral data, group processing to adapt to different soils, and rely on fusion and verification to control monitoring errors. Third, it meets the actual needs, not only retains the advantages of high efficiency and flexibility of unmanned aerial vehicles, but also improves the reliability of the data, and the output water content distribution map can support agricultural irrigation and environmental management, providing a feasible path for complex land monitoring, and has systematicness and practicality.

[0035] Reference Figure 2 The second embodiment of the application provides an unmanned aerial vehicle remote sensing soil moisture inversion precision monitoring system, comprising: A data collection module: a spectrum sensor carried by a mobile platform is used to collect original spectrum data of different physicochemical properties of soil types and different environmental conditions, and to obtain an initial multispectral image data set containing near-infrared and short-wave infrared bands; A soil grouping module: according to the pixel reflection characteristics of the initial multispectral image data set, a support vector machine classification method is used to divide the soil types of the pixel area of the land, and the area where the clay component accounts for more than a preset proportion threshold is marked as a first grouping, and the remaining area is marked as a second grouping, to form a soil type grouping set; The candidate wave band screening module: for each group in the soil type group set, the reflectivity variance of the multispectral wave band is calculated respectively, and the multispectral wave band with a reflectivity variance exceeding a preset variance threshold is selected as a candidate wave band to construct a corresponding candidate wave band set; The wave band optimization module: a preset random forest regression model is used to perform data processing on the candidate wave band set, environmental condition variables including illumination and vegetation coverage are input to dynamically adjust the wave band weight, and a wave band combination covering the water sensitive peak value and ranking the top three in the weight order is iteratively screened to form an optimized wave band combination sequence; The weighted fusion module: the spectral features corresponding to the optimized wave band combination sequence are extracted from the initial multispectral image data set as a data set feature, and the weighted average method is used for feature fusion of the soil variation area of the complex plot to output a processed image with uniform pixel value distribution and enhanced water sensitive response; The image output module: based on the pixel reflectivity of the processed image, a linear inversion model is used to calculate the soil water content value, and if the deviation of the soil water content value from the historical verification data is less than a preset water content threshold, a soil water content distribution map is output. The spectral features include wave band reflectivity, spectral feature parameters, and wave band difference and ratio, and the historical verification data include soil water content measured values measured by field sampling in past monitoring of the target plot, corresponding sampling point coordinates, and synchronous recorded illumination intensity and vegetation coverage.

[0036] It should be noted that the unmanned aerial vehicle remote sensing soil water content inversion precision monitoring system provided by the embodiments of the present application is used to execute all process steps of the unmanned aerial vehicle remote sensing soil water content inversion precision monitoring method of the above-mentioned embodiments, and the working principles and beneficial effects of the two are one-to-one corresponding, so they will not be repeated.

[0037] The above-described specific embodiments further illustrate the purpose, technical solutions and beneficial effects of the present application. It should be understood that the above-described specific embodiments are only examples of the present application and do not limit the protection scope of the present application. It is particularly pointed out that any modification, equivalent replacement, improvement, etc. made by those skilled in the art within the spirit and principles of the present application should be included in the protection scope of the present application.

Claims

1. A method for precision monitoring of soil moisture content by unmanned aerial vehicle remote sensing inversion, characterized in that, The method comprises the following steps: acquiring initial multispectral image data sets containing near-infrared and short-wave infrared bands by collecting original spectral data of different soil types and different environmental conditions through a spectral sensor mounted on a mobile platform; dividing the soil types of the pixel regions of the field according to the pixel reflection characteristics of the initial multispectral image data sets by using a support vector machine classification method, marking the regions where the proportion of clay composition exceeds a preset proportion threshold as a first group, and marking the remaining regions as a second group, to form a soil type grouping set; calculating the reflectivity variance of the multispectral bands for each group in the soil type grouping set, and selecting the multispectral bands whose reflectivity variance exceeds a preset variance threshold as candidate bands to construct a corresponding candidate band set; processing the candidate band set by using a preset random forest regression model, inputting environmental condition variables including illumination and vegetation coverage to dynamically adjust the band weight, iteratively screening the band combination that ranks the top three in the weight order and covers the water-sensitive peak value, and forming an optimized band combination sequence; extracting the spectral features corresponding to the optimized band combination sequence from the initial multispectral image data set as the data set features, and using a weighted average method to fuse the features of the soil variation regions of the complex field, and outputting a processed image with uniform pixel value distribution and enhanced water-sensitive response; calculating the soil water content value based on the pixel reflectivity of the processed image by using a linear inversion model, and outputting a soil water content distribution map if the deviation of the soil water content value from the historical verification data is less than a preset water content threshold. The spectral features include band reflectivity, spectral feature parameters, and band-to-band difference and ratio, and the historical verification data include soil water content measured values measured by field sampling in past monitoring of the target field, corresponding sampling point coordinates, and synchronous recorded illumination intensity and vegetation coverage. 2.The unmanned aerial vehicle remote sensing inversion soil moisture content precision monitoring method according to claim 1, characterized in that, The method according to the pixel reflection characteristics of the initial multispectral image data set, using a support vector machine classification method to divide the soil types of the pixel regions of the field, marking the regions where the proportion of clay composition exceeds a preset proportion threshold as a first group, and marking the remaining regions as a second group to form a soil type grouping set, comprises the following steps: inputting the pixel reflection characteristics into a pre-trained support vector machine classification model to determine the proportion of soil composition of each pixel region through operation analysis; identifying the pixel regions where the proportion of clay composition exceeds a preset proportion threshold as clay-dominant type regions according to the proportion of soil composition; marking the regions identified as clay-dominant type as a first group, and marking the regions that do not exceed the preset proportion threshold as a second group; integrating the spatial distribution information of all regions marked as a first group and a second group to construct a soil type grouping set reflecting the soil type zoning characteristics of the target field; The support vector machine classification model has been trained by using sample spectral data covering clay and other soil types, and has built-in spectral response discrimination rules for different soil types. 3.The unmanned aerial vehicle remote sensing inversion soil moisture content precision monitoring method according to claim 1, characterized in that, The reflectance variance of the multi-spectral band is calculated for each group in the soil type grouping set respectively, and the multi-spectral band with reflectance variance exceeding a preset variance threshold is selected as a candidate band to construct a corresponding candidate band set, including: According to each group in the soil type grouping set, determine the pixel area range covered by each group and the corresponding multi-spectral band; For each group containing all pixel regions, respectively calculate the reflectance variance of the corresponding group on each multi-spectral band; Compare the reflectance variance of each multi-spectral band with the preset variance threshold, select the multi-spectral band exceeding the preset variance threshold and include the candidate range; Based on the candidate range, the candidate range band is respectively summarized and filtered according to the soil type grouping to form a candidate band set corresponding to each group; Wherein, the reflectance variance is used to represent the sensitivity of the band to the moisture change of the soil type. 4.The unmanned aerial vehicle remote sensing inversion soil moisture content precision monitoring method according to claim 1, characterized in that, The preset random forest regression model is used to process the candidate band set, the environmental condition variable including illumination and vegetation coverage is input to dynamically adjust the band weight, the band combination covering the water sensitive peak value is iteratively selected and sorted in the top three, and an optimized band combination sequence is formed, including: The band data in the candidate band set is input into the random forest regression model, and the environmental condition variable including illumination intensity and vegetation coverage is imported as an adjustment parameter; Through operation and analysis, the sensitive contribution of each band to soil moisture under the current environmental conditions is dynamically generated and adjusted to generate and adjust the weight value of each band; According to the weight value, the bands are sorted, a preset proportion of bands are selected as preliminary candidate bands from high to low, and it is checked whether the top three bands in the preliminary candidate bands cover the water sensitive peak value; If yes, the preliminary candidate band is determined as an effective optimization combination; If not, the environmental condition variable weight influence factor is readjusted, and the band weight calculation and sorting selection are iteratively performed; The final determined effective optimization combination is arranged according to the soil type grouping to form the optimized band combination sequence; Wherein, the random forest regression model is trained by associating the band data and water content samples under different soil types and environmental conditions, and the weight calculation rule of the band to the water sensitive degree is built in. 5.The unmanned aerial vehicle remote sensing inversion soil moisture content precision monitoring method according to claim 1, characterized in that, The spectral features corresponding to the optimized band combination sequence are extracted from the initial multi-spectral image data set as the data set features, and the weighted average method is used for feature fusion in the soil variation area of the complex plot to output the processed image with uniform pixel value distribution and enhanced water sensitive response, including: According to the optimized band combination sequence, determine the target band, and extract the spectral features corresponding to the target band from the initial multi-spectral image data set to form a spectral feature data set matched with the optimized band combination; Based on the spatial distribution of the spectral feature data set, identify the soil variation area composed of soil type transition zone and micro-topographic change area, and mark the pixel coordinate range and the corresponding spectral feature subset; The spectral feature weight value is set according to the proportion of different soil types at different positions for the marked soil variation area and the corresponding spectral feature subset; Based on the spectral feature weight value, a weighted average algorithm is used to perform fusion operation on the spectral feature subset to generate a unified feature value covering all pixels of the soil variation area; The unified feature value is embedded in the spectral feature data set according to the pixel coordinate range to form a fused image data; Uniformity verification is performed on the fused image data, the gradient change rate of pixel value distribution is calculated, and when the gradient change rate is less than a preset threshold, it is determined that the verification is passed, and the verified image data is output as a processed image; The target waveband is a multi-spectral waveband with high sensitivity to soil moisture change after weight screening, and covers the water-sensitive peak value of near-infrared and short-wave infrared, which reflects the difference in soil moisture content. The spectral feature subset refers to the spectral feature data extracted from the spectral feature data set and completely corresponding to the pixel coordinate range of the marked soil variation area. 6.The unmanned aerial vehicle remote sensing inversion soil moisture content precision monitoring method according to claim 1, characterized in that, Based on the pixel reflectivity of the processed image, a linear inversion model is used to calculate the soil moisture content value, and if the deviation of the soil moisture content value from the historical verification data is less than a preset moisture content threshold, a soil moisture content distribution map is output, including: Extract the reflectivity data of each pixel region in the processed image and input it into the linear inversion model to obtain the corresponding soil moisture content calculation value of each pixel region through calculation; The historical verification data of the target plot is called, the soil moisture content calculation value is compared with the historical verification data of the same region, and the deviation between the two is calculated; When the deviation is less than a preset deviation threshold, the effective calculation result is visualized according to the spatial distribution of the pixel region to generate a soil moisture content distribution map reflecting the difference in soil moisture content in the plot; The reflectivity data is spectral reflectivity information enhanced by feature fusion and sensitive to water response, the linear inversion model is trained by correlating samples of soil reflectivity and measured moisture content, and the historical verification data includes measured soil moisture content and corresponding spectral reflectivity features at different positions. 7.The unmanned aerial vehicle remote sensing inversion soil moisture content precision monitoring method according to claim 1, characterized in that, After outputting the soil moisture content distribution map, it further includes: Obtain the regional moisture variation data of the soil moisture content distribution map, group them according to the continuity of moisture gradient using clustering algorithm, mark the stable area and the area to be optimized, and form an extended environment adaptive grouping set; Based on the environment adaptive grouping set, calculate the waveband sensitivity index of spectral reflectivity dynamic change, select the wavebands with sensitivity exceeding a preset sensitivity threshold to construct an optimized enhanced waveband set; Through the enhanced waveband set, the processed image is extracted again, and adaptive fusion is performed on the area to be optimized to synthesize the enhanced image with uniform distribution after fusion and the original distribution map to obtain the final soil moisture content distribution map. 8.The unmanned aerial vehicle remote sensing inversion soil moisture content precision monitoring method according to claim 1, characterized in that, The acquisition of the regional moisture variation data of the soil moisture content distribution map, the grouping according to the continuity of moisture gradient using clustering algorithm, the marking of the stable area and the area to be optimized, and the formation of an extended environment adaptive grouping set, include: Extracting the moisture content value and spatial position information of each pixel from the soil moisture content distribution map, and integrating the moisture content absolute value, spatial variation gradient and position correlation to form a regional moisture variation data set by calculating the moisture content difference between adjacent pixels; Based on the regional moisture variation data set, the distance determination standard and the minimum aggregation unit of the clustering algorithm are set according to the soil hydraulic conductivity characteristics of the land, and the density clustering algorithm is used to perform aggregation grouping on adjacent pixels according to the natural continuity of moisture content gradient; For each group formed by clustering, calculate the moisture content dispersion degree and spatial distribution uniformity, and mark as a stable area when both meet the preset standard, and mark as an optimization area when any one does not meet the preset standard; The labeled stable area and optimization area are associated with the corresponding soil type grouping information, and the terrain undulation and vegetation coverage are supplemented as environmental tags to form an environmental adaptation grouping set containing regional boundaries, moisture content statistical characteristics and environmental tags. 9.The unmanned aerial vehicle remote sensing inversion soil moisture content precision monitoring method according to claim 1, characterized in that, The processed image is extracted again based on the enhanced band set, and adaptive fusion is performed on the optimization area to obtain the final soil moisture content distribution map, including: Based on the optimized enhanced band set, perform secondary spectral feature extraction on the processed image, and directionally obtain fine spectral feature data of the optimization area under the enhanced band; Based on the fine spectral feature data, call the adaptive fusion algorithm, dynamically adjust the fusion weight according to the soil type and terrain characteristics of the optimization area, and complete feature fusion; Perform distribution uniformity verification on the fused image data, match the natural distribution characteristics of soil moisture by analyzing the pixel value gradient change rule, and mark the image data that meets the distribution rule as an enhanced image; Align the marked enhanced image with the initial output soil moisture content distribution map in spatial coordinates, retain the original data of the stable area and the fused data of the optimization area, and generate the final soil moisture content distribution map.

10. An unmanned aerial vehicle remote sensing inversion soil moisture content precision monitoring system, characterized in that, It includes: Data acquisition module: through the spectral sensor carried by the mobile platform, the original spectral data of different physicochemical properties of soil types and different environmental conditions are collected, and the initial multispectral image data set containing near-infrared band and short-wave infrared band is obtained; Soil grouping module: according to the pixel reflection characteristics of the initial multispectral image data set, the soil type of the land pixel area is divided by using support vector machine classification method, the area with clay content exceeding the preset proportion threshold is marked as the first grouping, and the remaining area is marked as the second grouping, forming a soil type grouping set; Candidate band screening module: for each grouping in the soil type grouping set, the reflectivity variance of the multispectral band is calculated respectively, and the multispectral band with reflectivity variance exceeding the preset variance threshold is selected as the candidate band to construct the corresponding candidate band set; The band optimization module: a preset random forest regression model is used to process the candidate band set, environmental condition variables including illumination and vegetation coverage are input to dynamically adjust the band weight, and the band combination covering the water-sensitive peak value and ranking the top three in the weight order is iteratively screened to form an optimized band combination sequence; The weighted fusion module: the spectral features corresponding to the optimized band combination sequence are extracted from the initial multispectral image dataset as dataset features, and the weighted average method is used for feature fusion of the soil variation area of the complex plot to output a processed image with uniform pixel value distribution and enhanced water-sensitive response; The image output module: based on the pixel reflectivity of the processed image, a linear inversion model is used to calculate the soil moisture content value, and if the deviation of the soil moisture content value from the historical verification data is less than a preset moisture content threshold, a soil moisture content distribution map is output. The spectral features include band reflectivity, spectral feature parameters, and band difference and ratio, and the historical verification data include soil moisture content measured values measured by field sampling in past monitoring of the target plot, corresponding sampling point coordinates, and synchronous recorded illumination intensity and vegetation coverage.

Images