Irrigation water requirement prediction method based on unmanned aerial vehicle remote sensing in irrigation district

Abstract

The application relates to the technical field of irrigation water demand prediction, and discloses an irrigation water demand prediction method based on unmanned aerial vehicle remote sensing, which comprises the following steps: step S1: data acquisition; step S2: cluster optimization calculation; and step S3: irrigation water demand prediction based on the cluster result. An optimization factor is constructed, different features are given different weights to reflect their different indicating effects on the irrigation water demand, and the correlation between the spectral features is considered. Specifically, the correlation between the spectral features is used to determine the weight of each wave band, and the wave band with higher correlation will be given a higher weight; and a higher weight is also given to the spectral feature due to its close relationship with the crop growth condition and the water condition. The optimization factor constructed in this way can more accurately reflect the importance of different features in the irrigation water demand prediction, and the application accuracy of the K-means clustering algorithm in the scene is improved.

Description

Technical Field

[0001] This invention belongs to the field of irrigation water demand prediction technology, specifically, it relates to a method for predicting irrigation water demand in irrigation districts based on UAV remote sensing. Background Technology

[0002] Precision agriculture is an inevitable trend in modern agricultural development. Its core lies in the meticulous management of agricultural production processes based on crop growth and environmental conditions, thereby improving crop yield and quality while reducing resource waste and environmental pollution. Irrigation is one of the most crucial aspects of agricultural production, and accurately predicting irrigation water demand is essential for achieving precision irrigation. Unmanned aerial vehicle (UAV) remote sensing technology, with its advantages of speed, flexibility, and high resolution, has been widely applied in agriculture. By carrying multispectral or hyperspectral cameras, UAVs can acquire high-resolution images of farmland areas, extracting information such as the spectral and textural characteristics of crops. Combined with ground observation data, crop growth models and irrigation water demand prediction models can be established.

[0003] K-means clustering is a commonly used unsupervised machine learning algorithm that divides data into different categories based on their inherent similarity. In agricultural applications, clustering is used to identify different crop types, crop growth stages, soil types, etc. Clustering is also used for irrigation water demand prediction, dividing farmland areas into different irrigation zones based on the spectral characteristics of crops.

[0004] In the scenario of predicting irrigation water demand in irrigation districts based on UAV remote sensing, the traditional K-means clustering algorithm based on Euclidean distance fails to consider the differences in the indicative significance of different features for water demand and the correlation between features. This results in low clustering accuracy when using multidimensional remote sensing data to divide irrigation water demand status areas, making it difficult to accurately reflect the actual irrigation needs of different areas and thus affecting the formulation of precision irrigation strategies. Specifically, Euclidean distance treats all features equally, ignoring the varying degrees of indicative significance of different spectral bands, vegetation indices, and environmental factors for irrigation water demand. For example, NDVI is usually more indicative of crop water status than the simple visible light band. At the same time, Euclidean distance treats each feature dimension as independent, ignoring the potential correlations between spectral bands, vegetation indices, and environmental factors. For example, there is a strong correlation between the reflectance of the near-infrared band and the red light band, and this correlation is closely related to crop growth status and the degree of water stress. Ignoring these factors means that Euclidean distance cannot accurately capture the complex information contained in the data, reducing the reliability and practicality of the clustering results.

[0005] In view of this, the present invention is proposed. Summary of the Invention

[0006] To solve the above-mentioned technical problems, the basic concept of the technical solution adopted by the present invention is as follows:

[0007] A method for predicting irrigation water demand in irrigation districts based on UAV remote sensing includes the following steps:

[0008] Step S1: Data Acquisition. Using a drone equipped with a multispectral or hyperspectral camera, aerial photography of the target irrigation area is conducted under clear weather and good lighting conditions to acquire high-resolution image data, and GPS information is recorded simultaneously. Data Preprocessing. Radiometric and geometric corrections are performed on the drone images to eliminate errors caused by atmospheric, lighting, and drone attitude factors, and the images are registered to a unified geographic coordinate system. Next, the Normalized Difference Vegetation Index (NDVI) is calculated. The preprocessed spectral data, including red band reflectance (R), green band reflectance (G), blue band reflectance (B), near-infrared band reflectance (NIR), and NDVI, are fused to construct a dataset: each pixel is a data point, and its attributes include spectral reflectance and vegetation index value, ultimately forming an N×(m+n) two-dimensional matrix, where N is the number of pixels, m is the number of spectral bands, and n is the number of vegetation indices. This matrix will serve as the input data for K-means clustering in subsequent steps.

[0009] Step S2: Clustering optimization calculation based on multi-feature indicative weighting and adaptive adjustment of spatial information, specifically including the following steps:

[0010] a. Based on the differences and correlations in indicative effects among features, an optimization factor α is constructed to perform preliminary optimization of the distance metric;

[0011] b. Based on the fusion of local spatial autocorrelation and global variation information, the optimization factor β is constructed, and the distance metric is optimized in a secondary manner;

[0012] c. Perform K-means clustering calculations using a quadratically optimized distance metric;

[0013] Step S3: Irrigation water demand prediction based on clustering results.

[0014] In a preferred embodiment of the present invention, the construction of the optimization factor in step S2 aims to optimize the distance metric based on the indicative differences in irrigation water demand of different features and the correlation between features; the specific steps are as follows:

[0015] First, all m+n features are divided into two groups: spectral features S and vegetation indices V; where spectral features S = (S1, S2, ..., S...). m The spectral reflectance includes m bands, namely the red band reflectance R, the green band reflectance G, the blue band reflectance B, and the near-infrared band reflectance NIR; the vegetation index V = (NDVI) contains only one normalized vegetation index NDVI.

[0016] In a preferred embodiment of the present invention, the indicative weights of each group of features are then calculated; for spectral features S, the indicative weights of each band S are calculated. i Pearson correlation coefficient with Normalized Difference Vegetation Index (NDVI)

[0017]

[0018] Among them, corr(S) i ,NDVI) represents the calculation of S i The correlation coefficient between NDVI and NDVI, i = 1, 2, ..., m; NDVI is an indicator of vegetation growth and water status; the calculated correlation coefficient... After normalization, the indicative weight vector r of the spectral features is obtained. s :

[0019]

[0020] Normalization ensures that the sum of the weights for all spectral bands is 1.

[0021] In a preferred embodiment of the present invention, the weight vector r of the spectral features needs to be weighted next. s The weight r of the vegetation index v Integrate into a unified weight vector; due to r s The length is 4, corresponding to four spectral bands, and r v Expand into a vector of length 1. Its only element is r v The value; that is Finally, calculate the optimization factor α; and set r s and Concatenate them into a vector and convert it into a diagonal matrix α:

[0022]

[0023] in, Indicates that r s and The vectors are concatenated and converted into a diagonal matrix. The optimization factor α is a diagonal matrix, with the elements on the diagonal representing the indicative weights of each feature. When calculating the distance, the optimization factor α is multiplied by the feature vectors to weight different features. Features with strong indicative power for irrigation water demand will be given greater weights, while features with weak indicative power will be given smaller weights.

[0024] As a preferred embodiment of the present invention, the construction of the optimization factor β mainly considers the local spatial correlation of the data and the global data distribution characteristics. The optimization factor β is constructed by calculating the spatial autocorrelation index of each data point in its local neighborhood and the global variation coefficient of each feature.

[0025] First, the local neighborhood is determined: for each data point x, its k nearest neighbors are selected to form a local neighborhood N. k (x); Local neighborhood N k (x) reflects the spatial information around data point x; the value of k needs to be selected by cross-validation based on the specific study area and data characteristics.

[0026] Next, we calculate the local spatial autocorrelation index: for the local neighborhood N k For each feature f within (x), calculate its local spatial autocorrelation index Moran's I. f (x); Moran's I index is a commonly used spatial autocorrelation index used to measure the similarity between a certain attribute value in a region and its surrounding regions; its calculation formula is:

[0027]

[0028] Where: n is the local neighborhood N k The number of pixels within (x); w pq These are elements of the spatial weight matrix, representing the spatial relationship between pixel p and pixel q; x p,f Indicates the value of pixel p on feature f; x q,f μ represents the value of pixel q on feature f; f This represents the global mean of feature f.

[0029] As a preferred embodiment of the present invention, Moran's I f The larger the value of (x), the more similar the data point x is to its surrounding data points on feature f, i.e., the stronger the local spatial correlation. Moran's I index measures the spatial autocorrelation of a feature in a region by calculating the product of the differences between each data point and its surrounding data points on a certain feature. The global coefficient of variation (CV) is calculated for each feature f. f CV is used to measure the dispersion of feature values ​​across the entire dataset; f The calculation formula is:

[0030]

[0031] Where, σ f and μ fLet represent the global standard deviation and global mean of feature f, respectively. CV f The larger the value, the greater the global variation of feature f; finally, the optimization factor β(x) is calculated to obtain the final optimized distance metric: the optimization factor β(x) is a diagonal matrix, whose diagonal elements are determined by the local spatial autocorrelation index and global variation coefficient of each feature; the formula for calculating β(x) is:

[0032] β(x)=diag((Moran's I) local (x).^u).*(CV.^z));

[0033] Here, u and z are two adjustable parameters used to control the influence of the local spatial autocorrelation index and the global coefficient of variation, respectively. These two parameters are determined through a data-driven approach. The median or mean of the local spatial autocorrelation index of all data points is used as the u value, and the median or mean of the coefficient of variation of all features is used as the z value. .^ indicates element-wise exponentiation; .* indicates element-wise multiplication; diag() indicates converting a vector into a diagonal matrix.

[0034] Compared with the prior art, the present invention has the following advantages:

[0035] This invention constructs an optimization factor α, assigning different weights to different features to reflect their varying indicative roles in irrigation water demand, while also considering the correlation between spectral features. Specifically, the weight of each band is determined by the correlation between spectral features and NDVI; bands with higher correlation to NDVI are assigned higher weights. Furthermore, NDVI itself is given a higher weight due to its close relationship with crop growth and water conditions. The optimization factor α constructed in this way can more accurately reflect the importance of different features in irrigation water demand prediction, improving the accuracy of the K-means clustering algorithm in this scenario.

[0036] The specific embodiments of the present invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0037] In the attached diagram:

[0038] Figure 1 This is a flowchart of a method for predicting irrigation water demand in irrigation districts based on UAV remote sensing. Detailed Implementation

[0039] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments will be clearly and completely described below with reference to the accompanying drawings. The following embodiments are used to illustrate the present invention.

[0040] A method for predicting irrigation water demand in irrigation districts based on UAV remote sensing includes the following steps:

[0041] Step S1: Data acquisition and preprocessing.

[0042] Using drones equipped with multispectral or hyperspectral cameras, aerial photography of the target irrigation area was conducted under clear weather and good lighting conditions to acquire high-resolution image data, while simultaneously recording GPS information. Radiometric and geometric corrections were then performed on the drone images to eliminate errors caused by atmospheric conditions, lighting, and drone attitude, and the images were registered to a unified geographic coordinate system. The Normalized Difference Vegetation Index (NDVI) was then calculated.

[0043] Finally, the preprocessed spectral data (red band reflectance R, green band reflectance G, blue band reflectance B, and near-infrared band reflectance NIR) and vegetation index (NDVI) are fused to construct a dataset: each pixel is a data point, and its attributes include spectral reflectance and vegetation index value, ultimately forming an N×(m+n) two-dimensional matrix, where N is the number of pixels, m is the number of spectral bands (m=4 in this example), and n is the number of vegetation indices (n=1 in this example). This matrix will serve as the input data for K-means clustering in subsequent steps.

[0044] Step S2: K-means clustering optimization calculation based on multi-feature indicative weighting and spatial information adaptive adjustment.

[0045] a. Based on the differences and correlations in indicative effects among features, an optimization factor α is constructed to perform preliminary optimization of the distance metric.

[0046] Detailed logic:

[0047] In the scenario of irrigation water demand prediction based on UAV remote sensing, the traditional K-means clustering algorithm based on Euclidean distance has a significant problem: it treats all features participating in the clustering equally, ignoring the differences in the strength of these features' indicative role in irrigation water demand and the correlation between features. This indiscriminate treatment leads to clustering results that cannot accurately reflect the true differences in irrigation water demand across different regions. Specifically, UAV remote sensing data typically contains multiple spectral bands and one or more vegetation indices. These features have different roles in indicating crop irrigation water demand. For example, the Normalized Difference Vegetation Index (NDVI), calculated by combining the reflectance of the near-infrared (NIR) and red (R) bands, is generally considered a good indicator of crop growth and water status. A high NDVI value usually indicates vigorous vegetation growth, strong transpiration, and relatively high water demand; while a low NDVI value indicates poor vegetation growth or water stress, and relatively low water demand. In contrast, the reflectivity of the blue light band (B) is more affected by atmospheric scattering and its direct relationship with crop water requirements is not as strong as that of NDVI. However, traditional Euclidean distance metrics calculate the reflectivity differences between two pixels equally across the NIR, R, and B bands and simply sum these differences. This results in the reflectivity differences in the B band having the same impact on the final distance calculation as the NIR and R bands, even though the B band should actually be a weaker indicator of water requirements than the NIR and R bands.

[0048] Furthermore, traditional Euclidean distance ignores the correlations between features. For example, in multispectral data, there is often a strong negative correlation between the reflectance of the near-infrared and red bands, which forms the basis for NDVI's indication of vegetation growth status. In hyperspectral data, there is also often a strong positive correlation between the spectral reflectance of adjacent bands. These correlations contain information about crop growth status and water requirements. Traditional Euclidean distance treats each feature dimension as independent when calculating distance, ignoring these correlations and resulting in the loss of useful information.

[0049] In summary, because the traditional Euclidean distance metric treats all features equally, it ignores the differences in the indicative role of different features in irrigation water demand and the correlation between features. Therefore, when applied to irrigation water demand prediction scenarios based on UAV remote sensing, it leads to inaccurate clustering results and fails to effectively distinguish regions with different irrigation water demand states.

[0050] To address the aforementioned issues, an optimization factor α is constructed, assigning different weights to different features to reflect their varying indicative roles in irrigation water demand, while also considering the correlation between spectral features. Specifically, the weight of each band is determined by the correlation between spectral features and NDVI; bands with higher correlation to NDVI are assigned higher weights. Furthermore, NDVI itself is given a higher weight due to its close relationship with crop growth and water conditions. This optimization factor α, constructed in this way, can more accurately reflect the importance of different features in irrigation water demand prediction, improving the accuracy of the K-means clustering algorithm in this scenario.

[0051] Formula logic:

[0052] The optimization factor α was constructed to optimize the distance metric based on the indicative differences in irrigation water demand among different features and the correlation between features. The specific steps are as follows:

[0053] First, all m+n features are divided into two groups: spectral features S and vegetation indices V. Wherein, spectral features S = (S1, S2, ..., S...). m The spectral reflectance includes m bands, where m = 4 in this example, representing the reflectance R in the red band, the reflectance G in the green band, the reflectance B in the blue band, and the reflectance NIR in the near-infrared band; the vegetation index V = (NDVI) contains only one normalized vegetation index NDVI.

[0054] Then, the indicative weights for each group of features are calculated separately. For spectral feature S, the indicative weights for each band S are calculated. i Pearson correlation coefficient with Normalized Difference Vegetation Index (NDVI)

[0055]

[0056] Among them, corr(S) i ,NDVI) represents the calculation of S i The correlation coefficient between NDVI and NDVI, i = 1, 2, ..., m. NDVI is a widely recognized indicator of vegetation growth and water status. If a spectral band has a high correlation coefficient with NDVI, it indicates that the band is also closely related to vegetation growth and water status, thus suggesting that the band has a strong indicative role in irrigation water demand. The calculated correlation coefficients... After normalization, the indicative weight vector r of the spectral features is obtained. s :

[0057]

[0058] Normalization ensures that the sum of the weights for all spectral bands is 1. For the vegetation index V, since NDVI is closely related to crop growth and water conditions, and in this example there is only one vegetation index, NDVI, we set the weight r of NDVI to... v Set it to a fixed value. Based on experience, r... v Set as spectral feature weight vector r s The maximum value in the middle.

[0059] Next, we need to weight the spectral features into vector r. s The weight r of the vegetation index v Integrate them into a unified weight vector. Because r s The length is 4 (corresponding to four spectral bands), and we will... v Expand into a vector of length 1. Its only element is r v The value of . That is

[0060] Finally, calculate the optimization factor α. Let r s and Concatenate them into a vector and convert it into a diagonal matrix α:

[0061]

[0062] in, Indicates that r s and The vectors are concatenated and then converted into a diagonal matrix. The optimization factor α is a diagonal matrix whose diagonal elements represent the indicative weights of each feature. When calculating the distance, the optimization factor α is multiplied by the feature vectors to weight the different features. Features that are highly indicative of irrigation water demand will be assigned greater weights, while features that are less indicative will be assigned smaller weights.

[0063] b. Based on the fusion of local spatial autocorrelation and global variation information, the optimization factor β is constructed, and the distance metric is optimized in a secondary manner.

[0064] Detailed logic:

[0065] In step a, an optimization factor α was constructed to address the problem that traditional Euclidean distance treats all features equally, ignoring the differences in the indicative weights of different features on irrigation water demand and the correlation between features. However, even considering the indicative weights and correlations of features, a problem still exists: the spatial distribution characteristics of the data are ignored, especially the spatial correlation of local areas and the degree of variation of global data. This can lead to spatially discontinuous and unreasonable clustering results, such as the "salt and pepper" phenomenon.

[0066] Specifically, in the scenario of predicting irrigation water demand in irrigation districts, adjacent plots often have similar crop planting patterns and irrigation conditions, so their irrigation water demands are often quite similar, exhibiting strong spatial correlation. For example, in an area where the same crops are planted and irrigation conditions are consistent, the irrigation water demands of different plots within it are usually quite similar, showing strong spatial correlation. However, the traditional K-means algorithm only considers the distance between data points in the feature space during clustering, without considering the spatial positional relationship of the data points. This may lead to a "salt and pepper" phenomenon in the clustering results, where regions of the same category are divided into multiple scattered small patches, or spatially adjacent regions with slightly different features are incorrectly classified into different categories, resulting in poor spatial continuity of the clustering results.

[0067] Furthermore, the data distribution characteristics may differ across regions. For example, in some regions, crop growth is relatively uniform, with minimal variation in characteristics such as spectral reflectance and vegetation indices across different bands. However, in other regions, due to differences in field management practices or other factors, even crops of the same category may exhibit significant differences in growth, leading to greater variability in characteristics. Traditional K-means algorithms do not consider these differences in data distribution characteristics, potentially resulting in over-clustering in areas of high variability and under-clustering in areas of low variability.

[0068] For example, suppose there is a cornfield where, due to differences in field management practices, corn in one area grows well while corn in another area grows poorly. These two areas differ in spectral characteristics and vegetation indices, but because they are spatially adjacent and both fall under the "moderate irrigation requirement" category, they should be classified into the same group when zoning irrigation. However, if clustering is performed solely based on the optimization factor α, without considering spatial information, some scattered pixels with slightly different spectral characteristics in these two areas may be incorrectly classified into other categories. This results in a "salt and pepper" phenomenon in the clustering results, where the "moderate irrigation requirement" area is divided into multiple discontinuous small patches. This will affect the formulation and implementation of irrigation plans.

[0069] To address the aforementioned issues, an optimization factor β is constructed to further refine the distance metric. This factor reflects both the spatial correlation of data within local regions and the degree of global data variability. Specifically, for each data point, the spatial autocorrelation index of each feature is calculated within its local neighborhood to measure the spatial correlation of that feature within that local region. Simultaneously, the global coefficient of variation for each feature is calculated to measure its variability across the entire study area. By combining the local spatial autocorrelation index and the global coefficient of variation, the optimization factor β refines the distance metric, better reflecting the spatial distribution characteristics of the data. This improves the accuracy of the K-means clustering algorithm in predicting irrigation demand in irrigation districts and reduces the occurrence of the "salt and pepper" phenomenon.

[0070] Formula logic:

[0071] The construction of the optimization factor β mainly considers the local spatial correlation of the data and the global data distribution characteristics. The optimization factor β is constructed by calculating the spatial autocorrelation index of each data point in its local neighborhood and the global coefficient of variation of each feature.

[0072] First, the local neighborhood is determined: for each data point x, its k nearest neighbors are selected to form a local neighborhood N. k (x). This local neighborhood N k (x) reflects the spatial information surrounding data point x. The value of k needs to be selected through cross-validation based on the specific study area and data characteristics.

[0073] Next, we calculate the local spatial autocorrelation index: for the local neighborhood N k For each feature f within (x) (in this example, the features include four spectral bands: R, G, B, NIR, and a vegetation index: NDVI), calculate its local spatial autocorrelation index Moran's I. f (x). Moran's I index is a commonly used spatial autocorrelation index used to measure the similarity between a given attribute value within a region and its surrounding regions. Its calculation formula is:

[0074]

[0075] in:

[0076] n is the local neighborhood N k The number of pixels within (x).

[0077] w pq These are elements of the spatial weight matrix, representing the spatial relationship between pixel p and pixel q. In this method, if pixel p and pixel q are among each other's k nearest neighbors (i.e., q belongs to N...), then...k (x) and p belongs to N k(q) ), then w pq =1, otherwise w pq =0.

[0078] x p,f This represents the value of pixel p on feature f.

[0079] x q,f This represents the value of pixel q on feature f.

[0080] μ f This represents the global mean of feature f.

[0081] Moran's I f The larger the value of (x), the more similar the data point x is to its surrounding data points on feature f, indicating a stronger local spatial correlation. The Moran's I index measures the spatial autocorrelation of a feature in a region by calculating the product of the differences between each data point and its surrounding data points on that feature. If a data point is relatively close to its surrounding data points on that feature (i.e., the differences are small), then the sum of their products will be larger, and the Moran's I index will also be larger, indicating a strong spatial correlation in that region on that feature.

[0082] Then, the global coefficient of variation is calculated: the global coefficient of variation (CV) for each feature f is calculated. f This is used to measure the dispersion of feature values ​​across the entire dataset. CV f The calculation formula is:

[0083]

[0084] Where, σ f and μ f Let represent the global standard deviation and global mean of feature f, respectively. CV f The larger the value of , the greater the global variability of feature f. The coefficient of variation (CV) is the ratio of the standard deviation to the mean. It eliminates the influence of dimensions and can be used to compare the degree of variation between different features. The larger the CV value, the greater the variability of that feature, meaning the more dispersed the data distribution on that feature.

[0085] Finally, the optimization factor β(x) is calculated to obtain the final optimized distance metric: the optimization factor β(x) is a diagonal matrix, whose diagonal elements are determined by the local spatial autocorrelation index and global coefficient of variation of each feature. The formula for calculating β(x) is:

[0086] β(x)=diag((Moran's I) local (x).^u).*(CV.^z))

[0087] in:

[0088] Moran's I local (x)=(Moran's I R (x), Moran's I G (x), Moran's I B (x), Moran's I NIR (x), Moran's I NDVI (x) is the local spatial autocorrelation index vector, corresponding to the five features R, G, B, NIR and NDVI respectively.

[0089] CV = (CV) R ,CV G ,CV B ,CV N IR, CV N DVI is the coefficient of variation vector, which corresponds to the five features R, G, B, NIR and NDVI respectively.

[0090] u and z are two adjustable parameters used to control the influence of the local spatial autocorrelation index and the global coefficient of variation, respectively. These two parameters are determined through a data-driven approach. In this invention, the median or mean of the local spatial autocorrelation index of all data points is used as the u value, and the median or mean of the coefficient of variation of all features is used as the z value.

[0091] .^ represents exponentiation element by element.

[0092] * indicates element-wise multiplication.

[0093] diag() converts a vector into a diagonal matrix.

[0094] Each element on the diagonal of the optimization factor β(x) is the product of the u-th power of the local spatial autocorrelation index of the corresponding feature and the z-th power of the global coefficient of variation. This design allows β(x) to simultaneously consider both the local spatial correlation and the degree of global variation of the data. For features with strong local spatial correlation (i.e., Moran's I...), f (x) is larger), and its corresponding weight will increase; for features with a large degree of global variation (i.e., CV) f If the feature is larger, its corresponding weight will also increase. The purpose of this is to enhance the influence of spatially clustered and highly variable features on the clustering results.

[0095] After constructing the optimization factors α and β, they are applied to the distance metric of K-means clustering to obtain the optimized distance metric formula:

[0096]

[0097] in:

[0098] x and y are two data points, each containing spectral features and vegetation indices.

[0099] α is an optimization factor constructed based on feature indicative weights and correlations. It is a diagonal matrix whose diagonal elements reflect the indicative weights of different features on irrigation water demand and take into account the correlations between spectral features.

[0100] β(x) is an optimization factor constructed based on local spatial correlation and data distribution characteristics. It is also a diagonal matrix, and the elements on its diagonal reflect the local spatial correlation around each data point and the degree of global variation of each feature.

[0101] sqrt() represents the square root.

[0102] (xy) T This represents the transpose of vector (xy).

[0103] c. Perform K-means clustering calculations using a quadratically optimized distance metric.

[0104] Detailed logic:

[0105] After obtaining the new distance metric d optimized according to the above two steps αβ After (x, y), K-means clustering is performed based on this distance metric. First, we need to determine the number of clusters K. This can be determined based on prior knowledge or the elbow rule. The elbow rule is a commonly used heuristic for determining K. Simply put, as K increases, the sum of squared errors (SSE) of the clustering results gradually decreases; when K reaches a certain critical point, the rate of decrease in SSE slows significantly, forming an inflection point similar to an "elbow." The K value corresponding to this inflection point is usually considered the optimal number of clusters. Since the elbow rule and traditional K-means clustering calculations are existing methods, detailed procedures and formulas will not be described here.

[0106] Summary of major steps:

[0107] This completes the optimized construction of the distance metric and the use of the optimized distance metric for clustering.

[0108] Step S3: Irrigation water demand prediction based on clustering results.

[0109] Based on the K-means clustering results from step S2, i.e., the division of the irrigation area into K different categories, irrigation water demand is predicted. Each category region corresponds to a set of cluster centers with specific spectral characteristics and vegetation indices, representing the unique crop growth status of that region. We will directly utilize the data from the cluster centers, combined with expert experience, to predict irrigation water demand. The specific steps are as follows:

[0110] 1. Interpretation of Cluster Center Data: For each category k, the cluster center data contains the average value of that category across various features, directly reflecting the typical characteristics of that category. For example, R... k G k B k NIR k Average reflectance of category k in the red, green, blue, and near-infrared bands. NDVI k The mean NDVI of category k reflects the average growth condition of crops in that category.

[0111] 2. Expert assessment and determination of irrigation water demand: Based on the cluster center data (R) for each category... k G k B k NIR k NDVI k Based on local historical irrigation experience and the water requirements of different crops, the irrigation water demand status of each category is assessed. Based on the assessment results, the irrigation water demand per unit area (W) for each category is given. k Recommended value.

[0112] 3. Calculate the total irrigation water requirement: Calculate the area A for each category. k Then calculate the irrigation water requirement W per unit area for each category. k Multiplying the water demand by its area and then summing up the water demand for all categories, we get the total water demand for the entire irrigation district.

[0113]

[0114] Through the above steps, based on the optimized K-means clustering results and combined with expert experience, we achieved the prediction of irrigation water demand in the irrigation district. Since the optimized distance metric more accurately reflects the irrigation water demand status of different regions, the irrigation water demand prediction based on this clustering result is also more accurate and reliable.

Claims

1. An irrigation water requirement prediction method based on unmanned aerial vehicle remote sensing, characterized in that, Comprising the following steps: Step S1: Data acquisition, using a multi-spectral or hyper-spectral camera-equipped unmanned aerial vehicle to take aerial photographs of the target irrigation area under sunny and well-lit conditions, obtaining high-resolution image data, and synchronously recording GPS information; data preprocessing, performing radiation correction and geometric correction on the unmanned aerial vehicle images to eliminate errors caused by factors such as atmosphere, light, and unmanned aerial vehicle attitude, and registering the images to a unified geographic coordinate system; then calculating the normalized vegetation index NDVI; fusing the pre-processed spectral data, red band reflectivity R, green band reflectivity G, blue band reflectivity B, near-infrared band reflectivity NIR, and vegetation index NDVI to construct a dataset: each pixel is a data point, its attributes include spectral reflectivity and vegetation index value, and finally an N x (m + n) two-dimensional matrix is formed, where N is the number of pixels, m is the number of spectral bands, and n is the number of vegetation indexes, which will be used as input data for K-means clustering in the subsequent steps; Step S2: Cluster optimization calculation based on multi-feature indicative weighting and spatial information adaptive adjustment, comprising the following steps: a. Constructing an optimization factor a according to the differences in indicative effects and correlations between features to preliminarily optimize the distance metric, the construction of the optimization factor aims to optimize the distance metric according to the differences in indicative effects of different features on irrigation water demand and the correlations between features; the specific steps are as follows: Firstly, all m+n features are divided into two groups: spectral features S and vegetation indices V; where spectral features S = (S1, S2, …, S m ) contain spectral reflectance of m wavebands, which are red waveband reflectance R, green waveband reflectance G, blue waveband reflectance B and near-infrared waveband reflectance NIR respectively; vegetation indices V = (NDVI) contain only one normalized difference vegetation index NDVI; Then the indicative weight of each group of features is calculated respectively; for the spectral feature S, the indicative weight of each band S i Pearson correlation coefficient between the normalized difference vegetation index (NDVI) wherein corr(S i , NDVI) represents a correlation coefficient between S i and NDVI, i = 1, 2, …, m; NDVI is an indicator factor of vegetation growth condition and water condition; the correlation coefficient obtained by calculation is normalized to obtain an indicative weight vector r s of spectral features: The normalization processing ensures that the sum of the weights of all spectral bands is 1; Next, we need to weight the spectral features into vector r. s The weight r of the vegetation index v Integrate into a unified weight vector; due to r s The length is 4, corresponding to four spectral bands, and r v Expand into a vector of length 1. Its only element is r v The value; that is Finally, calculate the optimization factor α; and set r s and Concatenate them into a vector and convert it into a diagonal matrix α: in, Indicates that r s and The vectors are concatenated and converted into a diagonal matrix; the optimization factor α is a diagonal matrix whose diagonal elements are the indicative weights of each feature; when calculating the distance, the optimization factor α will be multiplied by the feature vectors to achieve weighting of different features; b. Based on the fusion of local spatial autocorrelation and global variation information, the optimization factor β is constructed to perform secondary optimization of the distance metric, the optimization factor β(x) is calculated and the final optimized distance metric is obtained: the optimization factor β(x) is a diagonal matrix, the elements on the diagonal are determined by the local spatial autocorrelation index and the global variation coefficient of each feature; the calculation formula of β(x) is: beta(x) = diag((Moran's I local (x).^u).*(CV.^z)); Where u and z are two adjustable parameters used to control the influence degree of the local spatial autocorrelation index and the global variation coefficient; these two parameters are determined in a data-driven manner; the median or mean of the local spatial autocorrelation index of all data points is used as the value of u, and the median or mean of the variation coefficient of all features is used as the value of z; ^ represents element-wise power; * represents element-wise multiplication; diag() represents converting a vector to a diagonal matrix; c. Using the secondary optimized distance metric to perform K-means clustering calculation; Step S3: Irrigation water demand prediction based on the clustering results. 2.The method of claim 1, wherein, The construction of the optimization factor β considers the local spatial correlation and global data distribution characteristics of the data, and constructs the optimization factor β by calculating the spatial autocorrelation index of each data point in its local neighborhood and the global variation coefficient of each feature; First, the local neighborhood is determined: for each data point x, its k nearest neighbors are selected to form the local neighborhood N k (x); the local neighborhood N k (x) reflects the spatial information around the data point x; the value of k needs to be selected according to the specific research area and data characteristics through cross-validation. Next, the calculation of the local spatial autocorrelation index is carried out: for each feature f within the local neighborhood N k (x), its local spatial autocorrelation index Moran's I f (x) is calculated; the Moran's I index is a commonly used spatial autocorrelation index for measuring the similarity of a property value in a region with its surrounding regions; The calculation formula is: where: n is the local neighborhood N k the number of pixels within (x) ; w pq is an element of the spatial weight matrix, indicating the spatial relationship between pixel p and pixel q; x p,f denotes the value of pixel p on feature f; x q,f denotes the value of pixel q on feature f; μ f denotes the global mean of feature f. 3.The method of claim 2, wherein, Local spatial autocorrelation index Moran's I f The larger the value of (x) is, the more similar the data point x is to its surrounding data points in the feature f, i.e. the stronger the local spatial correlation is; the Moran's I index measures the spatial autocorrelation of a feature in a region by calculating the product of the differences of each data point and its surrounding data points in the feature; the calculation of the global coefficient of variation is performed: the global coefficient of variation CV of each feature f is calculated f to measure the dispersion degree of the feature value over the entire data set; the calculation formula of CV f is: Where, σ f and μ f Let CV represent the global standard deviation and global mean of feature f, respectively; f The larger the value of f, the greater the global variability of feature f.

Images