Knowledge-driven irrigation farmland sample generation method and system

Abstract

The invention discloses a knowledge-driven irrigation farmland sample generation method and system, and the method comprises the steps: excavating the mechanism knowledge information of an irrigation activity based on a constructed optical gradient model, and carrying out the inversion to obtain the surface soil water content in image data; calculating the water content of the surface soil based on a surface water balance process, and determining a distribution space range of a potential irrigation farmland through a threshold value method; the method comprises the following steps: mining prior knowledge of a target irrigation farmland drawing product based on a distribution space range of a potential irrigation farmland, and performing time filtering and spatial filtering on the target irrigation farmland drawing product to obtain an irrigation farmland coarse sample; and screening the irrigated farmland coarse samples based on an uncertainty theory to generate high-quality irrigated farmland samples with representativeness and reliability. According to the method, multi-source remote sensing image data and existing irrigation farmland drawing products are utilized, mechanism knowledge mining, priori knowledge mining and uncertainty analysis are combined, and high-quality irrigation farmland samples are screened and generated.

Description

Technical Field

[0001] This invention relates to the field of irrigated farmland sample processing, and more specifically, to a knowledge-driven method and system for generating irrigated farmland samples. Background Technology

[0002] Existing methods for obtaining irrigated farmland samples mainly include layer-by-layer reporting and sampling surveys. However, these methods generally suffer from problems such as poor timeliness, insufficient accuracy, and high input costs. Furthermore, they are easily affected by human factors, which limits the accuracy and efficiency of crop monitoring.

[0003] Remote sensing technology, with its advantages of large information capacity, high timeliness, wide coverage, and low cost, provides a new technical means for obtaining spatial distribution data of irrigated farmland. In early irrigated farmland monitoring research, although traditional image extraction methods could achieve higher classification accuracy compared to simple data processing methods, they suffered from drawbacks such as cumbersome operation and high labor intensity in the process of preparing training samples, making it difficult to meet the needs of large-scale applications.

[0004] In recent years, deep learning models have demonstrated significant advantages in remote sensing image interpretation due to their powerful self-learning and fault-tolerant capabilities. Through complex neural network structures, deep learning can uncover the intrinsic connections between information in different spectral bands and extract spectral, texture, and shape feature patterns exhibited by crops at different growth stages during irrigation activities, forming abstract and representative feature representations that effectively overcome the limitations of traditional methods in feature extraction. However, the effective application of deep learning methods heavily relies on the support of large-scale labeled image sets, and such training samples usually need to be constructed through visual interpretation, a time-consuming, labor-intensive process. In large-scale remote sensing data processing, the consistency and accuracy of sample labeling results are often difficult to maintain, leading to insufficient sample quality and representativeness, thus limiting the widespread application of deep learning in irrigated farmland mapping and monitoring.

[0005] To overcome these limitations, some scholars have proposed a sample extraction strategy based on mathematical statistics, which constructs samples by mathematically modeling the mechanism or feature distribution of the target object. However, the parameters of this method have clear physical meanings and are easily affected by complex environments and the diversity of geographical elements, which reduces stability. Another type of method is based on data migration strategy, which uses historical samples and data products to construct new sample points. This can alleviate the problem of insufficient samples to some extent, but it is limited by the low spatial resolution of existing irrigated farmland mapping products, which makes it difficult to meet the application needs of precision agricultural production.

[0006] Therefore, existing technologies still suffer from problems such as high sample acquisition costs, insufficient accuracy, and poor representativeness. There is an urgent need to propose a technical solution that can accurately extract data from irrigated farmland in order to provide reliable data support for agricultural production management and related decision-making. Summary of the Invention

[0007] This invention provides a knowledge-driven method and system for generating irrigated farmland samples. Based on existing medium- and low-resolution irrigated farmland mapping products, it can efficiently acquire a sufficient number of spatially high-resolution and representative irrigated farmland sample data, providing reliable data support for remote sensing irrigated farmland monitoring research and application.

[0008] The technical solution adopted in this invention is: A knowledge-driven method for generating irrigated farmland samples includes the following steps: Acquire multi-source image data and irrigation farmland mapping products for the study area; wherein, the multi-source image data includes optical remote sensing image data, SMAP image data, and precipitation data; preprocess the optical remote sensing image data and irrigation farmland mapping products containing red band, near-infrared band (NIR), and short-wave infrared band (SWIR) respectively to obtain target optical remote sensing image data and target irrigation farmland mapping products; Based on the constructed optical gradient model, the target optical remote sensing image data and SMAP image data are analyzed to mine the mechanism knowledge information of irrigation activities and to retrieve the surface soil moisture content in the image data. The surface soil moisture content is estimated based on the surface water balance process, and the spatial distribution range of potential irrigated farmland is determined by the threshold method. Based on the spatial distribution range of potential irrigated farmland, prior knowledge of the target irrigated farmland mapping products is mined, and temporal and spatial filtering is applied to the target irrigated farmland mapping products to obtain coarse samples of irrigated farmland. Based on uncertainty theory, a coarse sample of irrigated farmland is screened to generate a high-quality irrigated farmland sample that is representative and reliable.

[0009] Furthermore, the preprocessing includes: 2) Preprocess the optical remote sensing image data, including geometric correction, radiometric correction and projection transformation, to obtain the target optical remote sensing image data, which is used to record the physical reflection characteristics of ground objects to invert irrigation activities. 2) Resample the irrigated farmland mapping products to obtain the target irrigated farmland mapping products, which are used to ensure that the resolution of the irrigated farmland mapping products is consistent with that of the optical remote sensing image data.

[0010] Furthermore, the construction process of the optical gradient model is as follows: 1) Construct an optical gradient model and calculate the Normalized Difference Vegetation Index (NDVI) based on the Red band and NIR band of the selected target optical remote sensing image data; the calculation formula is as follows:

[0011] 2) Based on the shortwave infrared (SWIR) band, the conversion reflectance (STR) of the shortwave infrared band is calculated; the calculation formula is as follows:

[0012] 3) Plot a scatter plot of all effective pixels with the Normalized Difference Vegetation Index (NDVI) on the horizontal axis and the Transform Reflectivity (STR) in the shortwave infrared band on the vertical axis to obtain the optical gradient space. Feature space; 4) Filter out the pixels in extreme humidity states from the scatter plot, and fit them to obtain the dry boundary line and wet boundary line, i.e., dry edge and wet edge; where dry edge represents the STR value corresponding to the pixel set fitting line in the dry environment of the region's surface; wet edge represents the STR value corresponding to the pixel set fitting line in the wet environment of the region's surface. 5) Based on From the feature space, dry edge, and wet edge, we obtain the intercepts and slopes of the dry and wet edges; 6) Based on the Normalized Difference Vegetation Index (NDVI), the intercepts and slopes of dry and wet edges, the maximum and minimum STR values ​​of the regional surface under arid and humid environments were calculated.

[0013] 8) Based on the converted reflectance STR of the shortwave infrared band, the maximum and minimum STR values ​​of the regional surface under arid and humid environments, the normalized soil moisture content is calculated.

[0014] 8) Based on SMAP image data, the maximum and minimum soil moisture contents in the SMAP image data are obtained and used as the maximum and minimum surface soil moisture contents in the region under arid and humid environments. 9) Based on normalized soil moisture content, the maximum and minimum surface soil moisture content under arid and humid conditions in the region, the surface soil moisture content is calculated; the calculation formula is as follows:

[0015] Right now

[0016] In the formula, This refers to the moisture content of the topsoil. and These represent the maximum and minimum soil moisture contents obtained from SMAP image data, respectively, as the maximum and minimum values ​​of surface soil moisture content under humid and arid conditions in the region. Vegetation coverage; It is in the near-infrared band; Red band; It is in the shortwave infrared band; The converted reflectivity is for the shortwave infrared band. and They represent The intercept and slope of the dry edge in the feature space. and They represent The intercept and slope of wet edges in the feature space; where dry edges represent regions where the land surface is in an arid environment. Value; wet edge indicates that the area's surface is in a humid environment. value.

[0017] Furthermore, the calculation of surface soil moisture content based on the surface water balance process specifically involves: 1) Construct a surface water balance model using hydrological data and soil physical property parameters; 2) The soil moisture content obtained by inversion through the optical gradient model is used as the initial condition and input into the surface water balance model; 3) Through model simulation, the dynamic changes in soil moisture content at different time scales are calculated.

[0018] Furthermore, the spatial distribution range of potential irrigated farmland is determined using a threshold method, specifically: 1) Set a threshold range for soil moisture content. This threshold range is determined based on the soil type, climate conditions, and irrigation activity characteristics of different regions. 2) Determine the calculated soil moisture content against the set threshold range: if the soil moisture content is within the threshold range, the area is determined to be potential irrigated farmland; otherwise, the area is determined to be non-potential irrigated farmland; based on the determination results, determine the spatial distribution range of potential irrigated farmland.

[0019] Furthermore, the formula for determining the soil moisture content within the set threshold range is as follows:

[0020] In the formula, Indicates the amount of irrigation water supplied; This represents the amount of precipitation, obtained from precipitation data. It indicates the range of change in soil moisture content, which is obtained by the difference in soil moisture content changes between adjacent time series.

[0021] Furthermore, the temporal and spatial filtering of the target irrigated farmland mapping product specifically involves: Pixels that have been identified as irrigated farmland for three consecutive years in the target irrigated farmland mapping product are used as candidate samples for irrigated farmland, and pixels that have not shown any signs of irrigation activity for ten consecutive years in the target irrigated farmland mapping product are used as candidate samples for other land types, thus realizing time filtering of the irrigated farmland mapping product. An eight-neighbor strategy is adopted to retain the center points of the same category of eight pixels around the candidate samples of irrigated farmland and other land use candidate samples, thereby realizing spatial filtering of irrigated farmland mapping products. After time and space filtering, the irrigated farmland mapping product is processed to obtain a coarse sample of irrigated farmland.

[0022] Furthermore, the screening of coarse samples of irrigated farmland based on uncertainty theory specifically involves: The daily average NDVI of a coarse sample of irrigated farmland was calculated. and standard deviation Exclude abnormally irrigated farmland samples that deviate from the mean NDVI, and retain The irrigated farmland samples within the interval were used to obtain high-quality irrigated farmland samples that are representative and reliable.

[0023] Another technical solution adopted by the present invention is: A knowledge-driven irrigated farmland sample generation system includes: The data acquisition and preprocessing module is used to acquire multi-source remote sensing image data and irrigated farmland mapping products and perform preprocessing. The mechanism knowledge mining module is used to analyze the preprocessed target optical remote sensing image data based on the constructed optical gradient model, invert the soil moisture content in the target optical remote sensing image data, and estimate the surface soil moisture content based on the surface water balance process, and determine the distribution space range of potential irrigated farmland through the threshold method. The prior knowledge mining module is used to mine prior knowledge of irrigation farmland mapping products based on the spatial distribution range of potential irrigated farmland, and to perform spatial and temporal filtering on the irrigation farmland mapping products to obtain coarse samples of irrigated farmland. The uncertainty analysis module is used to screen coarse samples of irrigated farmland based on uncertainty theory to obtain high-quality irrigated farmland samples that are representative and reliable.

[0024] Another technical solution adopted in this invention is: A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the knowledge-driven irrigated farmland sample generation methods described above.

[0025] Compared with the prior art, the present invention has the following advantages: This invention utilizes multi-source remote sensing imagery data and existing irrigated farmland mapping products, combined with mechanistic knowledge mining, prior knowledge mining, and uncertainty analysis, to screen and generate high-quality irrigated farmland samples. Specifically: First, a simple, effective, and easy-to-implement framework for generating irrigated farmland samples is constructed, which can automatically acquire high-quality samples without extensive manual annotation. Second, in the sample generation process, not only is a physical model of irrigation activities introduced as mechanistic knowledge, but also the surface water balance process is combined to more accurately determine the potential distribution range of irrigated farmland. Finally, while considering the impact of complex environments and noise interference from remote sensing data, prior knowledge from existing irrigated farmland mapping products is used to collaboratively optimize coarse sample data based on uncertainty analysis, thereby significantly improving the quality and representativeness of the samples. Attached Figure Description

[0026] The present invention will be further described below with reference to the accompanying drawings and specific embodiments: Figure 1 This is a flowchart illustrating a knowledge-driven method for generating irrigated farmland samples according to the present invention. Figure 2 This is the process of constructing the optical gradient model in this invention. Feature space diagram; Figure 3 This is a schematic diagram of the surface water balance process in this invention; Figure 4 This is a schematic diagram comparing the clustering boundaries of irrigated and non-irrigated farmland samples generated by the method of this invention in the feature space with other methods; Figure 5 This is a schematic diagram illustrating the verification accuracy of the proportion of samples randomly drawn from the irrigated farmland sample bank in this invention, where proportions (a) to (e) are 10%, 30%, 50%, 70%, and 100%, respectively; Figure 6 This is a schematic diagram of the overall distribution results of irrigated farmland extracted by the sample generation strategy of the present invention, wherein (a) to (g) represent Sentinel-2 true color image, irrigated farmland field distribution data, IRRIMAP mapping product, IRRIMAP as sample input result, OPTRAM recognition result, OPTRAM as sample input and sample generation strategy result of the present invention, respectively. Figure 7 This is a schematic diagram illustrating the extraction results of a local area of ​​irrigated farmland using the sample generation strategy of this invention. Figure 8 This is a schematic diagram of the structure of a knowledge-driven irrigated farmland sample generation system in this invention. Detailed Implementation

[0027] Example 1

[0028] like Figures 1 to 3 As shown, a knowledge-driven method for generating irrigated farmland samples includes: S1. Acquire multi-source image data and irrigation farmland mapping products of the study area; wherein, the multi-source image data includes optical remote sensing image data, SMAP image data and precipitation data; preprocess the optical remote sensing image data and irrigation farmland mapping products containing red band, near-infrared band (NIR) and short-wave infrared band (SWIR) respectively to obtain target optical remote sensing image data and target irrigation farmland mapping products; Preprocessing includes: 1) Preprocess the optical remote sensing image data, including geometric correction, radiometric correction and projection transformation, to obtain the target optical remote sensing image data, which is used to record the physical reflection characteristics of ground objects to invert irrigation activities; 2) Resample the irrigated farmland mapping products to obtain the target irrigated farmland mapping products, which are used to ensure that the resolution of the irrigated farmland mapping products is consistent with that of the optical remote sensing image data.

[0029] Note: In step S1, acquiring multi-source image data is crucial, providing a rich information foundation for subsequent sample generation: optical remote sensing image data captures the spectral characteristics of the land surface, providing key clues for identifying irrigated farmland; SMAP image data provides important information such as soil moisture, helping to further confirm the extent of irrigated farmland; and precipitation data reflects the regional moisture status, aiding in the identification of irrigated farmland. Geometric correction of the optical remote sensing image data eliminates image distortion caused by sensor attitude, Earth curvature, and other factors, ensuring the geometric accuracy of the image; radiometric correction eliminates the influence of atmospheric and solar altitude angle factors on image radiometric values, ensuring that the image radiometric values ​​accurately reflect the reflection or emission characteristics of ground objects; projection transformation converts the image from one projection method to another to adapt to different application needs; and resampling of the irrigated farmland mapping product is performed to match its resolution with the optical remote sensing image data. Since data from different sources may differ in resolution, direct fusion or analysis may lead to information distortion or increased error. Therefore, resampling can ensure the consistency of the two in terms of spatial scale, providing a more accurate data basis for subsequent sample generation.

[0030] S2. Based on the constructed optical gradient model, the target optical remote sensing image data and SMAP image data are analyzed to mine the mechanism knowledge information of irrigation activities and to retrieve the surface soil moisture content in the image data. The construction process of the optical gradient model is as follows: 1) Construct an optical gradient model and calculate the Normalized Difference Vegetation Index (NDVI) based on the Red band and NIR band of the selected target optical remote sensing image data; the calculation formula is as follows:

[0031] 2) Based on the shortwave infrared (SWIR) band, the conversion reflectance (STR) of the shortwave infrared band is calculated; the calculation formula is as follows:

[0032] 3) Plot a scatter plot of all effective pixels with the Normalized Difference Vegetation Index (NDVI) on the horizontal axis and the Transform Reflectivity (STR) in the shortwave infrared band on the vertical axis to obtain the optical gradient space. Feature space; 4) Filter out the pixels in extreme humidity states from the scatter plot, and fit them to obtain the dry boundary line and wet boundary line, i.e., dry edge and wet edge; where dry edge represents the STR value corresponding to the pixel set fitting line in the dry environment of the region's surface; wet edge represents the STR value corresponding to the pixel set fitting line in the wet environment of the region's surface. 5) Based on From the feature space, dry edge, and wet edge, we obtain the intercepts and slopes of the dry and wet edges; 6) Based on the Normalized Difference Vegetation Index (NDVI), the intercepts and slopes of dry and wet edges, the maximum and minimum STR values ​​of the regional surface under arid and humid environments were calculated.

[0033] 9) Based on the converted reflectance STR of the shortwave infrared band, the maximum and minimum STR values ​​of the regional surface under arid and humid environments, the normalized soil moisture content is calculated.

[0034] 8) Based on SMAP image data, the maximum and minimum soil moisture contents in the SMAP image data are obtained and used as the maximum and minimum surface soil moisture contents in the region under arid and humid environments. 9) Based on normalized soil moisture content, the maximum and minimum surface soil moisture content under arid and humid conditions in the region, the surface soil moisture content is calculated; the calculation formula is as follows:

[0035] Right now

[0036] In the formula, This refers to the moisture content of the topsoil. and These represent the maximum and minimum soil moisture contents obtained from SMAP image data, respectively, as the maximum and minimum values ​​of surface soil moisture content under humid and arid conditions in the region. Vegetation coverage; It is in the near-infrared band; Red band; It is in the shortwave infrared band; The converted reflectivity is for the shortwave infrared band. and They represent The intercept and slope of the dry edge in the feature space. and They represent The intercept and slope of wet edges in the feature space; where dry edges represent regions where the land surface is in an arid environment. Value; wet edge indicates that the area's surface is in a humid environment. value.

[0037] Note: The optical gradient model described above comprehensively considers various factors, such as the maximum and minimum soil moisture content under different environments, vegetation cover, specific parameters, and shortwave infrared band conversion reflectance. By utilizing the intercept and slope information of dry and wet edges in the feature space, it can accurately extract the mechanism knowledge information of irrigation activities based on target optical remote sensing image data and SMAP image data, and then invert the soil moisture content in the image data, providing key data support for the subsequent knowledge-driven generation of irrigated farmland samples.

[0038] S3. Based on the surface water balance process, the surface soil moisture content is estimated, and the spatial distribution range of potential irrigated farmland is determined by the threshold method. Specifically, the calculation of surface soil moisture content based on the surface water balance process is as follows: 1) Construct a surface water balance model using hydrological data and soil physical property parameters; 2) The soil moisture content obtained by inversion through the optical gradient model is used as the initial condition and input into the surface water balance model; 3) Through model simulation, the dynamic changes in soil moisture content at different time scales are calculated; The spatial distribution range of potential irrigated farmland is determined using the threshold method, specifically as follows: 1) Set a threshold range for soil moisture content. This threshold range is determined based on the soil type, climate conditions, and irrigation activity characteristics of different regions. Specifically, the formula for determining the soil moisture content within the set threshold range is as follows:

[0039] In the formula, Indicates the amount of irrigation water supplied; This represents the amount of precipitation, obtained from precipitation data. It indicates the range of change in soil moisture content, which is obtained by the difference in soil moisture content changes between adjacent time series; 2) Determine the calculated soil moisture content against the set threshold range: if the soil moisture content is within the threshold range, the area is determined to be potential irrigated farmland; otherwise, the area is determined to be non-potential irrigated farmland; based on the determination results, determine the spatial distribution range of potential irrigated farmland.

[0040] like Figure 3 The diagram illustrates the surface water balance process. Changes in surface soil moisture content include both water input and output. Water input includes precipitation, irrigation water supply, and upwelling recharge, while water output includes evaporation and transpiration losses, soil infiltration, and surface runoff losses. This can be expressed as:

[0041] In the formula, Indicates precipitation (in millimeters). Indicates irrigation water supply (in millimeters). This indicates that the water level is rising to replenish the water supply. This represents evaporation and transpiration loss (mm). This indicates the range of change in soil moisture content. This represents the loss from soil infiltration and surface runoff; currently, high-resolution and accurate estimation of evapotranspiration losses is possible. and loss from soil infiltration and surface runoff It remains difficult, especially without thermal infrared bands; furthermore, under normal circumstances, groundwater recharge occurs via upwelling. It can be ignored, that is =0. In this embodiment, the objective is only to extract potential irrigated farmland as candidate samples for irrigated farmland generated based on the optical gradient model, rather than to accurately calculate irrigation water supply. Therefore, the above theory can be simplified to the soil of the farmland requiring irrigation, where, within a certain period, the irrigation water supply is greater than the difference between the change in surface soil moisture content and precipitation, i.e.: .

[0042] In fact, if an unexpected increase in soil moisture content exceeds precipitation, this abnormal increase can only be attributed to irrigation activities. Therefore, pixels showing irrigation activity that satisfy this relationship can be considered potential irrigated farmland, i.e., candidate samples for irrigated farmland. These potential irrigated farmlands are extracted using a thresholding method (the threshold is denoted by T), i.e.: .

[0043] Considering that the accuracy of soil moisture content based on the optical gradient model is better than 0.1 cm 3 ·cm 3 The topsoil thickness is 5 cm. In this embodiment, an empirical threshold of 5 mm is set. That is, if the difference between the change in soil moisture content and precipitation is greater than 5 mm, farmland irrigation activity is considered to have occurred, and the farmland can be considered as potential irrigated farmland. At the same time, CHIRPS products are used as precipitation parameters. .

[0044] Explanation: Calculating topsoil moisture content using the surface water balance process fully considers the dynamic changes in soil moisture over time and space. Combined with soil physical properties and hydrological data, the calculation results are more accurate and reliable. By setting a reasonable threshold range for soil moisture content and comparing the calculated soil moisture content with this threshold using a specific formula, the spatial distribution range of potential irrigated farmland can be accurately determined, effectively improving the representativeness and accuracy of the sample.

[0045] S4. Based on the spatial distribution range of potential irrigated farmland, mine the prior knowledge of the target irrigated farmland mapping products, and perform time filtering and spatial filtering on the target irrigated farmland mapping products to obtain a coarse sample of irrigated farmland. Specifically, 1) pixels in the target irrigated farmland mapping product that have been marked as irrigated farmland for three consecutive years are used as candidate samples of irrigated farmland, and pixels in the target irrigated farmland mapping product that have not shown any signs of irrigation activity for ten consecutive years are used as candidate samples of other land types, so as to achieve time filtering of irrigated farmland mapping products. 2) An eight-neighbor strategy is adopted to retain the center points of the same category of eight pixels around the candidate samples of irrigated farmland and other land use candidates, thereby realizing spatial filtering of irrigated farmland mapping products. 3) After time filtering and spatial filtering, the irrigated farmland mapping products are used to obtain a rough sample of irrigated farmland.

[0046] Explanation: Temporal filtering effectively eliminates misjudged samples caused by short-term climate fluctuations or data errors by screening continuous time series, thereby improving the temporal stability of irrigated farmland samples; spatial filtering, based on the principle of spatial autocorrelation of ground features, uses an eight-neighborhood strategy to eliminate isolated noise points, ensuring the spatial continuity of irrigated farmland samples; the combination of the two can significantly improve the purity and reliability of coarse irrigated farmland samples.

[0047] S5. Based on uncertainty theory, the coarse samples of irrigated farmland are screened to generate high-quality irrigated farmland samples that are representative and reliable. Specifically, this was achieved by calculating the daily average NDVI of a coarse sample of irrigated farmland. and standard deviation Exclude abnormally irrigated farmland samples that deviate from the mean NDVI, and retain The irrigated farmland samples within the interval were used to obtain high-quality irrigated farmland samples that are representative and reliable.

[0048] Explanation: NDVI (Normalized Difference Vegetation Index) is an important indicator reflecting vegetation growth and cover. Its daily average value reflects the overall vegetation growth level of irrigated farmland over a certain period, while the standard deviation reflects the dispersion of vegetation growth among samples. By setting a reasonable range, outlier samples that deviate significantly from the average NDVI are excluded, effectively eliminating abnormal irrigated farmland areas caused by local special factors (such as pests, diseases, and human interference). This preserves irrigated farmland samples with relatively stable growth and typical characteristics, ensuring that the generated high-quality irrigated farmland samples accurately represent the general characteristics of irrigated farmland in the region, providing reliable data support for subsequent farmland monitoring and water resource management.

[0049] To visually demonstrate the superiority of the knowledge-driven irrigated farmland sample generation strategy in the task of irrigated farmland extraction, the t-SNE (t-distributed stochastic neighbor embedding) algorithm is further employed for visualization analysis. t-SNE is a nonlinear dimensionality reduction method that maps sample points in a high-dimensional, complex data space to a low-dimensional (e.g., two-dimensional) space while preserving the local similarity structure of the original dataset, ensuring that similar data points are as close as possible in the low-dimensional space. Figure 4As shown, 1000 irrigated farmland sample points (represented in red) and 1000 other types of farmland sample points (represented in green) were randomly selected, representing sample extraction strategies based on mathematical statistics, data transfer, and knowledge-driven irrigated farmland sample generation strategies, respectively. t-SNE visualization analysis revealed that samples generated using the proposed strategy exhibited significantly clearer clustering boundaries on the one-dimensional projection map, indicating higher distinguishability between irrigated farmland and other land types, with fewer instances of mixed misclassification. In contrast, when the data transfer strategy was directly used as sample input, the distribution of sample points from different categories in two-dimensional space was more mixed, making effective differentiation difficult. This suggests that the model may struggle to accurately distinguish irrigated farmland from other farmland due to feature confusion when processing these samples.

[0050] The sample generation method of this invention effectively solves the problem of poor sample quality caused by image noise, insufficient spatial resolution, and complex environments in traditional sample generation methods, achieving high-precision construction of irrigated farmland samples. Therefore, the method proposed in this invention can significantly improve the effectiveness and reliability of irrigated farmland sample generation, providing a feasible technical approach for large-scale irrigated farmland monitoring and mapping, and has high practical application value.

[0051] To verify the effectiveness of the method for generating irrigated farmland samples in this invention, a transformer model was trained using training samples of different proportions. Quantitative results are as follows: Figure 5 As can be seen, the model validation accuracy increases with the increase in the proportion of training samples; even using only 10% of the training samples, irrigated farmland shows high accuracy, with an F1 score of 97.3%. In contrast, for samples containing rain-fed farmland, paddy fields, and other land cover types with complex features such as residential areas, at least 50% of the training samples are needed to achieve a similar level of accuracy. This reveals the challenges the model faces when handling multiple land cover types: because these land cover types are less distinguishable from irrigated farmland, the model needs to learn richer data information to extract effective classification rules from complex environments and subtle features. With the increase of generated samples, when 30% and 70% of the training data are added respectively, the model gradually improves its ability to distinguish other land cover types. Nevertheless, as the sample proportion increases, the model's performance improves to varying degrees for both irrigated farmland and other complex land cover types, confirming the effectiveness of the knowledge-driven sample generation strategy in irrigated farmland extraction.

[0052] To verify the reliability of knowledge-driven irrigated farmland sample generation, a sample library was constructed using the generated irrigated farmland samples, irrigated farmland mapping products, and irrigated farmland extraction results based on OPTRAM. Quantitative and qualitative comparisons were then performed on the irrigated farmland extraction results. Townships A, B, C, and D were selected as the study areas. Figure 6 The predicted distribution of irrigated farmland across the entire study area is presented, with annotated local areas such as... Figure 7 As shown: 1) The overall distribution results are as follows Figure 6 As shown, the extracted irrigated farmland data accurately identifies the distribution of most irrigated farmland in Town C; furthermore, some irrigated farmland near the ditches in Township D can also be identified. On the other hand, compared to irrigated farmland mapping products and OPTRAM extraction results, directly using the identification results as training samples for deep learning model training does not significantly improve the irrigated farmland identification effect. In fact, deep learning models struggle to distinguish erroneous parts in the training samples, leading to biases in prediction. The knowledge-driven irrigated farmland sample generation strategy of this invention, to a certain extent, ensures the reliability of the training samples, allowing the extracted results to better align with the actual distribution of irrigated farmland.

[0053] 2) Local region extraction results are as follows Figure 7As shown: A localized area in Township A is an area affected by cloud contamination in the time-series imagery. Cloud occlusion and underlying surface effects both impact irrigated farmland identification, posing a significant challenge to its extraction. In the process of irrigated farmland extraction, OPTRAM-based extraction is affected by uncertainties such as cloud and fog occlusion. Even with enhancements and improvements using deep learning models, overfitting and uncertain inferences regarding cloud and fog areas occur during model training, leading to biased identification results. In contrast, the sample generation strategy of this invention utilizes prior knowledge and information from irrigated farmland mapping products. These products are unaffected by atmospheric conditions, thus mitigating cloud occlusion and underlying surface effects to some extent, providing more accurate irrigated farmland extraction results. The localized areas in Township B and Town C respectively show the distribution areas of rain-fed farmland and irrigated farmland in the imagery, both of which are continuously distributed. It is noteworthy that during the deep learning model training process for irrigated farmland, the inherent errors of existing irrigated farmland mapping products and the uncertainties inherent in OPTRAM-based irrigation extraction results are further amplified and highlighted. Specifically, samples generated from irrigated farmland mapping products use spatial and temporal filtering, resulting in a certain loss of relevant irrigated farmland features. This leads to missed classifications during the extraction of irrigated farmland, and the predicted results are often scattered. Furthermore, the coarse-resolution mapping products introduce pixel mixing issues, making it difficult to distinguish farmland areas adjacent to ditches. During training, the model incorrectly learns pixels containing ditches, increasing the difficulty of discrimination. Conversely, pixel results extracted from OPTRAM irrigation still contain significant misleading noise, causing overclassification during training. The model tends to classify rain-fed farmland as irrigated farmland. Therefore, extracting effective and representative feature information from existing prior knowledge for sample generation is crucial. The irrigated farmland sample generation strategy proposed in this invention reasonably integrates the characteristics of the above two approaches and specifically addresses these problems. The high-quality point samples obtained through this strategy significantly improve the model's ability to distinguish between irrigated and non-irrigated farmland, effectively reducing misclassification and omissions, thereby enhancing the accuracy and reliability of identifying the spatial distribution of irrigated farmland. In some areas of Township D, the low contrast between irrigated farmland and its surrounding environment (such as paddy fields) makes identifying these irrigated farmlands through remote sensing imagery a challenge, especially when the irrigated farmland is distributed in a strip pattern. Existing irrigated farmland mapping products and OPTRAM-based irrigation extraction results often struggle to distinguish these irrigated farmland areas with extremely high similarity to the background. In contrast, both the irrigated farmland generation strategy proposed in this invention and using the extraction results or products directly as training samples can improve the model's ability to distinguish between irrigated farmland and its surrounding environment.

[0054] In summary, compared with other solutions, the knowledge-driven irrigated farmland sample generation strategy proposed in this invention performs better in extracting irrigated farmland samples. This strategy extracts key features of irrigated farmland by deeply mining and utilizing prior knowledge, and generates high-quality sample data accordingly. This enables the more accurate capture of the boundary and internal structure information of irrigated farmland in low-contrast backgrounds, and provides reliable data support for the training of subsequent deep learning models, thereby ensuring the accuracy and stability of the models.

[0055] Example 2

[0056] like Figure 8 As shown, a knowledge-driven irrigated farmland sample generation system includes: The data acquisition and preprocessing module 601 is used to acquire multi-source remote sensing image data and irrigated farmland mapping products and perform preprocessing. The mechanism knowledge mining module 602 is used to analyze the preprocessed target optical remote sensing image data based on the constructed optical gradient model, invert the soil moisture content in the target optical remote sensing image data, and estimate the surface soil moisture content based on the surface water balance process, and determine the distribution space range of potential irrigated farmland through the threshold method. The prior knowledge mining module 603 is used to mine prior knowledge of irrigation farmland mapping products based on the distribution spatial range of potential irrigated farmland, and to perform spatial and temporal filtering on the irrigation farmland mapping products to obtain coarse samples of irrigated farmland. Uncertainty analysis module 604 is used to screen coarse samples of irrigated farmland based on uncertainty theory to obtain high-quality irrigated farmland samples that are representative and reliable. Example 3 Embodiment 3 of this application discloses a computer-readable storage medium. The computer-readable storage medium is, for example, a non-volatile memory, such as magnetic media (e.g., hard disks, floppy disks, and magnetic tapes), optical media (e.g., CD-ROMs and DVDs), magneto-optical media (e.g., optical discs), and hardware devices specifically configured for storing and executing computer-executable instructions (e.g., read-only memory (ROM), random access memory (RAM), flash memory, etc.). A computer program is stored on the computer-readable storage medium. The computer-readable storage medium can be executed by one or more processors or processing devices to implement a knowledge-driven irrigated farmland sample generation method according to the foregoing embodiments.

[0057] Furthermore, it is understood that the foregoing embodiments are merely illustrative examples of the present invention. Provided that the technical features do not conflict, the structure is not contradictory, and the purpose of the invention is not violated, the technical solutions of the various embodiments can be arbitrarily combined and used.

[0058] In the embodiments provided by this invention, it should be understood that the disclosed methods, systems, and devices can be implemented in other ways. For example, the modules included in the systems described above are merely illustrative, and the division of modules is only a logical functional division. In actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0059] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0060] Furthermore, in the various embodiments of the present invention, the functional units / modules can be integrated into one processing unit / module, or each unit / module can exist physically separately, or two or more units / modules can be integrated into one unit / module. The integrated unit / module described above can be implemented in hardware or in the form of hardware plus software functional units / modules.

[0061] The integrated units / modules implemented as software functional units / modules described above can be stored in a computer-readable storage medium. These software functional units, stored in a storage medium, include several instructions to cause one or more processors of a computer device (which may be a personal computer, server, or network device, etc.) to execute some steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0062] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the principles and essence of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A knowledge-driven based method for generating irrigated farmland samples, characterized in that, The method comprises the following specific steps: Obtain multi-source image data and irrigation farmland mapping products of a study area; wherein, the multi-source image data comprises optical remote sensing image data, SMAP image data and precipitation data; the optical remote sensing image data containing red band (Red), near-infrared band (Near Infra-Red, NIR) and short-wave infrared band (Short-Wave-InfraRed, SWIR) and the irrigation farmland mapping products are preprocessed respectively to obtain target optical remote sensing image data and target irrigation farmland mapping products; Based on the constructed optical gradient model, the target optical remote sensing image data and the SMAP image data are analyzed to mine the mechanism knowledge information of irrigation activities, and the surface soil water content in the image data is inversely obtained; Based on the surface water balance process, the surface soil water content is calculated, and the threshold method is used to determine the distribution spatial range of the potential irrigation farmland; Based on the distribution spatial range of the potential irrigation farmland, the prior knowledge of the target irrigation farmland mapping products is mined, and the target irrigation farmland mapping products are time-filtered and space-filtered to obtain irrigation farmland coarse samples; Based on the uncertainty theory, the irrigation farmland coarse samples are screened to generate high-quality irrigation farmland samples with representativeness and reliability.

2. The knowledge-driven based irrigation farmland sample generation method according to claim 1, wherein, The preprocessing comprises: 1) The optical remote sensing image data is preprocessed, including geometric correction, radiation correction and projection transformation processing, to obtain target optical remote sensing image data, which is used to record the physical meaning of the ground object reflection characteristics to inverse the irrigation activities; 2) The irrigation farmland mapping products are resampled to obtain target irrigation farmland mapping products, which are used to ensure that the resolution of the irrigation farmland mapping products is consistent with that of the optical remote sensing image data.

3. The knowledge-driven based method for generating irrigation field samples according to claim 1, wherein, The construction process of the optical gradient model is as follows: 1) The optical gradient model is constructed, and the normalized vegetation index NDVI is calculated based on the selected red band Red and near-infrared band NIR in the target optical remote sensing image data; the calculation formula is as follows: 2) The short-wave infrared band conversion reflectivity STR is calculated based on the short-wave infrared band SWIR; the calculation formula is as follows: 3) Scatter plot of all valid pixels with normalized difference vegetation index (NDVI) as horizontal axis and shortwave infrared band's single band terrain corrected reflectance (STR) as vertical axis, to get optical gradient space, i.e. feature space; 4) The pixel points of extreme humidity state are selected from the scatter diagram to fit the dry boundary line and the wet boundary line, i.e. the dry edge and the wet edge; wherein, the dry edge represents the STR value corresponding to the fitting line of the pixel set of the regional surface in the arid environment; the wet edge represents the STR value corresponding to the fitting line of the pixel set of the regional surface in the humid environment; 5) based on characteristic space, dry edge and wet edge, to obtain the intercept and slope of the dry edge and wet edge; 6) The maximum STR value and the minimum STR value of the regional surface in the arid and humid environments are calculated based on the normalized vegetation index NDVI, the intercept and the slope of the dry edge and the wet edge; 7) The normalized soil water content is calculated based on the short-wave infrared band conversion reflectivity STR, the maximum STR value and the minimum STR value of the regional surface in the arid and humid environments; 8) The maximum soil water content and the minimum soil water content in the SMAP image data are obtained based on the SMAP image data, which are taken as the maximum surface soil water content and the minimum surface soil water content of the regional surface in the arid and humid environments; 9) based on the normalized soil moisture, the maximum surface soil moisture and the minimum surface soil moisture of the region under dry and wet environment, the surface soil moisture is calculated, and the calculation formula is as follows: That is In the formula, This refers to the moisture content of the topsoil. and These represent the maximum and minimum soil moisture contents obtained from SMAP image data, respectively, as the maximum and minimum values ​​of surface soil moisture content under humid and arid conditions in the region. Vegetation coverage; It is in the near-infrared band; Red band; It is in the shortwave infrared band; The converted reflectivity is for the shortwave infrared band. and They represent The intercept and slope of the dry edge in the feature space. and They represent The intercept and slope of wet edges in the feature space; where dry edges represent regions where the land surface is in an arid environment. Value; wet edge indicates that the area's surface is in a humid environment. value.

4. The knowledge-driven based method for generating irrigation field samples according to claim 1, wherein, The surface soil moisture is calculated based on the surface water balance process, specifically: 1) using hydrological element data, combined with soil physical property parameters, a surface water balance model is constructed; 2) the soil moisture obtained by the optical gradient model is input into the surface water balance model as the initial condition; 3) through model simulation, the dynamic change of soil moisture under different time scales is calculated.

5. The knowledge-driven based method for generating an irrigated farmland sample according to claim 4, wherein, The distribution space range of the potential irrigated farmland is determined by the threshold method, specifically: 1) set the threshold range of soil moisture, which is determined according to different soil types, climate conditions and irrigation activity characteristics in different regions; 2) the calculated soil moisture is compared with the set threshold range: when the soil moisture is within the threshold range, the region is determined as potential irrigated farmland; otherwise, the region is determined as non potential irrigated farmland; According to the determination result, the distribution space range of the potential irrigated farmland is determined.

6. The knowledge-driven based method for generating an irrigated farmland sample according to claim 5, wherein, The determination formula of the soil moisture and the set threshold range is as follows: In the formula, represents the irrigation water supply amount; represents the precipitation amount, obtained based on precipitation data acquisition; represents the change range of soil moisture content, obtained by the difference of adjacent time series soil moisture content changes.

7. The knowledge-driven based method for generating an irrigation field sample according to claim 1, wherein, The target irrigated farmland mapping product is subjected to time filtering and spatial filtering, specifically: The pixel points in the target irrigated farmland mapping product that have been labeled as irrigated farmland for three consecutive years are taken as irrigated farmland candidate samples, and the pixel points in the target irrigated farmland mapping product that have shown no signs of irrigation activity for ten consecutive years are taken as other land type candidate samples, so as to realize time filtering of the irrigated farmland mapping product; An eight neighborhood strategy is adopted to retain the center points with the same class in the eight pixels around the irrigated farmland candidate samples and the other land type candidate samples, so as to realize spatial filtering of the irrigated farmland mapping product; After time filtering and spatial filtering, the irrigated farmland mapping product is obtained.

8. The knowledge-driven based method for generating an irrigated farmland sample according to claim 1, wherein, The irrigated farmland coarse sample is screened based on the uncertainty theory, specifically: The daily average NDVI of a coarse sample of irrigated farmland was calculated. and standard deviation Exclude abnormally irrigated farmland samples that deviate from the mean NDVI, and retain The irrigated farmland samples within the interval were used to obtain high-quality irrigated farmland samples that are representative and reliable.

9. A knowledge-driven based irrigated farmland sample generation system, characterized in that, It includes: A data acquisition and preprocessing module is used to acquire multi-source remote sensing image data and irrigated farmland mapping products and perform preprocessing; A mechanism knowledge mining module is used to analyze the target optical remote sensing image data after preprocessing based on the constructed optical gradient model, and to inverse the soil moisture in the target optical remote sensing image; And based on the surface water balance process, the surface soil moisture is calculated, and the distribution space range of the potential irrigated farmland is determined by the threshold method; A priori knowledge mining module is used to mine the priori knowledge of the irrigated farmland mapping product based on the distribution space range of the potential irrigated farmland, to perform spatial filtering and time filtering on the irrigated farmland mapping product, and to obtain the irrigated farmland coarse sample; An uncertainty analysis module is used to screen the irrigated farmland coarse sample based on the uncertainty theory, and to obtain high-quality irrigated farmland samples with representativeness and reliability.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, When the computer program is executed by the processor, the knowledge-driven irrigated farmland sample generation method is realized.

Images