Irrigation area detection method and device based on multi-source remote sensing images and storage medium

Abstract

The application provides an irrigation area detection method and device based on multi-source remote sensing images and a storage medium. The irrigation area detection method based on multi-source remote sensing images comprises the following steps: acquiring an original remote sensing image; preprocessing the original remote sensing image to obtain a multi-source remote sensing image; constructing a training sample; training a multi-task learning model by using the training sample to obtain an irrigation area detection model; inputting an EVI time sequence, an NDVI time sequence and an NDMI time sequence of a multi-source remote sensing image to be identified into the irrigation area detection model to obtain irrigation states and irrigation types of different areas. The irrigation area detection method and device based on multi-source remote sensing images and the storage medium can automatically detect small-farmer irrigation areas in remote sensing images and determine irrigation states and irrigation types of the small-farmer irrigation areas.

Description

Technical Field

[0001] This application relates to the field of agricultural remote sensing technology, and in particular to methods, apparatus and storage media for detecting irrigation areas based on multi-source remote sensing images. Background Technology

[0002] Globally, small-scale irrigation facilities are scattered and lack systematic documentation, making accurate identification crucial for agricultural water resource management, energy infrastructure planning, and food security assessment. Traditional ground-based survey methods are costly, inefficient, and struggle to achieve large-scale coverage. Satellite remote sensing technology offers a potential solution. Existing technologies have seen some studies attempting to map irrigation using high-resolution or medium-resolution time-series imagery. However, these methods generally suffer from limitations such as scarce labeled data, poor model generalization ability, and insufficient utilization of multi-scale information.

[0003] Existing methods are difficult to apply to small-scale, decentralized irrigation scenarios, especially due to limitations such as high annotation costs and weak generalization ability. Therefore, there is an urgent need for a method for automatic detection of irrigation areas for small-scale irrigation. Summary of the Invention

[0004] In view of the shortcomings of the prior art, the purpose of this application is to provide a method, device and storage medium for detecting irrigation areas based on multi-source remote sensing images, which can automatically detect small-scale irrigation areas in remote sensing images and determine the irrigation status and irrigation type of the small-scale irrigation areas.

[0005] Firstly, this application provides a method for detecting irrigation areas based on multi-source remote sensing imagery. The method includes:

[0006] Acquire raw remote sensing images from multiple remote sensing platforms;

[0007] The original remote sensing images are preprocessed to obtain multi-source remote sensing images;

[0008] Construct training samples, which include multi-source remote sensing images and irrigation labels for different regions in the multi-source remote sensing images. Irrigation labels include irrigation status and irrigation type. Irrigation status includes irrigation and non-irrigation, and irrigation type includes conventional irrigation and water-saving irrigation.

[0009] A multi-task learning model is trained using training samples to obtain an irrigation area detection model;

[0010] The EVI, NDVI, and NDMI time series of the multi-source remote sensing images to be identified are input into the irrigation area detection model to obtain the irrigation status and irrigation type of different areas.

[0011] In one implementation of the first aspect, the original remote sensing imagery includes at least two of the following: MODIS medium-resolution imagery, Sentinel-2 high-resolution optical imagery, and UAV imagery.

[0012] In one implementation of the first aspect, constructing training samples includes:

[0013] A hybrid pixel decomposition model was used to decompose the EVI time series of each pixel in the MODIS medium resolution image into a linear combination of three types of endmembers: periodic vegetation, evergreen vegetation, and non-vegetation, and the abundance of each endmember was calculated.

[0014] Potential irrigation areas are selected from MODIS medium-resolution images, where the abundance of periodic vegetation in the pixels of the potential irrigation areas is greater than or equal to a first threshold.

[0015] Cells outside the potential irrigation area are marked as non-irrigated;

[0016] For pixels within the potential irrigation area, the irrigation status and irrigation type are automatically labeled based on their EVI time series.

[0017] In one implementation of the first aspect, for pixels within the potential irrigation area, the irrigation status and type are automatically labeled based on their EVI time series, including:

[0018] If the maximum EVI value of a cell is greater than the second threshold in winter or spring, it is labeled as "irrigated"; otherwise, it is labeled as "non-irrigated".

[0019] For the region labeled "irrigation", calculate the spatial variation coefficient of its EVI image;

[0020] If the spatial variation coefficient is greater than the third threshold, the area is marked as conventional irrigation; otherwise, the area is marked as water-saving irrigation.

[0021] In one implementation of the first aspect, the method for generating endmembers is as follows:

[0022] The EVI time series of each pixel in the MODIS medium-resolution image is normalized to obtain the normalized EVI time series.

[0023] The normalized EVI time series is filtered to obtain a smoothed EVI time series;

[0024] Calculate the pixel purity index of each pixel based on the smoothed EVI time series;

[0025] Pixels with a pixel purity index greater than the fourth threshold are selected as candidate endmembers.

[0026] Clustering algorithms were used to divide candidate endmembers into three clusters based on the shape of the smoothed EVI time series: periodic vegetation, evergreen vegetation, and non-vegetation.

[0027] Centroids of periodic vegetation types were selected as periodic vegetation end-members, centroids of evergreen vegetation types were selected as evergreen vegetation end-members, and centroids of non-vegetation types were selected as non-vegetation end-members.

[0028] In one implementation of the first aspect, the original remote sensing image is preprocessed to obtain multi-source remote sensing imagery, including:

[0029] Obtain the blue band reflectance and near-infrared band reflectance of each pixel in the Sentinel-2 high-resolution optical image;

[0030] Pixels with blue band reflectance greater than the fifth threshold and near-infrared band reflectance less than the sixth threshold are marked as cloud pixels.

[0031] Multi-source remote sensing images were obtained by reconstructing the EVI, NDMI, and NDVI time series that were interfered with by cloud pixels.

[0032] In one implementation of the first aspect, before inputting the irrigation area detection model, the pixels of the multi-source remote sensing image to be identified are initially screened to exclude non-irrigated pixels.

[0033] Initial screening criteria include:

[0034] The 10th percentile EVI value of the EVI time series is less than the seventh threshold and is used to exclude evergreen pixels.

[0035] The 90th percentile EVI value of the EVI time series is greater than the eighth threshold and is used to exclude non-vegetation pixels.

[0036] The largest EVI value in the winter and spring EVI time series is greater than the ninth threshold and is used to exclude non-vegetation pixels.

[0037] If the ratio of the 90th percentile EVI value to the 10th percentile EVI value in an EVI time series is greater than the tenth threshold, it is used to exclude evergreen pixels.

[0038] The slope of the EVI time series in winter is less than -0.01 / day, and the slope of the EVI time series in spring is greater than 0.01 / day, which is used to exclude grassland pixels.

[0039] In one implementation of the first aspect, the multi-task learning model includes:

[0040] The input layer is used to input the EVI time series, NDVI time series, and NDMI time series of pixels that meet the admission criteria.

[0041] Convolutional layers are used to extract local spatiotemporal features from EVI time series, NDVI time series, and NDMI time series.

[0042] Long Short-Term Memory (LSTM) networks are used to extract long-term dynamic features from local spatiotemporal features;

[0043] The cross-attention module is used to generate fused features of long-term dynamic features of EVI time series, NDVI time series and NDMI time series.

[0044] The first task output head is used to determine the irrigation status of pixels based on the fusion features;

[0045] The second task output head is used to determine the irrigation type of the pixel based on the fusion features.

[0046] Secondly, this application provides an irrigation area detection device based on multi-source remote sensing imagery, including a memory and a processor. The processor is used to execute a computer program stored in the memory to enable the irrigation area detection device to perform an irrigation area detection method based on multi-source remote sensing imagery.

[0047] Thirdly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a method for detecting irrigation areas based on multi-source remote sensing images.

[0048] As described above, the irrigation area detection method based on multi-source remote sensing imagery proposed in this application achieves high-quality preprocessing and spatiotemporal consistency of remote sensing data under complex meteorological and topographical conditions by fusing multi-source remote sensing images such as MODIS, Sentinel-2, and UAV images and combining cloud pixel detection and time series reconstruction techniques. By constructing a hybrid pixel decomposition model based on EVI time series and combining it with clustering algorithms to extract three types of phenological endmembers, it achieves fine interpretation of vegetation dynamics in small-scale farming areas and accurate initial screening of potential irrigation areas. By designing automated annotation rules based on EVI seasonal features and spatial variation coefficients, it achieves efficient and objective annotation of irrigation status and irrigation type in training samples, reducing manual dependence and improving sample construction efficiency. By constructing a multi-task learning model that integrates convolutional neural networks, long short-term memory networks, and cross-attention mechanisms, it achieves in-depth mining of multi-dimensional temporal features of EVI, NDVI, and NDMI and simultaneous high-precision prediction of irrigation status and irrigation type. Attached Figure Description

[0049] Figure 1 The flowchart shown is a method for detecting irrigation areas based on multi-source remote sensing images according to an embodiment of this application.

[0050] Figure 2 This is shown as an embodiment of the present application. Figure 1Flowchart of step S300.

[0051] Figure 3 This is shown as an embodiment of the present application. Figure 2 Flowchart of step S340.

[0052] Figure 4 The flowchart shown is a process for obtaining the terminator in step S310 of an embodiment of this application.

[0053] Figure 5 The diagram shown is a structural schematic of a multi-task learning model in one embodiment of this application.

[0054] Figure 6 The diagram shown is a structural schematic of an electronic device according to an embodiment of this application. Detailed Implementation

[0055] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. This application can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, unless otherwise specified, the following embodiments and features in the embodiments can be combined with each other.

[0056] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this application. Therefore, the drawings only show the components related to this application and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0057] The following will describe in detail the principles and implementation methods of the irrigation area detection method, electronic equipment, and system based on multi-source remote sensing imagery in this embodiment, so that those skilled in the art can understand the irrigation area detection method, electronic equipment, and system based on multi-source remote sensing imagery in this embodiment without creative effort.

[0058] like Figure 1 As shown, this embodiment provides a method for detecting irrigation areas based on multi-source remote sensing imagery, including:

[0059] Step S100: Acquire raw remote sensing images from multiple remote sensing platforms;

[0060] Step S200: Preprocess the original remote sensing image to obtain multi-source remote sensing image;

[0061] Step S300: Construct training samples. The training samples include multi-source remote sensing images and irrigation labels for different regions in the multi-source remote sensing images. The irrigation labels include irrigation status and irrigation type. Irrigation status includes irrigation and non-irrigation, and irrigation type includes conventional irrigation and water-saving irrigation.

[0062] Step S400: Train a multi-task learning model using training samples to obtain an irrigation area detection model;

[0063] Step S500: Input the EVI time series, NDVI time series and NDMI time series of the multi-source remote sensing images to be identified into the irrigation area detection model to obtain the irrigation status and irrigation type of different areas.

[0064] The remote sensing image preprocessing method in this embodiment can employ existing technologies, such as radiometric calibration and atmospheric correction, geometric fine correction and spatial registration, cloud and cloud shadow detection and masking, time series reconstruction and interpolation, and spatiotemporal fusion and resampling of multi-source data, thereby obtaining spatiotemporally consistent, cloud interference-restored, radiometrically and geometrically normalized multi-source remote sensing images. The multi-source remote sensing images to be identified can also be obtained by processing the original remote sensing images to be identified using the same preprocessing method.

[0065] The sample labels in this embodiment can be obtained by computer calculation, manual identification by experts through visual remote sensing images, identification by drones capturing close-up images of the scene, or any combination of two or more of these methods.

[0066] The multi-task learning model in this embodiment includes at least two task output heads, which output the irrigation status and irrigation type, respectively. When the irrigation status is non-irrigation, no further identification is required. When the irrigation status is irrigation, the other task output head further identifies whether it is regular irrigation or water-saving irrigation.

[0067] This embodiment acquires and fuses raw remote sensing images from multiple remote sensing platforms to achieve comprehensive coverage and complementary enhancement of multi-scale, multi-temporal surface information of the target area, ensuring the integrity and accuracy of the raw image data and providing a reliable data foundation for subsequent model training. Then, a neural network is trained using training samples containing dual labels for irrigation status and irrigation type. The trained neural network is then used to identify small-scale irrigation areas and simultaneously predict irrigation status and type, quickly and accurately.

[0068] This embodiment utilizes vegetation phenological types to detect the irrigation status and type of the area to be identified. Due to the pixel mixing problem in remote sensing images, the EVI, NDVI, and NDMI time series of the multi-source remote sensing images to be identified are mixtures of EVI, NDVI, and NDMI time series of various typical phenological types. During the model training phase, the main phenological types and corresponding irrigation statuses of pixels are identified by decomposing the mixed pixels, thus obtaining labels. Through training, a mapping relationship is established between the EVI, NDVI, and NDMI time series of pixels and the irrigation status and type. Therefore, during the irrigation area detection phase, the irrigation status and type can be directly determined through the EVI, NDVI, and NDMI time series of pixels. The irrigation area detection model in this embodiment achieves collaborative perception of crop growth dynamics and soil moisture changes by simultaneously inputting EVI, NDVI, and NDMI time series, exhibiting high robustness and accuracy.

[0069] The steps S100 to S500 of the irrigation area detection method based on multi-source remote sensing images in this embodiment will be described in detail below.

[0070] In one specific embodiment, the original remote sensing image in step S100 includes at least two of the following: MODIS medium-resolution image, Sentinel-2 high-resolution optical image, and UAV image.

[0071] This embodiment uses MODIS medium-resolution imagery as the basic temporal data source, leveraging its high temporal resolution to achieve continuous monitoring of crop phenological dynamics and seasonal changes in irrigation behavior. This embodiment also introduces Sentinel-2 high-resolution optical imagery, utilizing its high spatial resolution and multispectral band combination advantages to achieve precise characterization of irrigation patterns at the field scale. Furthermore, this embodiment integrates UAV imagery, utilizing its ultra-high spatial resolution and flexible acquisition capabilities to achieve microstructural analysis and ground truth supplementation for small farmland plots in typical sample areas or complex terrains. Finally, this embodiment uses at least two remote sensing sources in conjunction during the training and inference phases to achieve data redundancy and cross-validation, reducing detection blind spots caused by cloud cover, revisit cycles, or resolution limitations of a single sensor.

[0072] In one specific embodiment, the MODIS medium-resolution imagery has a temporal resolution of 16 days and a spatial resolution of 250 meters. The temporal range of the MODIS medium-resolution imagery covers at least one complete vegetation growth cycle, for example, from June 1, 2014 to June 1, 2024. The MODIS medium-resolution imagery is used to acquire EVI time series, NDVI time series, and NDMI time series. The Sentinel-2 high-resolution optical imagery has a temporal resolution of 10 days and a spatial resolution of 10 meters. The Sentinel-2 high-resolution optical imagery is used to acquire EVI time series, NDVI time series, and NDMI time series. High-resolution UAV imagery is used for visual observation of irrigation facilities and confirmation of irrigation type.

[0073] like Figure 2 As shown, in one specific embodiment, step S300 includes:

[0074] Step S310: Using a hybrid pixel decomposition model, the EVI time series of each pixel in the MODIS medium resolution image is decomposed into a linear combination of three types of endmembers: periodic vegetation, evergreen vegetation, and non-vegetation, and the abundance of each endmember is calculated.

[0075] Step S320: Select potential irrigation areas from the MODIS medium-resolution image, where the abundance of periodic vegetation in the pixels of the potential irrigation areas is greater than or equal to a first threshold.

[0076] Step S330: Mark the pixels outside the potential irrigation area as non-irrigated;

[0077] Step S340: For pixels within the potential irrigation area, automatically label the irrigation status and irrigation type based on their EVI time series.

[0078] Small-scale irrigation areas typically exhibit strong seasonal vegetation dynamics, with spring sowing, summer growth, and autumn harvest. Their EVI curves show a single or double peak shape, distinguishing them from areas with stable evergreen vegetation or bare land with no vegetation change throughout the year. Therefore, this embodiment utilizes the abundance of periodic vegetation to exclude pixels that are clearly not irrigated areas, achieving automated and unsupervised initial screening of irrigation potential areas. This avoids the high cost and subjective bias of manual annotation, effectively filtering out non-target land types such as cities, forests, and deserts, narrowing the scope of subsequent fine-grained annotation, improving sample construction efficiency, providing physically meaningful prior spatial constraints for subsequent automatic annotation, and enhancing label quality and model generalization ability.

[0079] In one specific embodiment, the range of the first threshold is 0.6 to 0.9.

[0080] In one specific embodiment, a manually defined typical phenological curve is used as an endmember template. Based on this, the time series of each pixel is represented as a linear combination of endmembers using a linear mixture model, and the abundance ratio of each endmember corresponding to each pixel is obtained by inversion using non-negative least squares (NNLS).

[0081] like Figure 3 As shown, in one specific embodiment, step S340 includes:

[0082] Step S341: If the maximum EVI value of a pixel is greater than the second threshold in winter or spring, it is marked as "irrigated"; otherwise, it is marked as "non-irrigated".

[0083] Step S342: For the area labeled "irrigation", calculate the spatial variation coefficient of its EVI image;

[0084] Step S343: If the spatial variation coefficient is greater than the third threshold, then the area is marked as conventional irrigation; otherwise, the area is marked as water-saving irrigation.

[0085] Winter and spring are critical periods for initiating irrigation in most dryland farming areas. If the maximum EVI value of a pixel during this period exceeds the second threshold, it indicates human-induced water input and classifies the area as irrigated. Conventional irrigation exhibits strong spatial heterogeneity due to varying farmer management levels, dispersed fields, and diverse irrigation equipment. In contrast, water-saving irrigation (such as drip irrigation and sprinkler irrigation) is typically accompanied by standardized management, resulting in uniform vegetation growth and low spatial variability within the field. This embodiment uses the maximum EVI value to perform a secondary screening of pixels in potential irrigated areas, excluding non-irrigated regions. This embodiment also utilizes the coefficient of spatial variation to distinguish between conventional and water-saving irrigation, achieving intelligent labeling without the need for on-site investigations. This significantly lowers the data labeling threshold, provides high-quality, structured monitoring signals for multi-task models, supports simultaneous prediction of irrigation status and type, and improves the final classification accuracy.

[0086] In one specific embodiment, the spatial variation coefficient is calculated using the following formula:

[0087]

[0088] In the formula, CV is the spatial variation coefficient. The population standard deviation is 1. This is the average value.

[0089] In one specific embodiment, the range of the second threshold is 0.2 to 0.4.

[0090] In one specific embodiment, the third threshold ranges from 0.2 to 0.5.

[0091] like Figure 4As shown, in a specific embodiment, the method for generating endmembers in step S310 is as follows:

[0092] Step S311: Normalize the EVI time series of each pixel in the MODIS medium resolution image to obtain the normalized EVI time series.

[0093] Step S312: Filter the normalized EVI time series to obtain a smoothed EVI time series.

[0094] Step S313: Calculate the pixel purity index of each pixel based on the smoothed EVI time series;

[0095] Step S314: Select pixels with a pixel purity index greater than the fourth threshold as candidate endmembers.

[0096] Step S315: Using a clustering algorithm, candidate endmembers are divided into three clusters according to the shape of the smoothed EVI time series: periodic vegetation, evergreen vegetation, and non-vegetation.

[0097] Step S316: Select the centroid of periodic vegetation as the periodic vegetation end-member, select the centroid of evergreen vegetation as the evergreen vegetation end-member, and select the centroid of non-vegetation as the non-vegetation end-member.

[0098] In one specific embodiment, the pixel purity index is calculated using the following formula:

[0099]

[0100] In the formula, Let M be the purity index of the i-th pixel, and M be the number of random projections. is an indicator function for whether pixel i is an extreme pixel in the k-th random projection. The higher the PPI value, the purer the pixel and the closer its spectral characteristics are to the endmember.

[0101] In one specific embodiment, the clustering algorithm employs a Gaussian mixture model (GMM).

[0102] In one specific embodiment, the fourth threshold ranges from 0.7 to 0.9.

[0103] In one specific embodiment, step S200 includes:

[0104] Step S210: Obtain the blue band reflectance and near-infrared band reflectance of each pixel in the Sentinel-2 high-resolution optical image.

[0105] Step S220: Mark the pixels with blue band reflectance greater than the fifth threshold and near-infrared band reflectance less than the sixth threshold as cloud pixels;

[0106] Step S230: Reconstruct the EVI time series, NDMI time series, and NDVI time series that are interfered with by cloud pixels to obtain multi-source remote sensing images.

[0107] This embodiment utilizes the spectral characteristics of clouds to construct a cloud mask, removing invalid data obscured by clouds from the EVI, NDMI, and NDVI time series. This provides clean, complete, and high signal-to-noise ratio input features for the multi-task model, significantly improving the model's inference stability and generalization performance.

[0108] In one specific embodiment, the fifth threshold ranges from 0.12 to 0.18, and the sixth threshold ranges from 0.15 to 0.3.

[0109] In one specific embodiment, for Sentinel-2 EVI time series, when consecutive missing values ​​are less than a preset threshold (e.g., 3 time steps), linear interpolation is used for completion. When consecutive missing values ​​exceed the preset threshold, a regression method based on regional time endmember phenological curves is used for completion, i.e.:

[0110] Based on the results of the previous mixed pixel decomposition, the main phenological type (such as periodic vegetation, evergreen vegetation, etc.) to which the pixel belongs is determined.

[0111] Use the standardized endmember curve of the corresponding phenological type as a reference template;

[0112] By utilizing the effective observation points of this pixel during non-cloudy periods, a specific phenological model is constructed by fitting its mapping relationship with the endmember curve through least squares regression.

[0113] EVI values ​​for missing periods are predicted based on fitted curves to ensure that the reconstructed data conforms to the typical growth rhythm of the region.

[0114] For missing values ​​in Sentinel-2 NDMI and NDVI time series, linear interpolation was used to complete them. Because NDMI and NDVI are highly tolerant of short-term fluctuations and their trends are usually highly correlated with EVI, linear interpolation can maintain data availability while ensuring efficiency.

[0115] Preprocessing was performed on the EVI, NDVI, and NDMI time series, including outlier removal, normalization, and Savitzky-Golay filtering (window size 7, polynomial order 2) for noise reduction, to ensure the spatial, temporal, and numerical consistency of the multi-source time series data and to provide high-quality input data for subsequent phenological analysis and smallholder irrigation identification.

[0116] This embodiment achieves accurate identification and intelligent repair of cloud pollution data, significantly reducing the interference of invalid observations on model training and inference. The EVI time series in this embodiment employs a dual-mode reconstruction of "short interpolation + long regression," balancing computational efficiency and phenological fidelity while avoiding the introduction of false signals that do not conform to regional agricultural patterns. The NDMI and NDVI time series in this embodiment utilize simplified processing strategies to improve the overall system efficiency while ensuring accuracy. Subsequent unified preprocessing further enhances the numerical stability, spatial comparability, and temporal smoothness of the multi-source time series, providing high-quality, highly consistent input features for irrigation status and type identification models, thereby comprehensively improving detection accuracy, robustness, and generalization ability.

[0117] In one specific embodiment, before inputting the irrigation area detection model, the pixels of the multi-source remote sensing image to be identified are initially screened to exclude non-irrigation pixels.

[0118] Initial screening criteria include:

[0119] The 10th percentile EVI value of the EVI time series is less than the seventh threshold and is used to exclude evergreen pixels.

[0120] The 90th percentile EVI value of the EVI time series is greater than the eighth threshold and is used to exclude non-vegetation pixels.

[0121] The largest EVI value in the winter and spring EVI time series is greater than the ninth threshold and is used to exclude non-vegetation pixels.

[0122] If the ratio of the 90th percentile EVI value to the 10th percentile EVI value in an EVI time series is greater than the tenth threshold, it is used to exclude evergreen pixels.

[0123] The slope of the EVI time series in winter is less than -0.01 / day, and the slope of the EVI time series in spring is greater than 0.01 / day, which is used to exclude grassland pixels.

[0124] In one specific embodiment, the seventh threshold ranges from 0.18 to 0.25.

[0125] In one specific embodiment, the range of the eighth threshold is 0.2 to 0.3.

[0126] In one specific embodiment, the range of the ninth threshold is 0.25 to 0.35.

[0127] In one specific embodiment, the tenth threshold ranges from 0.15 to 0.3.

[0128] like Figure 5 As shown, in one specific embodiment, the multi-task learning model includes:

[0129] The input layer is used to input the EVI time series, NDVI time series, and NDMI time series of pixels that meet the admission criteria.

[0130] Convolutional layers are used to extract local spatiotemporal features from EVI time series, NDVI time series, and NDMI time series.

[0131] Long Short-Term Memory (LSTM) networks are used to extract long-term dynamic features from local spatiotemporal features;

[0132] The cross-attention module is used to generate fused features of long-term dynamic features of EVI time series, NDVI time series and NDMI time series.

[0133] The first task output head is used to determine the irrigation status of pixels based on the fusion features;

[0134] The second task output head is used to determine the irrigation type of the pixel based on the fusion features.

[0135] In one specific embodiment, the input layer receives pixel-level time-series data that meets the admission criteria. The input dimension is: 3 vegetation / moisture indices (EVI, NDVI, NDMI) × 36 time steps (covering the complete crop growth cycle). Each time step corresponds to one observation day, forming a three-dimensional tensor input, which serves as the starting point of the model's backbone network.

[0136] The convolutional layer (CNN module) uses a two-layer one-dimensional convolutional neural network (1D-CNN), with the following structural configuration:

[0137] First layer: 32 filters, kernel size = 3, stride = 1, ReLU activation;

[0138] Second layer: 64 filters, kernel size = 3, stride = 1, ReLU activation;

[0139] By using a sliding window mechanism, local temporal patterns (such as sudden peak increases and slow recovery) and implicit spatial texture response features (indirectly reflected by multi-temporal changes) in the time series of each index are extracted, enhancing the model's sensitivity to vegetation dynamic disturbances caused by irrigation, differences in field structure, and seasonal phenological turning points.

[0140] The Long Short-Term Memory (LSTM) network consists of two stacked layers with 64 hidden units and a Dropout rate of 0.2 to prevent overfitting. LSTM captures long-term dependencies spanning weeks or even months in time series, modeling the continuous impact of irrigation behavior on crop growth (e.g., a steady increase in EVI after irrigation, a sharp drop in NDMI during drought), thus enhancing the model's ability to understand dynamic evolution patterns.

[0141] The Cross-Attention module introduces a multi-head cross-attention mechanism with four attention heads, 32 key and value dimensions, and a Softmax weight normalization function. By calculating the cross-channel correlations among three time-series features—EVI, NDVI, and NDMI—information complementarity and adaptive weighted fusion are achieved. For example, during periods of water stress, the NDMI weight automatically increases; during rapid growth, EVI dominates decision-making. This gives the model a "dynamic focusing" capability, improving the analytical accuracy for complex irrigation response patterns.

[0142] Based on shared core features, two independent task branches are established to achieve cascaded prediction of irrigation status and type. The first task output head (Irrigation Status Head) determines whether a pixel is in an irrigated state (binary classification: "irrigated" or "non-irrigated"). The activation function of the first task output head is Sigmoid, and its output dimension is 1D. The second task output head (Irrigation Type Head) only applies to pixels identified as "irrigated" in the first task, further distinguishing whether they belong to "regular irrigation" or "water-saving irrigation." The activation function of the second task output head is Softmax, and its output dimension is 2D. A conditional activation mechanism is adopted to avoid invalid type predictions for non-irrigated areas, improving system efficiency and the reasonableness of results.

[0143] In one specific embodiment, the multi-task learning model is trained using the Adam optimizer (lr=0.001, β1=0.9, β2=0.999), with a batch size of 32, 100 iterations, and an early stopping patience of 5. During training, a random time offset (±30 days) is used to augment the input sequence, increasing the model's robustness to inconsistent observation times or missing data. Class weights and a balanced region sampling strategy are employed to mitigate the imbalanced irrigation type sample problem. Simultaneously, a region-based cross-validation approach is used, with the F1 score outside the observed region serving as the early stopping metric to ensure the model's generalization ability in unseen regions.

[0144] The performance metrics of the trained model on the test set are as follows:

[0145]

[0146] In one specific embodiment, the prediction results are post-processed, including: converting the probability values ​​into binary labels (irrigated / non-irrigated) using a threshold of 0.5; applying area filtering, setting a minimum mapping unit threshold (0.1 hectares), any patch with an area smaller than this threshold will be reclassified as "non-irrigated" to improve the spatial coherence and rationality of the results. Multi-level result products are output: irrigation probability map, irrigation type map, and reports on irrigation area and interannual variation.

[0147] The scope of protection of the irrigation area detection method based on multi-source remote sensing images in this application is not limited to the order of steps listed in this embodiment. Any solution implemented by adding, subtracting, or replacing steps in the prior art based on the principles of this application is included within the scope of protection of this application.

[0148] like Figure 6 As shown, this embodiment also provides an electronic device, which is a user's mobile device such as a mobile phone, PAD, wearable device, or smart AI device; the electronic device includes a memory for storing computer programs; and a processor for running the computer programs to implement the irrigation area detection method based on multi-source remote sensing images in the above embodiments.

[0149] The memory is connected to the processor via the system bus and communicates with it. The memory stores computer programs, and the processor runs the computer programs to enable electronic devices to perform actions such as... Figures 1 to 4 The method for detecting irrigation areas based on multi-source remote sensing imagery is shown.

[0150] It should also be noted that the system bus mentioned above can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This system bus can be divided into address bus, data bus, control bus, etc. For ease of representation, only one thick line is used in the diagram, but this does not indicate that there is only one bus or one type of bus. The communication interface is used to enable communication between the database access device and other devices (such as clients, read-write databases, and read-only databases).

[0151] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0152] In addition, this embodiment also provides a storage medium storing program instructions, which, when executed by a processor, implement the irrigation area detection method based on multi-source remote sensing images described in the above embodiments.

[0153] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing a processor, and the program can be stored in a computer-readable storage medium. The storage medium is a non-transitory medium, such as random access memory, read-only memory, flash memory, hard disk, solid-state drive, magnetic tape, floppy disk, optical disk, and any combination thereof. The aforementioned storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., digital video disc (DVD)), or a semiconductor medium (e.g., solid-state disk (SSD)).

[0154] The descriptions of the processes or structures corresponding to the above figures each have their own emphasis. For parts of a process or structure that are not described in detail, please refer to the relevant descriptions of other processes or structures.

[0155] The above embodiments are merely illustrative of the principles and effects of this application and are not intended to limit this application. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of this application. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in this application should still be covered by the claims of this application.

Claims

1. A method for detecting irrigation areas based on multi-source remote sensing imagery, characterized in that, The irrigation area detection method based on multi-source remote sensing imagery includes: Acquire raw remote sensing images from multiple remote sensing platforms; the raw remote sensing images include MODIS medium-resolution images, Sentinel-2 high-resolution optical images, and UAV images; The original remote sensing images are preprocessed to obtain multi-source remote sensing images; A training sample is constructed, comprising the multi-source remote sensing image and irrigation labels for different regions within the image. The irrigation labels include irrigation status and irrigation type; the irrigation status includes irrigated and non-irrigated, and the irrigation type includes conventional irrigation and water-saving irrigation. The construction of the training sample includes: employing a hybrid pixel decomposition model to decompose the EVI time series of each pixel in the MODIS medium-resolution image into a linear combination of three endmembers: periodic vegetation, evergreen vegetation, and non-vegetation, and calculating the abundance of each endmember; selecting potential irrigation areas from the MODIS medium-resolution image, where the abundance of periodic vegetation in pixels within these potential irrigation areas is greater than or equal to a first threshold; labeling pixels outside the potential irrigation areas as non-irrigated; and automatically labeling the irrigation status and irrigation type of pixels within the potential irrigation areas based on their EVI time series. A multi-task learning model was trained using the training samples to obtain an irrigation area detection model. The EVI time series, NDVI time series, and NDMI time series of the multi-source remote sensing images to be identified are input into the irrigation area detection model to obtain the irrigation status and irrigation type of different areas.

2. The irrigation area detection method based on multi-source remote sensing imagery according to claim 1, characterized in that, The automatic labeling of irrigation status and type for pixels within the potential irrigation area based on their EVI time series includes: If the maximum EVI value of a cell is greater than the second threshold in winter or spring, it is labeled as "irrigated"; otherwise, it is labeled as "non-irrigated". For the area labeled "irrigation", calculate the spatial variation coefficient of its EVI image; If the spatial variation coefficient is greater than the third threshold, the area is marked as conventional irrigation; otherwise, the area is marked as water-saving irrigation.

3. The irrigation area detection method based on multi-source remote sensing imagery according to claim 1, characterized in that, The method for generating the terminator is as follows: The EVI time series of each pixel in the MODIS medium-resolution image is normalized to obtain the normalized EVI time series. The normalized EVI time series is filtered to obtain a smoothed EVI time series; The pixel purity index of each pixel is calculated based on the smoothed EVI time series. Pixels with a pixel purity index greater than the fourth threshold are selected as candidate endmembers. A clustering algorithm is used to divide the candidate endmembers into three clusters according to the shape of the smoothed EVI time series: periodic vegetation, evergreen vegetation, and non-vegetation. Centroids of periodic vegetation types were selected as periodic vegetation end-members, centroids of evergreen vegetation types were selected as evergreen vegetation end-members, and centroids of non-vegetation types were selected as non-vegetation end-members.

4. The irrigation area detection method based on multi-source remote sensing imagery according to claim 1, characterized in that, The preprocessing of the original remote sensing image to obtain multi-source remote sensing imagery includes: Obtain the blue band reflectance and near-infrared band reflectance of each pixel in the Sentinel-2 high-resolution optical image; Pixels with blue band reflectance greater than the fifth threshold and near-infrared band reflectance less than the sixth threshold are marked as cloud pixels. The multi-source remote sensing image is obtained by reconstructing the EVI time series, NDMI time series, and NDVI time series that are interfered with by the cloud pixels.

5. The irrigation area detection method based on multi-source remote sensing imagery according to claim 4, characterized in that, Before inputting the irrigation area detection model, the pixels of the multi-source remote sensing image to be identified are initially screened to exclude non-irrigated pixels. Initial screening criteria include: The 10th percentile EVI value of the EVI time series is less than the seventh threshold and is used to exclude evergreen pixels. The EVI value at the 90th percentile of the EVI time series is greater than the eighth threshold and is used to exclude non-vegetation pixels. The largest EVI value in the winter and spring EVI time series is greater than the ninth threshold and is used to exclude non-vegetation pixels. If the ratio of the 90th percentile EVI value to the 10th percentile EVI value in the EVI time series is greater than the tenth threshold, it is used to exclude evergreen pixels. The slope of the EVI time series in winter is less than -0.01 / day, and the slope of the EVI time series in spring is greater than 0.01 / day, which is used to exclude grassland pixels.

6. The irrigation area detection method based on multi-source remote sensing imagery according to claim 5, characterized in that, The multi-task learning model includes: The input layer is used to input the EVI time series, NDVI time series, and NDMI time series of pixels that meet the admission criteria. Convolutional layers are used to extract local spatiotemporal features of the EVI time series, the NDVI time series, and the NDMI time series; Long Short-Term Memory (LSTM) network is used to extract long-term dynamic features of the local spatiotemporal features; A cross-attention module is used to generate fused features of the long-term dynamic features of the EVI time series, the NDVI time series, and the NDMI time series; The first task output head is used to determine the irrigation status of the pixel based on the fusion features; The second task output head is used to determine the irrigation type of the pixel based on the fusion features.

7. An irrigation area detection device based on multi-source remote sensing imagery, comprising a memory and a processor, characterized in that, The processor is used to execute the computer program stored in the memory, so that the irrigation area detection device performs the irrigation area detection method based on multi-source remote sensing imagery as described in any one of claims 1 to 6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the irrigation area detection method based on multi-source remote sensing imagery as described in any one of claims 1 to 6.

Images