Farmland water demand accurate prediction method based on remote sensing data

Abstract

The invention provides a farmland water demand accurate prediction method based on remote sensing data, and the method comprises the steps: firstly, synchronously collecting optical, thermal infrared, microwave remote sensing and ground data, carrying out the time-space alignment based on a dynamic time warping algorithm, and constructing a four-dimensional unified data set; then, an improved asymmetric automatic encoder integrating multi-branch encoding and a space attention mechanism is utilized to generate a deep fusion feature vector; and inputting the feature vector into an STCN-LSTM hybrid model, mining spatial features through multi-scale convolution, capturing a time evolution rule in combination with a bidirectional long short-term memory network, and embedding DSSAT crop physiological parameters to optimize the model weight. And finally, the uncertainty is quantized and predicted through Monte Carlo Dropout, and the error is further reduced by adopting a space-time layering correction strategy. The method breaks through the limitation of a single data source, the prediction error is smaller than 5%, a reliable basis is provided for precise irrigation, and the water resource utilization efficiency is remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of agricultural water resource management and remote sensing technology, and particularly relates to a farmland water demand accurate prediction method based on remote sensing data. BACKGROUND

[0002] Improving farmland irrigation efficiency is a key link to promote the process of agricultural modernization. Traditional farmland water resource management often relies on ground observation and experience, lacking accurate grasp of the spatial and temporal distribution of water resources. Remote sensing technology can provide large-scale, high-resolution surface information, which can accurately monitor and evaluate the water status of farmland, providing support for the development of precision agriculture.

[0003] The existing farmland water demand prediction method has the following disadvantages: first, single satellite remote sensing data is limited by the spatial and temporal resolution of the sensor, which restricts the accurate monitoring of farmland water demand, and the spectral radiation value is greatly affected by the weather, with large geographical spatial error; second, the traditional remote sensing-based crop water demand estimation model has high dependence on initial conditions, and small deviations in initial data may lead to large errors in estimation results, and the model parameters are extremely sensitive to setting, and different parameter combinations often lead to significant differences in estimation results; in addition, farmland water demand is affected by crop type, growth stage, soil moisture, temperature, humidity and other factors, and the existing model still has limitations in comprehensively and accurately integrating these complex factors. SUMMARY

[0004] In view of the above technical problems, the present application provides a farmland water demand accurate prediction method based on remote sensing data, to solve the problem that the prior art relies on single type remote sensing data and cannot fully reflect the complex topography of farmland, and to overcome the technical problems such as insufficient adaptability of existing models to spatial and temporal changes.

[0005] In order to achieve the above purpose, the technical scheme of the embodiments of the present application is: In a first aspect, the present application provides a farmland water demand accurate prediction method based on remote sensing data, which comprises: spatial and temporal alignment processing of the multi-source heterogeneous data of the target farmland region collected synchronously, to obtain multi-source spatio-temporal data with consistent spatial and temporal resolution; the multi-source heterogeneous data includes remote sensing data and ground monitoring data; inputting the multi-source spatio-temporal data into three independent branch encoding channels of the improved asymmetric autoencoder corresponding to optical, thermal infrared and microwave data respectively to extract features, and generating a fusion feature vector containing the correlation between vegetation growth, canopy temperature and soil moisture; The fused feature vector is input into the trained STCN-LSTM model for water demand prediction; wherein, the STCN-LSTM model includes: an STCN network, which uses multi-scale convolutional kernels to extract micro, meso and macro spatial features in parallel; and a bidirectional LSTM network, which learns the temporal evolution trend of water demand and embeds crop physiological parameters to dynamically adjust the gating weights, and finally outputs a preliminary predicted value of farmland water demand. The uncertainty of the predicted farmland water demand is quantified, and spatial and temporal error corrections are performed on the quantification results using Kriging interpolation optimization and sliding window method, respectively, to generate target farmland water demand values ​​with confidence intervals.

[0006] In some embodiments, the step of spatiotemporally aligning the multi-source heterogeneous data of the synchronously acquired target farmland area to obtain multi-source spatiotemporal data with consistent spatiotemporal resolution includes: The multi-source heterogeneous data are sequentially standardized, outlier removed, and converted to the WGS84 coordinate system. The preprocessed remote sensing data is used as a long-interval low-sampling sequence, and the preprocessed ground monitoring data is used as a short-interval high-sampling sequence. The two sequences are then non-linearly time-aligned. The Euclidean distance matrix between the aligned sequences is calculated using a dynamic time warping algorithm, and dynamic programming is used to find the optimal path with the minimum cumulative distance. Based on the optimal path, the remote sensing data is interpolated to generate remote sensing sequence data with the same temporal resolution as the ground monitoring data. In the WGS84 coordinate system, each remote sensing sequence data is assigned to the corresponding ground sensor within the pixel's geographic boundary range to generate the multi-source spatiotemporal data.

[0007] In some embodiments, the remote sensing data includes optical remote sensing data, thermal infrared remote sensing data, and microwave remote sensing data; the improved asymmetric autoencoder includes a multi-branch coding layer, a shared fusion layer, and a unified decoding layer; the multi-branch coding layer includes an optical coding branch, a thermal infrared coding branch, and a microwave coding branch; The multi-branch coding layer is used to input the optical remote sensing data, the thermal infrared remote sensing data, and the microwave remote sensing data into their respective branch coding channels. Each branch coding channel includes a cascaded convolutional feature extraction module and a spatial attention enhancement module. The convolutional feature extraction module is used to extract deep spatial features of each remote sensing data through a multi-layer convolutional neural network to obtain a high-dimensional feature map. The spatial attention enhancement module is used to optimize the spatial dimension weights of the high-dimensional feature map to enhance the feature response of key areas and obtain an enhanced feature map. Ground monitoring data is introduced as a conditional input into the spatial attention module of each branch to adjust the attention weights. The shared fusion layer is used to align the enhanced feature maps of each branch in terms of feature map size and then concatenate them along the channel dimension to obtain a fused feature map. The unified decoding layer is used to perform global pooling and fully connected mapping on the fused feature map to generate the fused feature vector.

[0008] In some embodiments, inputting the fused feature vector into a trained STCN-LSTM model for water demand prediction includes: The fused feature vector is input into the STCN network, and 1×1, 3×3, and 5×5 convolutional kernels are used to process the fused feature vector in parallel. Specifically, the 1×1 convolutional kernel is used to extract local-level water demand distribution feature maps of individual crop plants or micro-regions; the 3×3 convolutional kernel is used to extract plot-level water demand distribution feature maps; and the 5×5 convolutional kernel is used to extract macro-level water demand trend feature maps of farmland. The local-level water demand distribution feature maps, plot-level water demand distribution feature maps, and macro-level water demand trend feature maps are concatenated at the channel level, and the concatenated feature maps are convolved using 1×1 convolutional kernels to obtain a multi-level spatial scale water demand feature matrix. The multi-level spatial scale water demand feature matrix is ​​arranged into time-series data according to the time dimension and input into a bidirectional LSTM network to obtain a time feature vector. Based on the multi-level spatial scale water demand feature matrix and the time feature vector, water demand is predicted to obtain the predicted water demand value for the farmland.

[0009] In some embodiments, the forward LSTM layer of the bidirectional LSTM network processes the time-series data in chronological order to learn the historical cumulative trend of water demand; the reverse LSTM layer of the bidirectional LSTM network processes the time-series data in reverse chronological order to trace back the driving factors of historical meteorological and soil conditions on water demand; the outputs of the forward LSTM layer and the reverse LSTM layer are concatenated in the feature dimension and fused through a fully connected layer to obtain a time feature vector containing historical trends and future predictions.

[0010] In some embodiments, the output of the preset STCN-LSTM model further includes a spatiotemporal attention gating mechanism; the step of predicting water demand based on the multi-level spatial scale water demand feature matrix and the temporal feature vector to obtain the predicted water demand value for farmland includes: Using the aforementioned spatiotemporal attention gating mechanism, a temporal attention weight vector is generated based on the crop growth calendar and leaf area index time-series data retrieved from remote sensing. The spatiotemporal attention gating mechanism is then used to perform global average pooling and a Sigmoid activation function on the multi-level spatial scale water demand feature map to calculate a spatial attention weight matrix. The temporal attention weight vector and the spatial attention weight matrix are then multiplied element-wise to generate a spatiotemporal attention weight matrix. This matrix is ​​then multiplied element-wise with the temporal feature vector to obtain a water demand prediction feature enhanced with spatiotemporal attention. This water demand prediction feature is input into a fully connected layer and mapped to the predicted water demand value for farmland.

[0011] In some embodiments, the crop physiological parameters include leaf area index (LAI) retrieved from remote sensing and temperature and humidity measured at ground weather stations; the method further includes: The crop physiological parameters are calculated based on the DSSAT crop growth model; the stomatal conductance correction coefficient is calculated in real time based on the crop physiological parameters; and the stomatal conductance correction coefficient is used as an adjustment factor to dynamically adjust the weights of the input gate or forget gate of the bidirectional LSTM network.

[0012] In some embodiments, the uncertainty quantification of the predicted water demand of the target farmland, and the spatial error correction and temporal error correction of the quantification results by Kriging interpolation optimization and sliding window method respectively, to generate a target water demand value for farmland carrying a confidence interval, includes: The fused feature vector is input into the preset STCN-LSTM model for N forward propagation predictions to obtain N water demand prediction values; where N is an integer greater than 1; the mean of the calculated N water demand prediction values ​​is used as the preliminary prediction baseline value; the variance of the calculated N water demand prediction values ​​is used as the prediction uncertainty threshold; the preliminary prediction baseline value is compared with the uncertainty threshold. If the preliminary prediction baseline value is greater than the uncertainty threshold, the current prediction uncertainty is determined to be high, triggering the subsequent hierarchical error correction process; if the preliminary prediction baseline value is less than or equal to the uncertainty threshold, the current prediction uncertainty is determined to be low, and the preliminary prediction baseline value is directly output to the result verification stage. When the hierarchical error correction process is triggered, the following operations are performed: Based on the N predicted water demand values, the spatial residuals are fitted using the Kriging interpolation optimization algorithm to generate a spatial error correction field, and the preliminary prediction baseline value is spatially corrected to obtain the spatially optimized prediction value. The sliding window method is used to perform a local weighted average of the spatially optimized prediction value in the time dimension to smooth time series noise and obtain the time-optimized prediction value. Based on the uncertainty threshold, a confidence interval is assigned to the time-optimized prediction value to generate the farmland water demand target value carrying the confidence interval.

[0013] This invention provides a method for accurate prediction of farmland water demand based on remote sensing data. On one hand, it significantly improves the traditional autoencoder structure by designing an asymmetric architecture of multiple input branches, a shared fusion layer, and a unified decoding layer. Under this architecture, each independent encoding branch extracts differentiated features based on the physical characteristics of optical, thermal infrared, and microwave data, and embeds a spatial attention mechanism to adaptively enhance the information response of key areas. Through feature map concatenation and residual connection techniques, deep complementarity and coupling of multi-source information at the feature level are achieved. The resulting fused feature vector effectively uncovers the inherent implicit correlation between vegetation growth, canopy temperature, and soil moisture. Compared to traditional methods, its feature representation capability is significantly improved, providing high-quality feature input for the core prediction task. On the other hand, it couples a spatiotemporal convolutional network with a bidirectional long short-term memory network (LSTM). This model utilizes the multi-scale convolutional kernels of the STCN to extract spatial features from micro to macro levels in parallel, while simultaneously using a bidirectional LSTM to capture the complex temporal dynamics of water demand. Furthermore, the model embeds physiological parameter calculation logic derived from the DSSAT crop growth model, ensuring that the prediction process is simultaneously constrained and optimized by both data trends and crop physiological mechanisms. Furthermore, the introduced spatiotemporal attention gating mechanism can dynamically focus on key spatiotemporal nodes. This construction method effectively solves the inherent limitations of traditional data-driven models that emphasize data over physiological aspects, significantly improving the accuracy and agronomic rationality of the prediction results. Thirdly, Monte Carlo Dropout technology is used to quantify the uncertainty of the preliminary prediction results. A hierarchical correction strategy is implemented for different sources of error: Kriging interpolation is used to optimize the spatial distribution error of the predicted values, while a sliding window method is used to calibrate the cumulative deviation in the time series. The final output is precise water demand data with confidence intervals. This method achieves a leap from overall correction to source-based hierarchical correction, not only controlling the prediction error to a lower level but also providing quantitative indicators of prediction reliability, offering higher-confidence data support for precision irrigation decisions. Attached Figure Description

[0014] Figure 1 This is a flowchart illustrating the method for accurately predicting farmland water demand based on remote sensing data provided in an embodiment of the present invention. Figure 2This is a schematic diagram of the structure of the improved asymmetric automatic encoder provided in the embodiment of the present invention; Figure 3 This is a flowchart of the collaborative modeling and prediction logic of the STCN-LSTM model provided in this embodiment of the invention. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings. The described embodiments should not be regarded as limitations on the present invention. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] In the following description, references to "some embodiments" refer to a subset of all possible embodiments; however, it is understood that "some embodiments" may be the same or different subsets of all possible embodiments and may be combined with each other without conflict. Unless otherwise defined, all technical and scientific terms used in the embodiments of the invention have the same meaning as commonly understood by one of ordinary skill in the art to which the embodiments of the invention pertain. The terminology used in the embodiments of the invention is for the purpose of describing the embodiments of the invention only and is not intended to limit the invention.

[0017] The following describes an exemplary application of the farmland water demand precision prediction device based on remote sensing data according to embodiments of the present invention. This device can be implemented as a terminal or a server. In one implementation, it can be implemented as a terminal such as a laptop, tablet, desktop computer, or mobile device. In another implementation, it can also be implemented as a server, which can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms. The terminal and server can be directly or indirectly connected via wired or wireless communication, without limitation in the embodiments of the present invention. The following will describe an exemplary application when the farmland water demand precision prediction device is used as a server.

[0018] This invention provides a method for accurately predicting farmland water demand based on remote sensing data. (See also...) Figure 1 , Figure 1This is a flowchart illustrating a method for accurately predicting farmland water demand based on remote sensing data, provided in an embodiment of the present invention. Figure 1 The steps shown are explained.

[0019] Step S110: Perform spatiotemporal alignment processing on the multi-source heterogeneous data of the target farmland area collected synchronously to obtain multi-source spatiotemporal data with consistent spatiotemporal resolution; the multi-source heterogeneous data includes remote sensing data and ground monitoring data.

[0020] In some implementations, synchronous acquisition refers to the operation of acquiring data on the same target farmland area using different sensors or monitoring devices within the same or very close time windows. The purpose is to ensure a high degree of consistency in the temporal dimension of data from different sources, thereby reducing the impact of time differences caused by factors such as crop growth and environmental changes on the analysis results.

[0021] In some implementations, multi-source heterogeneous data refers to raw datasets originating from different types of sensors or data acquisition devices, differing in data structure, format, physical meaning, and spatiotemporal resolution. In this application, it specifically refers to a dataset composed of both remote sensing data and ground monitoring data. In this application, remote sensing data includes optical remote sensing data, thermal infrared remote sensing data, and microwave remote sensing data; ground monitoring data includes data from meteorological stations deployed in fields and soil moisture data.

[0022] In some implementations, spatiotemporal alignment refers to unifying heterogeneous data from different sources with different spatial resolutions (pixel sizes) and temporal sampling frequencies to the same spatial grid system and time point through methods such as resampling, projection transformation, coordinate registration, and temporal interpolation.

[0023] In some implementations, spatiotemporal resolution consistency describes the state achieved by the preprocessed multi-source dataset, specifically meaning that all data layers have the same cell size and geographic coordinate reference in space, and the same observation time point or time interval in time.

[0024] In some implementations, multi-source spatiotemporal data refers to a data set that has been spatiotemporally aligned and integrates multiple information sources, while simultaneously containing spatial and temporal dimension information.

[0025] Step S120: Input the multi-source spatiotemporal data into the three independent branch encoding channels of the improved asymmetric autoencoder corresponding to optical, thermal infrared and microwave data respectively for feature extraction, and generate a fusion feature vector containing the correlation between vegetation growth, canopy temperature and soil moisture.

[0026] In some implementations, the improved asymmetric autoencoder is a pre-stored deep learning network structure designed specifically for processing multimodal data, representing an improvement over the standard autoencoder. Its asymmetry is reflected in the encoder portion, which includes multiple structurally independent branch encoding channels with non-shared parameters, each optimized for the characteristics of a specific data type; while the decoder portion can be shared or similarly asymmetric. Its improvement lies in the optimization of the structure or loss function specifically for the characteristics of farmland data.

[0027] In some implementations, independent branch coding channels refer to independent convolutional neural networks in an improved asymmetric autoencoder, each corresponding to a different data source such as optical, thermal infrared, and microwave. Each branch coding channel sequentially performs convolutional feature extraction and spatial attention weighting, then fuses the original features and attention-weighted features through residual connections. Subsequently, the output feature maps of the three branches are concatenated by channel and mapped through global pooling and fully connected layers to generate a fused feature vector.

[0028] In some implementations, a fused feature vector refers to a single, comprehensive feature vector formed by concatenating, weighting, or otherwise combining features extracted from three independent branch encoding channels. This vector integrates information from different data sources and can more comprehensively describe the complex relationship between vegetation growth, canopy temperature, and soil moisture.

[0029] Step S130: Input the fused feature vector into the trained STCN-LSTM model for water demand prediction; wherein, the STCN-LSTM model includes: an STCN network, used to extract micro, meso and macro spatial features in parallel using multi-scale convolutional kernels; and a bidirectional LSTM network, used to learn the temporal evolution trend of water demand and embed crop physiological parameters to dynamically adjust the gating weights, and finally output a preliminary predicted value of farmland water demand.

[0030] In this invention, water demand prediction refers to extracting spatial heterogeneity features from the fused feature vector using the STCN network of a trained STCN-LSTM model to obtain a multi-level spatial scale water demand feature map; inputting the multi-level spatial scale water demand feature map into a bidirectional LSTM network embedding crop physiological parameters to obtain a time feature vector; and based on the multi-level spatial scale water demand feature map and the time feature vector, performing water demand prediction through the output of the trained STCN-LSTM model to determine the predicted water demand value for the target farmland.

[0031] In some implementations, multi-scale convolutional kernels refer to the use of multiple convolutional filters with different receptive field sizes in parallel in the STCN network, which can extract features from different spatial scales simultaneously: small-scale convolutional kernels focus on microscopic local details, medium-scale kernels focus on mesoscopic regional patterns, and large-scale convolutional kernels capture macroscopic overall features.

[0032] In some implementations, micro-features characterize details at the pixel level or within a very small neighborhood, such as the texture of a single vegetation unit. Meso-features characterize patterns in localized areas within a field, such as patches of moisture distribution caused by differences in soil texture. Macro-features characterize the overall pattern of the entire target farmland area or its relationship with the surrounding environment.

[0033] In this invention, crop physiological parameters are calculated based on the DSSAT crop growth model. Specifically, the stomatal conductance correction coefficient is calculated in real time based on the leaf area index (LAI) retrieved from remote sensing and the temperature and humidity measured by ground weather stations. This correction coefficient is then used as a regulating factor to dynamically adjust the weights of the LSTM input gate or forget gate.

[0034] In some implementations, the preliminary water demand forecast for farmland is a direct output from the STCN-LSTM model, without subsequent error correction. It is typically expressed as water volume per unit time and unit area (e.g., mm / day).

[0035] Step S140: Quantify the uncertainty of the predicted farmland water demand, and use Kriging interpolation optimization and sliding window method to correct spatial error and temporal error of the quantification result, respectively, to generate a target value of farmland water demand carrying a confidence interval.

[0036] In some implementations, uncertainty quantification refers to the process of estimating and expressing the uncertainty or error range of model predictions, and its output includes not only a single prediction value, but also the possible distribution or confidence interval of that prediction value.

[0037] In some implementations, Kriging interpolation optimization is a geostatistical spatial interpolation method. In this invention, it is used to optimize the spatial distribution of preliminary prediction results, specifically by using the prediction error of the model prediction points and their spatial correlation to estimate the error at unknown locations, and spatially correcting the predicted values ​​to make the predicted surface more consistent with spatial continuity and correlation laws.

[0038] In some implementations, the sliding window method is a time series data processing technique that defines a fixed-length time window and slides it across the time series, performing operations such as calculating statistics and fitting models only on the data within the window at a time. For example, the prediction error sequence within the most recent time period can be used to estimate the temporal correlation error of the current prediction value and then correct it.

[0039] In some implementations, spatial error correction refers to the correction of systematic or random errors in the spatial distribution of prediction results, aiming to improve the accuracy of predictions in the spatial dimension, mainly achieved through Kriging interpolation optimization.

[0040] In some implementations, time error correction refers to the correction of lags, fluctuations, or non-stationarities in the prediction results over time, aiming to improve the accuracy and smoothness of the predicted values ​​over time, and is mainly achieved through the sliding window method.

[0041] In some implementations, the target value for farmland water demand refers to the final, optimized prediction result obtained after feature fusion, model prediction, uncertainty quantification, and spatiotemporal error correction. This prediction result also carries confidence interval information characterizing its reliability, which can be directly used to guide precision irrigation decisions.

[0042] This invention provides a method for accurate prediction of farmland water demand based on remote sensing data. On one hand, it significantly improves the traditional autoencoder structure by designing an asymmetric architecture of multiple input branches, a shared fusion layer, and a unified decoding layer. Under this architecture, each independent encoding branch extracts differentiated features based on the physical characteristics of optical, thermal infrared, and microwave data, and embeds a spatial attention mechanism to adaptively enhance the information response of key areas. Through feature map concatenation and residual connection techniques, deep complementarity and coupling of multi-source information at the feature level are achieved. The resulting fused feature vector effectively uncovers the inherent implicit correlation between vegetation growth, canopy temperature, and soil moisture. Compared to traditional methods, its feature representation capability is significantly improved, providing high-quality feature input for the core prediction task. On the other hand, it couples a spatiotemporal convolutional network with a bidirectional long short-term memory network (LSTM). This model utilizes the multi-scale convolutional kernels of the STCN to extract spatial features from micro to macro levels in parallel, while simultaneously using a bidirectional LSTM to capture the complex temporal dynamics of water demand. Furthermore, the model embeds physiological parameter calculation logic derived from the DSSAT crop growth model, ensuring that the prediction process is simultaneously constrained and optimized by both data trends and crop physiological mechanisms. Furthermore, the introduced spatiotemporal attention gating mechanism can dynamically focus on key spatiotemporal nodes. This construction method effectively solves the inherent limitations of traditional data-driven models that emphasize data over physiological aspects, significantly improving the accuracy and agronomic rationality of the prediction results. Thirdly, Monte Carlo Dropout technology is used to quantify the uncertainty of the preliminary prediction results. A hierarchical correction strategy is implemented for different sources of error: Kriging interpolation is used to optimize the spatial distribution error of the predicted values, while a sliding window method is used to calibrate the cumulative deviation in the time series. The final output is precise water demand data with confidence intervals. This method achieves a leap from overall correction to source-based hierarchical correction, not only controlling the prediction error to a lower level but also providing quantitative indicators of prediction reliability, offering higher-confidence data support for precision irrigation decisions.

[0043] In some embodiments, step S110 above can also be implemented by steps 111 to 115: Step 111: The multi-source heterogeneous data is sequentially standardized, outlier removed, and converted to the WGS84 coordinate system.

[0044] Step 112: The preprocessed remote sensing data is used as a long-interval low-sampling sequence, and the preprocessed ground monitoring data is used as a short-interval high-sampling sequence. The two types of sequences are then non-linearly time-aligned.

[0045] In some implementations, nonlinear time alignment refers to aligning two sequences that exhibit non-uniform, non-fixed-proportional scaling or distortion on the time axis, namely, a long-interval low-sampling sequence (remote sensing) and a short-interval high-sampling sequence (ground).

[0046] In this invention, nonlinear time alignment aims to solve the problem that time points cannot be directly correlated between remote sensing data (sparse) and ground data (dense) due to different sampling frequencies.

[0047] Step 113: Calculate the Euclidean distance matrix between the two aligned sequences using the dynamic time warping algorithm, and use dynamic programming to find the optimal path with the minimum cumulative distance.

[0048] In some implementations, the dynamic time warping algorithm establishes the optimal temporal matching relationship between sparse remote sensing data points and dense ground data points by calculating the Euclidean distance matrix between two sequences and using dynamic programming to find the optimal path with the minimum cumulative distance.

[0049] In some implementations, cumulative distance refers to the sum of the Euclidean distances of every pair of matching points (one from the remote sensing sequence and one from the ground sequence) along a possible time-aligned path in the dynamic time warping algorithm. The goal of dynamic programming is to find the path that minimizes this sum.

[0050] In some implementations, the optimal path refers to the path found through dynamic time warping algorithms and dynamic programming that connects the start and end points of the Euclidean distance matrix and minimizes the cumulative distance. This path explicitly defines which specific time points in the short-interval high-sampling sequence each remote sensing data point in the long-interval low-sampling sequence should correspond to, and is a key output for achieving high-precision time alignment.

[0051] Step 114: Interpolate the remote sensing data based on the optimal path to generate remote sensing sequence data with the same temporal resolution as the ground monitoring data.

[0052] Step 115: In the WGS84 coordinate system, each remote sensing sequence data is assigned to the corresponding ground sensor within the geographic boundary of the pixel to generate the multi-source spatiotemporal data.

[0053] This invention solves the spatiotemporal matching problem between optical, thermal infrared, microwave remote sensing, and ground data by employing a coordinate system and dynamic time warping algorithm. This significantly increases the effective data retention rate under cloudy weather conditions from approximately 30% in traditional methods to 92%, fundamentally addressing the issue of insufficient prediction reliability caused by data gaps. Simultaneously, the generated high-quality, spatiotemporally consistent multi-source spatiotemporal data provides optimal input for subsequent feature extraction. This enables the fusion features generated by the improved asymmetric autoencoder to comprehensively cover key water demand factors such as vegetation, temperature, and moisture. Its feature representation capability is improved by more than 40% compared to traditional PCA methods, laying a solid data foundation for achieving high-precision farmland water demand prediction.

[0054] In some embodiments, the remote sensing data includes optical remote sensing data, thermal infrared remote sensing data, and microwave remote sensing data; the improved asymmetric autoencoder includes a multi-branch coding layer, a shared fusion layer, and a unified decoding layer; the multi-branch coding layer includes an optical coding branch, a thermal infrared coding branch, and a microwave coding branch.

[0055] In some implementations, optical remote sensing data refers to remote sensing data used to reflect the spectral reflectance characteristics of the vegetation canopy in a target farmland area. Combinations of specific bands (such as the red band and near-infrared band) can quantitatively characterize key vegetation growth parameters such as vegetation cover, leaf area index, and biomass, and are one of the core inputs for estimating crop evapotranspiration.

[0056] In some implementations, thermal infrared remote sensing data refers to remote sensing data used to retrieve the surface radiant brightness temperature of the target farmland area. This data directly indicates the crop canopy temperature and is a key indicator for diagnosing crop water stress.

[0057] In some implementations, microwave remote sensing data refers to remote sensing data that is extremely sensitive to the surface and shallow soil moisture content of the target farmland area.

[0058] The multi-branch coding layer is used to input the optical remote sensing data, the thermal infrared remote sensing data, and the microwave remote sensing data into the corresponding branch coding channels. Each branch coding channel includes a cascaded convolutional feature extraction module and a spatial attention enhancement module. The convolutional feature extraction module is used to extract deep spatial features of each remote sensing data through a multi-layer convolutional neural network to obtain a high-dimensional feature map. The spatial attention enhancement module is used to optimize the spatial dimension weights of the high-dimensional feature map to enhance the feature response of key areas and obtain an enhanced feature map. Ground monitoring data is introduced as a conditional input into the spatial attention module of each branch to adjust the attention weights.

[0059] In some implementations, the convolutional feature extraction module is a core component in each branch coding channel, consisting of multiple convolutional operations. Its function is to automatically learn and extract deep spatial features related to water demand from the input remote sensing imagery. For example, it can extract texture features indicating uneven spatial distribution of vegetation from optical images, extract spatial heterogeneity patterns of temperature from thermal infrared images, and extract backscattering spatial patterns caused by differences in soil moisture from microwave images.

[0060] In some implementations, the spatial attention enhancement module is embedded in the intelligent feature selection mechanism within each branch. Based on the input feature map, a spatial weight map is automatically generated to enhance the feature response of key regions. In this scheme, its attention is directed to areas crucial for water demand prediction, such as vigorous crop areas, patches experiencing heat stress, or areas with significant changes in soil moisture.

[0061] It should be noted that the ground monitoring data is converted into an auxiliary feature map that is spatially aligned with the remote sensing data through a Kriging or depth interpolation network, and is used as a conditional input in the spatial attention module of each branch.

[0062] The shared fusion layer is used to align the enhanced feature maps of each branch by feature map size and then stitch them together along the channel dimension to obtain a fused feature map.

[0063] In some implementations, a shared fusion layer is used to receive enhanced feature maps from the three branches, each already focused by its respective spatial attention enhancement module. Through feature map size alignment and channel-dimensional stitching operations, multiple sets of features representing vegetation state, canopy temperature, and soil moisture are integrated at the pixel level to generate a fused feature map containing rich coupling information.

[0064] The unified decoding layer is used to perform global pooling and fully connected mapping on the fused feature map to generate the fused feature vector.

[0065] In some implementations, the unified decoding layer, a component at the end of the network, is responsible for aggregating and compressing the fused feature maps of the spatial dimension into a global, fixed-length feature representation. It eliminates the influence of spatial location through global pooling operations, extracts global statistical features, and then performs nonlinear transformations and dimensionality reduction via fully connected mapping, ultimately outputting a fused feature vector.

[0066] In some implementations, the aforementioned multi-branch coding layer is specifically used to perform the following operations: An optical coding branch, consisting of a first convolutional layer and a second convolutional layer, is used to extract local texture features of the crop canopy and plot-scale growth differences from optical remote sensing data. After spatial downsampling, a first initial feature map characterizing vegetation vitality, biomass, and spatial distribution is obtained. A thermal infrared coding branch, consisting of a third convolutional layer and a fourth convolutional layer, is used. The third convolutional layer performs channel compression and nonlinear transformation through 1×1 convolution. The fourth convolutional layer uses a 3×3 dilated convolution with a dilation rate of 2 to expand the receptive field and extract spatial correlations of temperature anomaly regions, resulting in a second initial feature map used to locate water stress boundaries and spatial distribution. A microwave coding branch, consisting of a fifth convolutional layer and a sixth convolutional layer, is used. The fifth convolutional layer uses a 3×3 convolutional kernel with a stride of 1 to extract local abrupt changes in soil moisture. The sixth convolutional layer uses two cascaded 3×3 convolutional kernels with a stride of 1 to model large-scale spatial changes and spatial gradients of soil moisture, resulting in a third initial feature map characterizing soil moisture content and spatial variability.

[0067] In some implementations, the optical coding branch refers to a subnetwork in an improved asymmetric autoencoder specifically designed for processing optical remote sensing data. Its structure is optimized for vegetation spectral characteristics, aiming to extract spatial features directly related to crop growth status from optical imagery in layers, providing data support for quantifying vegetation vitality and biomass.

[0068] In some implementations, the thermal infrared coding branch refers to a sub-network in an improved asymmetric autoencoder specifically designed for processing thermal infrared remote sensing data. Its structure is optimized for the characteristics of land surface temperature data, and its core task is to identify spatial patterns related to crop water stress from the temperature field.

[0069] In some implementations, the microwave coding branch refers to the size of the region on the input image corresponding to a point on the feature map in the convolutional neural network. The fourth convolutional layer expands the receptive field, enabling it to perceive the spatial relationship between temperature anomaly regions (such as a drought-stricken crop area) and surrounding normal regions, thereby more accurately defining the boundaries and extent of stress.

[0070] In some embodiments, the spatial attention enhancement module receives multiple input feature maps of size H×W×C, including a first initial feature map, a second initial feature map, and a third initial feature map. Specifically, the spatial attention enhancement module performs the following operations: First, it performs global average pooling and global max pooling on the input feature maps respectively, obtaining two pooled feature vectors of dimension 1×1×C. The two pooled feature vectors are then concatenated to obtain a fused description vector of dimension 1×1×2C. Second, the fused description vector is input into a shared fully connected layer for dimensionality reduction and feature transformation. Specifically, the dimension of the fused description vector is compressed from 2C to C / 4, and a ReLU activation function is used to perform a non-linear mapping on the dimensionality-reduced fused description vector, resulting in an intermediate feature vector of dimension 1×1×C / 4. Third, the spatial feature vector is restored to 1×1×C through a shared fully connected layer to obtain a channel attention vector. The channel attention vector is then normalized using a Sigmoid activation function to generate an H×W×1 spatial weight matrix. Finally, the spatial weight matrix is ​​multiplied element-wise with the input feature map of the corresponding branch at spatial locations to output the enhanced feature map.

[0071] In some embodiments, the step S130 above, in which the fused feature vector is input into the trained STCN-LSTM model for water demand prediction, can be achieved through the following steps 131 to 134: Step 131: Input the fused feature vector into the STCN network and process the fused feature vector in parallel using 1×1, 3×3, and 5×5 convolutional kernels; wherein, the 1×1 convolutional kernel is used to extract the local water demand distribution feature map of individual crop plants or micro-regions; the 3×3 convolutional kernel is used to extract the plot-level water demand distribution feature map; and the 5×5 convolutional kernel is used to extract the macro-level water demand trend feature map of farmland.

[0072] Step 132: The local water demand distribution feature map, the plot-level water demand distribution feature map, and the macro water demand trend feature map are spliced ​​together at the channel level, and the spliced ​​feature map is convolved using a 1×1 convolution kernel to obtain a multi-level spatial scale water demand feature matrix.

[0073] Step 133: Arrange the multi-level spatial scale water demand feature matrix into time series data according to the time dimension, and input it into a bidirectional LSTM network to obtain the time feature vector.

[0074] Step 134: Based on the multi-level spatial scale water demand feature matrix and the time feature vector, water demand is predicted to obtain the predicted value of farmland water demand.

[0075] It should be noted that the forward LSTM layer of the bidirectional LSTM network processes the time-series data in chronological order to learn the historical cumulative trend of water demand; the reverse LSTM layer of the bidirectional LSTM network processes the time-series data in reverse chronological order to trace back the driving factors of historical meteorological and soil conditions on water demand; the outputs of the forward LSTM layer and the reverse LSTM layer are concatenated in the feature dimension and fused through a fully connected layer to obtain a time feature vector containing historical trends and future predictions.

[0076] In some implementations, step S134 above can also be achieved through the following steps 1341 to 1344: Step 1341: Using the spatiotemporal attention gating mechanism, a temporal attention weight vector is generated based on the crop growth period calendar and the leaf area index time series data retrieved by remote sensing.

[0077] Step 1342: Using the spatiotemporal attention gating mechanism, global average pooling and Sigmoid activation function are applied to the multi-level spatial scale water demand feature map to calculate the spatial attention weight matrix.

[0078] Step 1343: Perform an outer product operation between the temporal attention weight vector and the spatial attention weight matrix to generate a spatiotemporal attention weight matrix; and multiply it element-wise with the temporal feature vector to obtain the water demand prediction feature enhanced by spatiotemporal attention.

[0079] Step 1344: Input the water demand prediction features into the fully connected layer and map them to the target farmland water demand prediction value.

[0080] In some embodiments, the crop physiological parameters include leaf area index (LAI) retrieved from remote sensing and temperature and humidity measured at ground weather stations; the method further includes: The crop physiological parameters are calculated based on the DSSAT crop growth model; the stomatal conductance correction coefficient is calculated in real time based on the crop physiological parameters; and the stomatal conductance correction coefficient is used as an adjustment factor to dynamically adjust the weights of the input gate or forget gate of the bidirectional LSTM network.

[0081] In some implementations, the shared fusion layer is specifically used to perform the following operations: First, the enhanced feature maps are concatenated along the channel dimension to generate a fused feature map with dimensions H×W×3C; second, the fused feature map is subjected to global average pooling to generate a global feature vector with dimensions 1×1×C; finally, the global feature vector is concatenated with the ground data to generate an associated feature vector.

[0082] In some implementations, the unified decoding layer is specifically used to perform the following operations: First, map the dimensions of the associated feature vectors to a preset dimension to obtain the mapped feature vectors, so as to eliminate feature redundancy and adapt to the input requirements of the subsequent time series prediction model; perform L2 normalization on the mapped feature vectors to obtain the fused feature vectors.

[0083] In some embodiments, step S140 above can also be implemented through steps 141 to 146: Step 141: Input the fused feature vector into the preset STCN-LSTM model to perform N forward propagation predictions to obtain N water demand prediction values; where N is an integer greater than 1.

[0084] Step 142: The mean of the calculated N water demand forecasts is used as the preliminary forecast baseline value; the variance of the calculated N water demand forecasts is used as the uncertainty threshold for the forecast.

[0085] Step 143: Compare the preliminary prediction benchmark value with the uncertainty threshold. If the preliminary prediction benchmark value is greater than the uncertainty threshold, it is determined that the current prediction uncertainty is high, and the subsequent hierarchical error correction process is triggered.

[0086] Step 144: If the preliminary prediction benchmark value is less than or equal to the uncertainty threshold, then the current prediction uncertainty is determined to be low, and the preliminary prediction benchmark value is directly output to the result verification stage.

[0087] Step 145: When the hierarchical error correction process is triggered, perform the following operations: Based on the N predicted water demand values, the spatial residuals are fitted using the Kriging interpolation optimization algorithm to generate a spatial error correction field, and the preliminary prediction baseline value is spatially corrected to obtain the spatially optimized prediction value. The sliding window method is used to perform a local weighted average of the spatially optimized prediction value in the time dimension to smooth time series noise and obtain the time-optimized prediction value. Based on the uncertainty threshold, a confidence interval is assigned to the time-optimized prediction value to generate the farmland water demand target value carrying the confidence interval.

[0088] In this invention, a hybrid STCN-LSTM model is used to achieve collaborative learning and prediction of spatiotemporal features: the STCN network extracts micro, meso, and macro spatial features in parallel using multi-scale convolutional kernels; the bidirectional LSTM network accurately learns the temporal evolution trend of water demand; and the model's internal weights are dynamically optimized by embedding DSSAT crop physiological parameters, and a spatiotemporal attention mechanism is superimposed to enhance the information contribution of key nodes and regions. Ultimately, the model's prediction accuracy is improved by 18% compared to the single STCN model, significantly enhancing its adaptability to differences in water demand patterns at different crop growth stages and in different plots.

[0089] The following will describe an exemplary application of the embodiments of the present invention in a practical application scenario.

[0090] Traditional methods often rely on single types of remote sensing data, making it difficult to comprehensively reflect the complex conditions of farmland. This embodiment integrates optical remote sensing, thermal infrared remote sensing, and microwave remote sensing data. In the data fusion stage, it abandons the simple overlay method and improves the autoencoder architecture in traditional deep learning to enhance the ability to represent factors affecting farmland water demand, overcoming the limitations of single data in reflecting complex farmland information.

[0091] To address the issue of poor adaptability of existing models to spatiotemporal variations, this embodiment constructs a spatiotemporally adaptive SCTN-LSTM model. The main body of the model employs a spatiotemporal convolutional network. In the spatial dimension, convolutional layers capture the spatial distribution characteristics of farmland water demand in different geographical locations. In the temporal dimension, a variant of recurrent neural networks—Long Short-Term Memory (LSTM)—is used to process time-series data. Simultaneously, an attention mechanism is introduced, enabling the model to automatically focus on spatiotemporal regions and time points that significantly impact water demand prediction when processing spatiotemporal data, thereby improving prediction accuracy. To address prediction uncertainties caused by factors such as fluctuations in remote sensing data quality and model structure simplification, an uncertainty quantification and dynamic correction mechanism is established. Traditional Kalman filtering only corrects the overall error; this embodiment employs a hierarchical correction strategy.

[0092] This embodiment provides another method for accurate prediction of farmland water demand based on remote sensing data. The method includes the following steps: Step S1: Perform spatiotemporal alignment processing on the multi-source heterogeneous data of the target farmland area collected synchronously to obtain multi-source spatiotemporal data with consistent spatiotemporal resolution, namely optical-thermal infrared-microwave-terrestrial four-dimensional dataset.

[0093] Step S2: The aligned multi-source data is input into the improved asymmetric autoencoder architecture. The optical data branch extracts crop growth features, the thermal infrared branch analyzes water stress signals, and the microwave branch captures soil moisture gradients to generate a fused feature vector containing the correlation between vegetation growth, canopy temperature, and soil moisture.

[0094] Step S3: Input the fused feature vector into the spatiotemporal adaptive model (SCTN-LSTM). First, extract the spatial heterogeneity features through a multi-scale convolutional layer. Then, learn the time series trend through a bidirectional LSTM optimized by embedding DSSAT crop physiological parameters and output the preliminary predicted value of farmland water demand. Step S4: First, the uncertainty of the predicted farmland water demand is quantified by Monte Carlo dropout, and then the error is corrected by a hierarchical correction strategy. Spatial error is optimized by Kriging interpolation, and temporal error is calibrated by sliding window. Finally, the target value of farmland water demand with confidence interval is obtained.

[0095] Step 1 involves synchronously collecting data using different hardware, specifically including the following: The hardware for acquiring optical remote sensing data is optical remote sensing satellites, represented by the Gaofen series satellites. These data include the Normalized Difference Vegetation Index (NDVI) and the Remote Sensing Retrieved Leaf Area Index (LAI) for crops. They can intuitively reflect the distribution and growth status of crops, provide basic information for assessing crop growth stages and health status, and are an important basis for judging crop water requirements.

[0096] The hardware for acquiring thermal infrared remote sensing data is the thermal infrared sensor carried on the Landsat series satellites, including the surface temperature of the crop canopy. By observing temperature changes, the crop's water stress status can be observed, and the real-time water demand of the crop can be determined.

[0097] The hardware for acquiring microwave remote sensing data is represented by microwave remote sensing satellites such as Sentinel-1. These satellites use microwave signals, which are sensitive to changes in soil moisture content, to retrieve soil moisture information. This data is used to directly obtain soil moisture status, reflecting the soil's water supply capacity, and is a key parameter for calculating farmland water demand.

[0098] The equipment for collecting ground meteorological monitoring data is the ground meteorological station, which includes hourly meteorological data such as temperature, humidity, and wind speed. This data provides meteorological support for water demand calculation and is an important influencing factor in crop water evaporation and transpiration processes.

[0099] The equipment for collecting soil monitoring data is soil sensors, including real-time soil moisture data, which accurately reflects the current soil moisture status and is cross-validated with soil moisture data retrieved from remote sensing, providing a real ground reference for water demand prediction.

[0100] In some implementations, satellite remote sensing data (optical, thermal infrared, microwave) serves as the core of macroscopic observation, reflecting crop growth status, water stress status, and soil moisture content, respectively. It offers broad spatial coverage but low temporal resolution. Ground-based data focuses on real-time microscopic monitoring. Ground weather stations provide hourly meteorological parameters such as temperature, humidity, and wind speed, while soil sensors output real-time soil moisture information. This approach offers high temporal resolution but limited spatial coverage. To address the discrepancy between the long-period satellite data and the high-frequency ground data, as well as the mismatch in spatial scale, this solution employs the DTW algorithm. The specific steps are as follows.

[0101] The collected data is spatiotemporally aligned using the DTW algorithm, which consists of three steps: preprocessing, temporal alignment, and spatial alignment. In the preprocessing stage, all data is first standardized, units are unified, and outliers are removed. Simultaneously, all data is converted to the WGS84 coordinate system to eliminate geographic bias, laying the foundation for subsequent alignment. Temporal alignment is crucial. Satellite data is initially treated as a long-interval, low-sampling sequence, while ground data is treated as a short-interval, high-sampling sequence. For example, the NDVI sequence from optical remote sensing and the ground temperature sequence create a sparse-dense contrast on the time axis. Then, the distance matrix between the two is calculated, quantifying the difference between each time point in the satellite data and each hourly point in the ground data using Euclidean distance. Next, dynamic programming is used to find the optimal path with the minimum cumulative distance, allowing for local stretching or compression of the satellite data time axis. Finally, interpolation is performed based on the path, combined with ground data trends, to generate hourly satellite data, improving the temporal resolution of the satellite data to the hourly level. Spatial alignment, based on a unified coordinate system, ensures precise matching between the interpolated satellite pixels and the ground sensor positions, achieving a spatial correspondence between pixels and measurement points. The final four-dimensional dataset processed by the DTW algorithm is synchronized with ground data at the hourly level in time and maintains the high resolution of satellite data in space. The dataset contains four key information categories: crop growth, canopy temperature, soil moisture, and meteorological moisture. This dataset solves the pain points of spatiotemporal asynchrony and type mismatch in traditional multi-source data, providing a high-quality foundation for the branch attention feature fusion in step S2. This ensures that subsequent water demand prediction can fully integrate complementary information from multiple sources, improving prediction accuracy from multiple dimensions such as crop growth, water stress, soil water supply, and meteorological influences.

[0102] Step S2, based on the spatiotemporally aligned four-dimensional dataset obtained in Step S1, improves the traditional autoencoder to achieve deep fusion of multi-source data. It abandons the symmetrical structure of the traditional autoencoder (single input, single encoder, single decoder) and adopts an asymmetric architecture with multiple input branches, a shared fusion layer, and a unified decoding layer to deeply explore the implicit correlations between vegetation growth, canopy temperature, and soil moisture. The improved asymmetric autoencoder's input layer receives the four-dimensional aligned datasets (optical, thermal infrared, microwave, and ground) generated in Step S1. Optical, thermal infrared, and microwave remote sensing data serve as core inputs, each connected to its corresponding branch encoding channel. Ground data serves as auxiliary features, integrated into the feature extraction process of each branch. Considering the differences in the physical characteristics of the three data types, an independent encoding path is designed for each type to accurately capture features and highlight key information. Each branch includes two levels of processing: convolutional feature extraction and spatial attention enhancement. Finally, the decoding layer reconstructs the associated features and outputs a fusion vector, focusing on feature association mining rather than data reconstruction. Figure 2 This is a schematic diagram of the structure of an improved asymmetric automatic encoder provided in an embodiment of the present invention.

[0103] First, the optical data branch focuses on the structured encoding of crop growth characteristics. Inputs include NDVI and leaf area index from high-resolution satellites. This branch employs a two-layer convolution. The first layer uses a 3×3 convolution kernel with a stride of 1 to extract local texture features of the crop canopy, outputting a feature map with the same size as the input, preserving spatial details. The second layer uses a 5×5 convolution kernel with a stride of 2 to expand the receptive field, capturing plot-scale growth differences. The feature map size is halved, achieving dimensionality compression, ultimately generating a feature matrix that characterizes vegetation vitality, biomass, and spatial distribution. Second, the thermal infrared data branch performs sensitive encoding of water stress signals. Inputs are Landsat series canopy temperature data. This branch also employs a two-layer convolution. The first layer uses a 1×1 convolution kernel to suppress high-frequency noise, preserving spatial continuity of temperature. The second layer uses a 3×3 convolution kernel (with dilated convolution) to expand the receptive field, capturing spatial correlations in temperature anomaly regions. Ultimately, it accurately locates the boundary between the temperature anomaly and normal zones, extracting the spatial distribution signal of water stress. Finally, the microwave data branch dynamically encodes the soil moisture gradient, with Sentinel-1 soil moisture inversion values ​​as input. It is sensitive to changes in soil surface (0-5cm) moisture and designs two convolutional layers. The first layer uses a 3×3 convolution kernel with a stride of 1 to extract local abrupt changes in soil moisture. The second layer uses a 7×7 convolution kernel with a stride of 1 to capture a larger range of soil moisture spatial gradients, taking into account both micro-differences and macro-trends. Finally, a composite feature map of soil moisture content, spatial variability rate, and correlation with crop root distribution is generated.

[0104] In the coding process of each branch, ground data (temperature, humidity, and real-time soil moisture) are introduced as constraints. For example, wind speed data from ground weather stations will affect the attention weight calculation of the thermal infrared branch; real-time soil moisture data from soil sensors will correct the feature extraction of the microwave branch.

[0105] To address the limitation of treating all spatial regions equally in traditional feature extraction, this scheme embeds a spatial attention mechanism into the coding layers of each branch of an improved asymmetric autoencoder. This mechanism dynamically calculates feature weights to adaptively enhance information in key regions. Each branch's coding layer follows a three-step logic: feature input → weight calculation → weighted enhancement. Each branch receives the feature map output from its corresponding convolutional layer, with a uniform size of H×W×C (H being height, W being width, and C being the number of channels), where each channel represents a specific feature. A dynamic weight matrix is ​​generated through global pooling and nonlinear transformation to accurately identify key regions: First, global average pooling and global max pooling are performed on each channel of the feature map to obtain two 1×1×C description vectors; the two vectors are concatenated and input into a shared fully connected layer (the hidden layer dimension is C / 4 to reduce computation), and the nonlinear expressive power is enhanced by the ReLU activation function, and then mapped to a 1×1×C fusion vector through the output layer; finally, the fusion vector is normalized by the Sigmoid activation function to generate an H×W×1 weight matrix with the same spatial size as the feature map (weight values ​​∈ [0,1]), where higher weights indicate a greater contribution of the feature at that spatial location to water demand prediction. The weight matrix is ​​multiplied element-wise with the original feature map of the corresponding branch according to spatial location to enhance features in high-weight regions and suppress features in low-weight regions; at the same time, differentiated strategies are designed for different data characteristics, and the weight calculation is dynamically adjusted according to the input data, and meaningless regions in the feature map are suppressed with low weights.

[0106] It should be noted that, based on the single-source feature enhancement of each branch, the optical, thermal infrared, and microwave features processed by branch attention can be transformed into a unified high-dimensional feature vector through a process of residual optimization, feature concatenation, and compression standardization. This achieves deep coupling of multi-source information while avoiding noise superposition. Here, to avoid the attention mechanism from excessively suppressing key features of weak signals, the original feature map and attention-weighted feature map of each branch are calculated first. Key weak features are retained through residual connection, and the weighting process is calibrated based on the original features to suppress the accumulation of single-source noise, resulting in the residual-optimized branch feature map, which is the enhanced feature map in the above embodiment.

[0107] The three residual-optimized branch feature maps (i.e., enhanced feature maps) are then concatenated along the channel dimension to generate a fused feature map of size H×W×3C. The first C channels correspond to optical features, the middle C channels to thermal infrared features, and the last C channels to microwave features. This clear channel partitioning preserves the physical independence of each branch feature while ensuring the correlation of multi-source features within the same spatial coordinates, laying the foundation for uncovering latent correlations.

[0108] Global average pooling is performed on the fused feature map of H×W×3C to compress the spatial dimension (H×W) to 1, resulting in a 1×1×3C global feature vector, realizing the transformation from spatially distributed features to globally correlated features. Then, a fully connected layer is used to map the 3C dimension to a preset dimension to eliminate dimensional redundancy and adapt to the input requirements of subsequent models. The mapped feature vector is then L2 standardized to ensure that the feature scale of each dimension is consistent, which is convenient for subsequent prediction model processing.

[0109] This step ultimately generates a fused feature vector that contains the correlation between vegetation growth, canopy temperature, and soil moisture. This vector not only integrates the advantages of the three types of data but also uncovers implicit correlations, breaking through the information limitations of single remote sensing data.

[0110] Figure 3 The flowchart of the collaborative modeling and prediction logic of the STCN-LSTM model provided in the embodiments of the present invention is shown.

[0111] Step S3 specifically includes the following: First, the fused feature vector output from step S2 is input into a spatiotemporal convolutional network (STCN) to capture the water demand distribution patterns at different spatial scales. Simultaneously, three different convolutional kernels of varying sizes (1×1, 3×3, and 5×5) are used to process the fused feature vector in parallel. Specifically, the 1×1 convolutional kernel focuses on the water demand characteristics of individual crop plants or micro-regions, preserving local details; the 3×3 convolutional kernel captures the water demand distribution at the plot level (such as the water demand gradient between high-biomass and sparse areas within the same plot); and the 5×5 convolutional kernel covers the macro-level water demand trend across the entire farmland. Then, the feature maps output from the three convolutional kernels are channel-concatenated, and the dimensionality is compressed to a uniform dimension using a 1×1 convolutional kernel, preserving multi-scale spatial information while avoiding computational redundancy caused by excessively high feature dimensions. Finally, the fused feature vectors are processed by the STCN network to generate a three-level spatial scale water demand feature matrix that includes micro-individual plants, meso-plots, and macro-farmland, solving the problem of the single spatial scale and inability to adapt to the heterogeneity of farmland in traditional models.

[0112] Based on the three-level spatial features extracted by STCN, a bidirectional LSTM network is connected to learn the temporal evolution of water demand on the one hand, and achieve synergistic optimization of data trends and physiological mechanisms by embedding crop physiological parameter modules on the other hand.

[0113] The three-level spatial feature matrix output by STCN is arranged into time-series data (dimension: T×H×W×C, where T is the time step) and input into a bidirectional LSTM network. The forward LSTM learns the cumulative trend of water demand from the past to the present (e.g., the decay pattern of water demand from high to low after irrigation); the backward LSTM traces the driving factors of water demand from the present to the past (e.g., whether high water demand in a certain period was caused by previous high temperatures). The outputs of both are fused through a fully connected layer to obtain a time feature vector containing historical trends and future predictions, thus solving the gradient vanishing problem in traditional RNNs when processing long-term series.

[0114] The gating mechanism (input gate, forget gate, output gate) of the LSTM unit is embedded with physiological parameter calculation logic based on the DSSAT crop growth model to achieve dynamic optimization of model parameters, including: Step 1: Using the LAI and surface meteorological monitoring data (temperature and humidity) obtained in step S1 as input, the stomatal conductance correction coefficient (g) is calculated in real time based on the photosynthesis-transpiration correlation formula of the DSSAT model. The formula is as follows: , in, The influence coefficient of LAI (increases linearly with increasing LAI). =0.1+0.2×LAI); Let be the function of temperature influence. , For suitable temperatures for crops, Parameters related to temperature influence; Let be the humidity effect function. RH represents the relative humidity measured at a ground weather station, expressed in % %. The base correction factor is a constant determined by the crop type, such as 1.0 for winter wheat and 0.9 for corn.

[0115] Step 2: Embed the stomatal conductance correction coefficient (g) into the LSTM gating mechanism. When the crop is in the grain-filling stage (high LAI, high photosynthetic demand), the correction coefficient increases, and the input gate ( The acceptance of high water demand characteristics is improved. The correction coefficient g directly adjusts the input gate weight, controlling the input intensity of water demand characteristics. The formula for calculating the corrected weight is: When crops are in dormancy (low LAI), the correction coefficient decreases, the forgetting gate weight increases, and the interference of non-critical water requirement signals is reduced. The corrected weight formula is as follows: ;in, The original weights are the forget gate weights of the LSTM. This embedding solves the problem of fixed parameters in traditional models, enabling the model to adaptively adjust parameters according to crop growth stages and environmental conditions, thus better reflecting the true water requirements of crops.

[0116] To further improve prediction accuracy, spatiotemporal attention gating is introduced at the output of the STCN-LSTM network, which assigns higher weights to features of key nodes in the reproductive period and high water demand areas, avoiding the dilution of key information caused by treating all time and space equally. Temporal attention strengthens key nodes in the crop growth cycle. The gating mechanism, based on the crop growth cycle calendar and combined with LAI changes retrieved from remote sensing, generates a temporal attention weight vector (dimension: T×1). It assigns a weight of 1.2-1.5 times to the time step of water-sensitive nodes such as the grain-filling and tasseling stages, and a weight of 0.5-0.7 times to low-water-demand nodes such as the sowing and ripening stages, thus reinforcing the water demand signal during critical periods. Spatial attention strengthens high-water-demand areas. The gating mechanism performs global average pooling on the spatial feature map output by STCN, extracts the water demand response values ​​of each region, and generates a spatial attention weight matrix (dimension: H×W×1) through the Sigmoid activation function. Differential weighting is applied to low-lying waterlogged areas (low water demand intensity, weight reduced to 0.6), arid areas (high water demand intensity, weight increased to 1.3), and high biomass areas (high water demand intensity, weight increased to 1.4), highlighting spatial areas that significantly influence water demand prediction. Based on the aforementioned spatiotemporal attention gating mechanism, the temporal attention weight vector and the spatial attention weight matrix are multiplied together to generate a spatiotemporal attention weight matrix. This matrix is ​​then multiplied element-wise with the output features of the STCN-LSTM to obtain water demand prediction features that focus on key spatiotemporal information. The spatiotemporally attention-enhanced features are then input into a fully connected layer, mapped to specific farmland water demand prediction values, and the output is a preliminary water demand prediction result containing both spatial distribution and time series information.

[0117] Step S4, based on the preliminary water demand forecast output in Step S3, quantifies the uncertainty and performs hierarchical error correction to eliminate prediction biases caused by factors such as fluctuations in remote sensing data quality and model simplification, ultimately outputting the target water demand value for farmland with confidence intervals. First, the uncertainty of the preliminary prediction results from Step S3 is assessed using the Monte Carlo dropout method. Step S4 continues the dropout layer setting from the training phase of the established STCN-LSTM prediction model, performing 10-20 repeated predictions on the same batch of input fused feature vectors. Each prediction uses a different combination of randomly deactivated neurons, generating 10-20 sets of water demand prediction results. Statistical indicators are calculated for multiple sets of prediction results, using the mean as the preliminary prediction baseline and variance as the uncertainty quantification indicator. A larger variance indicates greater fluctuation and higher uncertainty in the prediction results. An uncertainty threshold is set. If the calculated variance exceeds this threshold, the reliability of the current prediction result is insufficient, triggering the subsequent hierarchical correction process; if the variance is ≤ the threshold, the result verification stage is directly entered, reducing unnecessary computational consumption.

[0118] To address the uncertainties discovered after quantification, a multi-dimensional and targeted hierarchical correction method is adopted, taking into account the sources of error, to avoid the inaccuracy caused by the one-size-fits-all approach of traditional single correction methods.

[0119] Spatial Error Correction (Kriging Interpolation Optimization): Spatial errors mainly stem from differences in spatial resolution of remote sensing data, misjudgment of plot boundaries, and deviations in local water demand distribution caused by topographic relief. This manifests as a mismatch between the predicted spatial distribution of water demand and the actual characteristics of the plot. Spatial error correction is based on the spatial location information in the optical-thermal infrared-microwave-terrestrial four-dimensional dataset obtained in step S1. The preliminary predicted farmland water demand is used as the variable to be corrected, while real-time soil moisture data from ground soil sensors and observed water demand from on-site sampling are used as ground truth sample points. Kriging interpolation is employed to correct the spatial errors. By calculating the spatial correlation between sample points, the predicted values ​​are weighted and adjusted to ensure that the corrected spatial distribution of water demand better matches the actual terrain and crop distribution characteristics of the plot.

[0120] Time error correction: Time errors mainly stem from deviations in crop phenological stage predictions and incomplete meteorological data time series, manifesting as discrepancies between the predicted water demand trend and the actual water demand patterns of crops. Time error correction constructs a time sliding window, using the ground meteorological monitoring data obtained in step S1 and the remotely sensed crop growth parameters (i.e., remotely sensed leaf area index) as inputs to calculate the actual water demand trend coefficient. This coefficient is then compared with the preliminary farmland water demand prediction value from step S3, and calibration is performed on time nodes with significant deviations.

[0121] After the aforementioned uncertainty quantification and stratified correction, the final target value for farmland water demand is output. The data format includes spatial distribution, time series, and confidence intervals. Spatially, water demand is output in units of plot pixels; temporally, water demand forecasts for the next 1-7 days are output in hourly or daily increments; the confidence intervals are calculated based on the variance of Monte Carlo dropout, intuitively reflecting the reliability of the prediction results.

[0122] A hybrid STCN-LSTM model was used to achieve collaborative learning and prediction of spatiotemporal features: the STCN network extracts micro, meso, and macro spatial features in parallel using multi-scale convolutional kernels; the bidirectional LSTM network accurately learns the temporal evolution trend of water demand; and the model's internal weights are dynamically optimized by embedding DSSAT crop physiological parameters, and a spatiotemporal attention mechanism is superimposed to enhance the information contribution of key nodes and regions. Ultimately, the model's prediction accuracy is improved by 18% compared to the single STCN model, significantly enhancing its adaptability to differences in water demand patterns across different crop growth stages and different land parcel characteristics.

[0123] This embodiment quantifies the uncertainty of model prediction results by introducing the Monte Carlo Dropout method and combines it with a hierarchical correction strategy: Kriging interpolation is used to optimize spatial distribution errors, and a sliding window method is applied to calibrate time series errors, providing targeted correction to the predicted values. This strategy stabilizes the final prediction error within 5%, a 42% reduction compared to before correction. The water demand prediction value output by this embodiment comes with a clear confidence interval, fundamentally solving the problems of unquantifiable uncertainty and lack of targeted error correction in traditional models, providing highly reliable data support for irrigation decisions. The farmland water demand target value of this embodiment can be directly used to guide variable irrigation and precise scheduling, effectively avoiding water waste or localized water shortages under traditional uniform irrigation models. In practical applications, it can improve water resource utilization by at least 15%. At the same time, crop growth is significantly improved, successfully achieving synergistic development of water conservation and improved agricultural production efficiency, providing effective technical support for sustainable agricultural development.

[0124] The above description is merely an embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and scope of the present invention are included within the scope of protection of the present invention.

[0125] It should be understood that the phrase "one embodiment" or "an embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of the invention. Therefore, "in one embodiment" or "in an embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. It should be understood that in the various embodiments of the invention, the sequence numbers of the above-described processes do not imply a sequential order of execution; the execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the invention. The sequence numbers of the above-described embodiments of the invention are merely descriptive and do not represent the superiority or inferiority of the embodiments.

[0126] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. In the several embodiments provided by this invention, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components may be combined, or integrated into another system, or some features may be ignored or not performed.

[0127] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for accurate prediction of farmland water demand based on remote sensing data, characterized in that, The method includes: Multi-source heterogeneous data of the target farmland area collected synchronously are spatiotemporally aligned to obtain multi-source spatiotemporal data with consistent spatiotemporal resolution; the multi-source heterogeneous data includes remote sensing data and ground monitoring data. The multi-source spatiotemporal data is input into the three independent branch encoding channels of the improved asymmetric autoencoder, corresponding to optical, thermal infrared and microwave data respectively, for feature extraction, generating a fusion feature vector containing the correlation between vegetation growth, canopy temperature and soil moisture. The fused feature vector is input into the trained STCN-LSTM model for water demand prediction; wherein, the STCN-LSTM model includes: an STCN network, which uses multi-scale convolutional kernels to extract micro, meso and macro spatial features in parallel; and a bidirectional LSTM network, which learns the temporal evolution trend of water demand and embeds crop physiological parameters to dynamically adjust the gating weights, and finally outputs a preliminary predicted value of farmland water demand. The uncertainty of the predicted farmland water demand is quantified, and spatial and temporal error corrections are performed on the quantification results using Kriging interpolation optimization and sliding window method, respectively, to generate target farmland water demand values ​​with confidence intervals.

2. The method according to claim 1, characterized in that, The process of spatiotemporally aligning the multi-source heterogeneous data of the target farmland area collected synchronously to obtain multi-source spatiotemporal data with consistent spatiotemporal resolution includes: The multi-source heterogeneous data is sequentially standardized, outlier removed, and converted to coordinate system 1, and then transformed to the WGS84 coordinate system. The preprocessed remote sensing data is used as a long-interval low-sampling sequence, and the preprocessed ground monitoring data is used as a short-interval high-sampling sequence. The two types of sequences are then non-linearly time-aligned. The Euclidean distance matrix between the two aligned sequences is calculated using a dynamic time warping algorithm, and dynamic programming is used to find the optimal path with the minimum cumulative distance. Based on the optimal path, the remote sensing data is interpolated to generate remote sensing sequence data with the same temporal resolution as the ground monitoring data; In the WGS84 coordinate system, each remote sensing sequence data is assigned to the corresponding ground sensor within the geographic boundary of the pixel to generate the multi-source spatiotemporal data.

3. The method according to claim 1, characterized in that, The remote sensing data includes optical remote sensing data, thermal infrared remote sensing data, and microwave remote sensing data; the improved asymmetric autoencoder includes a multi-branch coding layer, a shared fusion layer, and a unified decoding layer; the multi-branch coding layer includes an optical coding branch, a thermal infrared coding branch, and a microwave coding branch; The multi-branch coding layer is used to input the optical remote sensing data, the thermal infrared remote sensing data, and the microwave remote sensing data into corresponding branch coding channels. Each branch coding channel includes a cascaded convolutional feature extraction module and a spatial attention enhancement module. The convolutional feature extraction module is used to extract deep spatial features of each remote sensing data through a multi-layer convolutional neural network to obtain a high-dimensional feature map. The spatial attention enhancement module is used to optimize the spatial dimension weights of the high-dimensional feature map to enhance the feature response of key areas and obtain an enhanced feature map. Ground monitoring data is introduced as a conditional input into the spatial attention module of each branch to adjust the attention weights. The shared fusion layer is used to align the enhanced feature maps of each branch by feature map size and then stitch them together along the channel dimension to obtain a fused feature map. The unified decoding layer is used to perform global pooling and fully connected mapping on the fused feature map to generate the fused feature vector.

4. The method according to claim 1, characterized in that, The step of inputting the fused feature vector into the trained STCN-LSTM model for water demand prediction includes: The fused feature vector is input into the STCN network, and the fused feature vector is processed in parallel using 1×1, 3×3, and 5×5 convolutional kernels; wherein, the 1×1 convolutional kernel is used to extract the local water demand distribution feature map of individual crop plants or micro-regions; the 3×3 convolutional kernel is used to extract the plot-level water demand distribution feature map; and the 5×5 convolutional kernel is used to extract the macro-level water demand trend feature map of farmland. The local water demand distribution feature map, the plot-level water demand distribution feature map, and the macro water demand trend feature map are spliced ​​together at the channel level, and the spliced ​​feature map is convolved using a 1×1 convolution kernel to obtain a multi-level spatial scale water demand feature matrix. The multi-level spatial scale water demand feature matrix is ​​arranged into time series data according to the time dimension and input into a bidirectional LSTM network to obtain a time feature vector. Water demand is predicted based on the multi-level spatial scale water demand feature matrix and the time feature vector to obtain the predicted value of water demand for farmland.

5. The method according to claim 4, characterized in that, The forward LSTM layer of the bidirectional LSTM network processes the time-series data in chronological order to learn the historical cumulative trend of water demand; the reverse LSTM layer of the bidirectional LSTM network processes the time-series data in reverse chronological order to trace back the driving factors of historical meteorological and soil conditions on water demand. The outputs of the forward LSTM layer and the reverse LSTM layer are concatenated along the feature dimension and fused through a fully connected layer to obtain a time feature vector containing historical trends and future predictions.

6. The method according to claim 5, characterized in that, The output of the STCN-LSTM model also includes a spatiotemporal attention gating mechanism; the water demand prediction based on the multi-level spatial scale water demand feature matrix and the temporal feature vector to obtain the predicted water demand value for farmland includes: Using the aforementioned spatiotemporal attention gating mechanism, a temporal attention weight vector is generated based on the crop growth period calendar and the leaf area index time series data retrieved from remote sensing. Using the spatiotemporal attention gating mechanism, global flat pooling and Sigmoid activation function are applied to the multi-level spatial scale water demand feature map to calculate the spatial attention weight matrix. The temporal attention weight vector is multiplied by the spatial attention weight matrix to generate a spatiotemporal attention weight matrix; and then multiplied element-wise with the temporal feature vector to obtain the water demand prediction feature enhanced by spatiotemporal attention. The water demand prediction features are input into the fully connected layer and mapped to the predicted water demand value of farmland.

7. The method according to claim 1, characterized in that, The crop physiological parameters include leaf area index (LAI) retrieved from remote sensing and temperature and humidity measured by ground weather stations; the method also includes: The crop physiological parameters were calculated based on the DSSAT crop growth model. The stomatal conductance correction coefficient is calculated in real time based on the crop physiological parameters; and the stomatal conductance correction coefficient is used as an adjustment factor to dynamically adjust the weights of the input gate or forget gate of the bidirectional LSTM network.

8. The method according to claim 1, characterized in that, The uncertainty quantification of the predicted water demand of the target farmland, and the spatial and temporal error corrections of the quantification results using Kriging interpolation optimization and the sliding window method respectively, generate target water demand values ​​for farmland carrying confidence intervals, including: The fused feature vector is input into the preset STCN-LSTM model for N forward propagation predictions to obtain N water demand prediction values; where N is an integer greater than 1. The mean of the calculated N water demand forecasts is used as the initial forecast baseline value; the variance of the calculated N water demand forecasts is used as the uncertainty threshold for the forecast. The preliminary prediction baseline value is compared with the uncertainty threshold. If the preliminary prediction baseline value is greater than the uncertainty threshold, the current prediction uncertainty is determined to be high, and the subsequent hierarchical error correction process is triggered. If the preliminary prediction benchmark value is less than or equal to the uncertainty threshold, the current prediction uncertainty is determined to be low, and the preliminary prediction benchmark value is directly output to the result verification stage. When the hierarchical error correction process is triggered, the following operations are performed: Based on the N predicted water demand values, the spatial residuals are fitted using the Kriging interpolation optimization algorithm to generate a spatial error correction field, and the preliminary prediction benchmark values ​​are spatially corrected to obtain spatially optimized prediction values. The sliding window method is used to perform a local weighted average of the spatial optimization prediction values ​​in the time dimension to smooth time series noise and obtain time optimization prediction values. Based on the uncertainty threshold, a confidence interval is assigned to the time-optimized predicted value, and the target value of farmland water demand carrying the confidence interval is generated.

Images