Agricultural drought monitoring and prediction system based on unmanned aerial vehicle microwave remote sensing and deep learning

Abstract

The present application relates to the technical field of agricultural drought information monitoring, and more particularly to an agricultural drought monitoring and prediction system based on unmanned aerial vehicle microwave remote sensing and deep learning, which comprises: a microwave remote sensing data fusion subsystem that carries multi-frequency microwave sensors on an unmanned aerial vehicle; dynamic adjustment of feature weights based on crop growth stages; elimination of geometric deviations between unmanned aerial vehicle flight strips through time series registration, and construction of a three-dimensional drought feature field by fusing multi-temporal data; a deep drought diagnosis subsystem that inputs the three-dimensional drought feature field into a space-time convolution network, extracts field patch-level anomalies in the spatial domain, and captures drought evolution patterns in the time domain; a drought response subsystem that divides drought levels based on the diagnosis results of the space-time convolution network; generation of differentiated irrigation schemes in combination with farmland Internet of Things data; and comparison of actual irrigation effects with prediction results. The present application realizes an upgrade of the agricultural drought prevention and control mode from passive monitoring to active prediction, and from extensive management to precise regulation.

Description

Technical Field

[0001] This invention relates to the field of agricultural drought information monitoring technology, and in particular to an agricultural drought monitoring and prediction system based on UAV microwave remote sensing and deep learning. Background Technology

[0002] Agricultural drought is a core threat to global food security. Traditional monitoring methods rely on weather station data and manual field surveys, much like trying to observe a desert with a magnifying glass—inefficient and limiting in scope. While satellite remote sensing can monitor large areas, it suffers from the dilemma of "it's too late when details become visible." The Landsat series satellites' 16-day revisit cycle is as slow as an old-fashioned clock, and MODIS data is like severe myopia—unable to see details in the fields clearly. Agricultural production requires real-time monitoring solutions that can "see both the forest and the trees."

[0003] Prior art 1, Chinese patent application number: 202410844745.X, discloses a daily agricultural drought prediction method based on drought propagation process, including the following steps: collecting and preprocessing daily precipitation, temperature, and soil moisture data of the target prediction area; accumulating data on a daily time scale to calculate meteorological drought index sequences and agricultural drought index sequences; calculating the time lag correlation between meteorological drought and agricultural drought based on the meteorological and agricultural drought index sequences to determine the drought propagation time under different crop growth stages, obtaining multi-period leading-lag meteorological drought indices under different crop growth stage lag times; constructing and training an agricultural drought prediction model ConvLSTM-SENet-Conv to predict the current agricultural drought index of the target prediction area. Although this method fully considers the crop growth process, quantifies drought propagation, incorporates previous meteorological drought data, and fills the gaps in soil data to mitigate the impact of continuous drought early warning, resulting in higher prediction accuracy, it is still difficult to achieve high spatiotemporal resolution monitoring of the entire farmland area due to limitations in station distribution density and data update frequency, as it relies on daily precipitation, temperature, and other ground sensor data.

[0004] Prior art two, Chinese patent application number 202411563821.6, discloses an agricultural drought prediction method and system integrating spatiotemporal feature learning, including: data collection and preprocessing; constructing a drought prediction model, including a spatial feature extraction module, a spatiotemporal feature learning module, and a sequence prediction module; constructing a drought event identification and prediction model; the drought event identification and prediction model analyzes the standardized soil moisture index (SMI) sequence for the next M consecutive days predicted by the sequence prediction module, extracts drought events, analyzes the spatial centroid movement characteristics of severe drought events, and predicts the future spatiotemporal changes of drought. Although it comprehensively considers the temporal and spatial attributes of drought, utilizes high-precision raster meteorological data, and uses multiple hydrometeorological elements as inputs to a pre-trained convolutional neural network coupled with a deep learning model to predict future drought indicators, it can efficiently provide early warning of drought while maintaining the model complexity within an appropriate range, providing technical guidance for drought prevention and water resource management; however, relying on high-precision raster meteorological data (such as 1km resolution) makes it difficult to capture drought anomalies at the farmland patch level (<100m), and cannot support precise irrigation decisions.

[0005] Prior art three, Chinese patent application number 202411904432.5, discloses a method for monitoring drought in alpine grasslands based on localized verification of the Palmer drought index. The method includes: calculating the daily meteorological drought index of the alpine grassland based on average daily precipitation data, and combining this with average daily soil moisture data. Based on drought level classification results, the accuracy of drought monitoring for different meteorological drought indices in alpine grasslands is calculated to obtain the Palmer drought index most suitable for monitoring drought in alpine grasslands. The correlation between the Palmer drought index and soil moisture is analyzed using the Pearson correlation coefficient method. A linear regression model between the Palmer drought index and soil moisture is constructed, and a linear fitting algorithm is used to fit the data, minimizing the difference between predicted and actual values. The coefficients in the linear regression model are calculated, and the Palmer drought index is verified based on these coefficients for drought monitoring in alpine grasslands. While this method effectively improves the accuracy of drought monitoring in alpine grasslands, the Palmer index verification method designed for alpine grasslands is difficult to directly apply to irrigated farmland.

[0006] Current technologies 1, 2, and 3 suffer from insufficient high-resolution dynamic data acquisition, low efficiency in multi-dimensional spatiotemporal feature fusion, and a lack of a closed-loop mechanism for early warning and intervention. Therefore, this invention provides an agricultural drought monitoring and prediction system based on UAV microwave remote sensing and deep learning. Summary of the Invention

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] One aspect of the present invention provides an agricultural drought monitoring and prediction system based on unmanned aerial vehicle (UAV) microwave remote sensing and deep learning, comprising:

[0009] The microwave remote sensing data fusion subsystem is used to simultaneously acquire polarization scattering matrix and interferometric coherence data by using a multi-band microwave sensor mounted on an UAV; dynamically adjust feature weights based on crop growth stages; eliminate geometric deviations between UAV flight strips through temporal registration; and fuse multi-temporal data to construct a three-dimensional drought feature field.

[0010] The deep drought diagnostic subsystem is used to input a three-dimensional drought feature field into a spatiotemporal convolutional network to extract farmland patch-level anomalies in the spatial domain and capture drought evolution patterns in the temporal domain.

[0011] The drought response subsystem is used to classify drought levels based on the diagnostic results of spatiotemporal convolutional networks and reversely locate key drought-causing factors; combine farmland IoT data to generate differentiated irrigation plans; and compare the actual irrigation effect with the prediction results to automatically correct the drought judgment threshold of the spatiotemporal convolutional network.

[0012] In one optional implementation, the microwave remote sensing data fusion subsystem includes:

[0013] The microwave data acquisition and processing module is used to acquire C, L, and X band data from multi-band microwave sensors, record polarization scattering matrices and interferometric coherence information, and perform spatiotemporal alignment of data from different frequency bands; perform radiometric calibration on C, L, and X band data; and correct geometric offsets between flight zones using a positioning system.

[0014] The adaptive feature extraction module is used to divide the monitoring area into key stages such as seedling stage, jointing stage, heading stage and maturity stage by combining historical agricultural data, and to dynamically allocate weights according to the key stages.

[0015] The temporal registration module is used to align the C, L, and X band data period by period using the initial C, L, and X band data as a reference, and eliminate the influence of UAV heading deviation and attitude fluctuation. It calculates the temporal coherence of the registered C, L, and X band data and removes outliers. It stacks the corrected microwave features of each period in a time series to form a three-dimensional drought feature field.

[0016] In one alternative implementation, the adaptive feature extraction module focuses on the following stages: seedling stage: topsoil moisture is emphasized, with high-resolution C-band data dominating; jointing to heading stage: vegetation structure parameters are enhanced, with L-band penetration data dominating, combined with InSAR monitoring of crop height changes; maturity stage: soil-vegetation mixed signals are balanced, with X-band providing detailed texture.

[0017] In one optional implementation, the timing registration module includes:

[0018] The feature point matching submodule is used to identify corresponding feature points in the C, L and X band data of each period by utilizing the natural features in the monitoring area, calculating the spatial offset, and adjusting the geometric position of the C, L and X band data period by period.

[0019] The temporal coherence calculation submodule is used to calculate the interferometric coherence at the same spatial location on the C, L, and X band data of each period after registration, and to evaluate the stability of data at different time points; if the coherence of C, L, and X band data of a certain period is significantly lower than the threshold, it is judged as an outlier and removed.

[0020] The microwave feature standardization submodule is used to extract key microwave features from the C, L and X band data of each period after registration and outlier removal, and to normalize them according to the crop growth stage.

[0021] The feature field construction submodule is used to stack the standardized microwave features from each period in chronological order to form a three-dimensional data field of space, time, and features.

[0022] In one optional implementation, data from the same location at different time points in the spatial dimension can be compared; the temporal dimension dynamically adjusts the weights according to the crop growth stage to reflect the changes in drought characteristics at different phenological stages; and the characteristic dimension integrates multi-band microwave parameters to form a comprehensive drought index.

[0023] In one optional implementation, the deep drought diagnostic subsystem includes:

[0024] The structured parsing module is used to take the three-dimensional drought feature field as the input data of the spatiotemporal convolutional network. The spatial dimension of the three-dimensional drought feature field has been temporally registered to ensure the consistency of geographic coordinates, the temporal dimension is dynamically weighted according to the crop growth stage, and the feature dimension integrates multi-band scattering and interference parameters.

[0025] The spatial anomaly detection module is used to slide the analysis window on the farmland spatial grid, extract the microwave feature statistics of each pixel within the window, and quantify the feature dispersion of the local area.

[0026] The temporal pattern capture module is used to slice the three-dimensional drought feature field along the time axis, extract temporal feature curves at the locations of spatially anomalous patches, suppress high-frequency noise through moving averages, and highlight the trend changes of drought.

[0027] The stage inflection point detection module is used to analyze the slope changes and curvature extremes of time-series curves according to crop growth stages, identify drought acceleration or mitigation periods, establish statistical correlations between early drought characteristics and later evolution, and predict drought chain reaction pathways.

[0028] In one optional implementation, the airspace anomaly detection module compares the characteristic statistics of adjacent farmland patches within the same growth stage. If the characteristics of a patch deviate from the threshold range of the surrounding area, it is marked as a potential drought anomaly area. By combining historical agricultural data to exclude interference from non-drought factors, drought-related abnormal patches are confirmed.

[0029] In one optional implementation, the airspace anomaly detection module includes:

[0030] The window space constraint analysis submodule is used to set a rectangular analysis window of a fixed size on the farmland grid based on the spatial dimension registered by the three-dimensional drought feature field. The size of the rectangular analysis window is determined according to the crop planting density and the resolution of the microwave sensor.

[0031] The multi-feature synchronous extraction submodule is used to calculate three types of statistics in parallel within each rectangular analysis window: central trend, dispersion, and extreme value difference.

[0032] The adaptive normalization submodule is used to perform stage-specific corrections for the three types of statistics—vegetative growth stage, reproductive growth stage, and maturity stage—by referencing the dynamic weights of crop growth stages over time.

[0033] The composite discrete index construction submodule is used to generate local discrete indices by linearly combining the corrected statistics. Spatial consistency is tested using the discrete indices of adjacent rectangular analysis windows. Isolated rectangular analysis windows with abrupt changes in discrete indices are marked as suspicious areas. The original discrete index is retained for areas that show gradient changes in three or more consecutive rectangular analysis windows. Combined with historical agricultural data from the spatial anomaly detection module, non-drought interference such as agricultural machinery operation trajectories is excluded.

[0034] In one optional implementation, the central tendency is the arithmetic mean of the backscattering intensity of all pixels within the window; the dispersion is the standard deviation of the coherence coefficients within the window; and the extreme value difference is the difference between the largest and smallest eigenvalues ​​of the polarization scattering matrix.

[0035] The vegetative growth stage contributes to the dispersion of the coherence coefficient; the reproductive growth stage enhances the weight of the extreme differences in polarization characteristics; and the maturation stage focuses on the central trend changes in backscattering intensity.

[0036] In one alternative implementation, the drought response subsystem includes:

[0037] The spatiotemporal feature decoding module is used to decompose the three-dimensional drought feature field output by the spatiotemporal convolutional network into the spatial dimension of farmland patch anomaly distribution and the temporal dimension of drought evolution trend; based on the gradient response of the features of each layer of the spatiotemporal convolutional network, the contribution of polarization scattering matrix, coherence coefficient and backscattering intensity to the diagnostic results is quantified, and the core features that dominate drought discrimination are identified.

[0038] The dynamic drought level classification module is used to classify regions into mild, moderate and severe drought levels based on the spatial continuity of abnormal intensity of farmland patches; combined with the historical drought event database, it matches the historical level labels of similar patterns to the current drought evolution pattern to correct the preliminary classification results;

[0039] The drought-causing factor reverse tracing module is used in polarization-dominant areas. If the abnormal contribution of the polarization scattering matrix eigenvalues ​​exceeds the threshold, it is located to vegetation structure variation or soil surface cracks. In coherence-dominant areas, when the coherence coefficient is abnormally high, it is associated with uneven canopy cover or differences in root water absorption. A continuous decrease in temporal coherence indicates soil water-holding capacity degradation. In scattering intensity-dominant areas, abrupt changes in the mean backscattering value are associated with a sharp drop in surface soil moisture content or interference from agricultural machinery operations.

[0040] The threshold adaptive correction module is used to compare the changes in the feature field after actual irrigation with the prediction results, adjust the drought sensitivity threshold of the spatiotemporal convolutional network for misjudged areas, and dynamically update the weights of the feature-factor association mapping based on the accuracy of drought-causing factor localization.

[0041] This invention utilizes a multi-band microwave remote sensing data fusion subsystem to achieve all-weather, high spatiotemporal resolution extraction of farmland drought features, overcoming the limitations of traditional optical remote sensing, which is susceptible to weather conditions and has long monitoring cycles. It dynamically adjusts feature weights based on crop growth stages, ensuring more targeted drought monitoring indicators for different growth stages and reducing misjudgments. The deep drought diagnosis subsystem employs a spatiotemporal convolutional network to simultaneously extract spatial anomalies (such as sudden drops in local soil moisture) and temporal evolution patterns (such as continuous drought trends) from a three-dimensional drought feature field, improving the ability to identify early signs of drought. Combined with farmland patch-level analysis, it avoids the problem of missed detection of localized droughts caused by traditional large-scale mean statistics. The drought response subsystem accurately classifies drought levels based on diagnostic results and reverse-locates key drought-causing factors (such as insufficient irrigation or poor soil water retention), providing a basis for precision agricultural management. It generates differentiated irrigation plans based on real-time IoT data, avoiding water waste caused by traditional uniform irrigation. Automatic correction of model thresholds through feedback from actual irrigation effects enables the system to continuously optimize and improve long-term prediction reliability. Attached Figure Description

[0042] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0043] Figure 1 This is a block diagram of the agricultural drought monitoring and prediction system based on UAV microwave remote sensing and deep learning provided in Embodiment 1 of the present invention;

[0044] Figure 2This is a block diagram of the microwave remote sensing data fusion subsystem provided in Embodiment 2 of the present invention;

[0045] Figure 3 This is a block diagram of the deep drought diagnostic subsystem provided in Embodiment 4 of the present invention;

[0046] Figure 4 This is a block diagram of the drought response subsystem provided in Embodiment 8 of the present invention;

[0047] Figure 5 A block diagram of the electronic device provided by the present invention;

[0048] Figure 6 A block diagram of a computer-readable storage medium provided for this invention. Detailed Implementation

[0049] The technical solutions of the present invention will now be described with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0050] Hereinafter, the terms "first," "second," etc., are used for descriptive convenience only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0051] In this invention, unless otherwise explicitly specified and limited, the term "connection" should be interpreted broadly. For example, "connection" can be a fixed mechanical connection, a detachable mechanical connection, or an integral part; or, "connection" can be a direct connection or an indirect connection through an intermediate medium. Furthermore, unless otherwise explicitly specified and limited, the term "coupling" should be interpreted broadly. For example, "coupling" can be a direct electrical connection, such as physical contact and electrical conduction between two components; it can also be understood as an electrical connection between different components in a circuit structure through physical lines capable of transmitting electrical signals, such as copper foil or wires on a printed circuit board (PCB), to transmit electrical signals; or, "coupling" can be an indirect electrical connection between two components through an intermediate medium; or, "coupling" can be an electrical connection between two components in a non-contact manner, such as an electrical connection between two components using capacitive coupling to transmit electrical signals.

[0052] In this embodiment of the invention, directional terms such as "up," "down," "left," and "right" may be defined relative to the orientation of the components shown in the accompanying drawings. It should be understood that these directional terms can be relative concepts, used for relative description and clarification, and can change accordingly depending on the orientation of the components in the accompanying drawings.

[0053] Example 1:

[0054] like Figure 1 As shown, this embodiment of the invention provides an agricultural drought monitoring and prediction system based on UAV microwave remote sensing and deep learning, comprising:

[0055] The microwave remote sensing data fusion subsystem is used to acquire polarization scattering matrix and interferometric coherence data simultaneously by using a multi-band microwave sensor mounted on an UAV; it dynamically adjusts feature weights based on crop growth stages, eliminates geometric deviations between UAV flight strips through temporal registration, and fuses multi-temporal data to construct a three-dimensional drought feature field;

[0056] The deep drought diagnostic subsystem is used to input a three-dimensional drought feature field into a spatiotemporal convolutional network to extract farmland patch-level anomalies in the spatial domain and capture drought evolution patterns in the temporal domain.

[0057] The drought response subsystem is used to classify drought levels based on the diagnostic results of spatiotemporal convolutional networks and reversely locate key drought-causing factors; combine farmland IoT data to generate differentiated irrigation plans; and compare the actual irrigation effect with the prediction results to automatically correct the drought judgment threshold of the spatiotemporal convolutional network.

[0058] In the above embodiments, this embodiment utilizes a multi-band microwave remote sensing data fusion subsystem to achieve all-weather, high spatiotemporal resolution extraction of farmland drought features, overcoming the limitations of traditional optical remote sensing, which is susceptible to weather conditions and has long monitoring cycles. It dynamically adjusts feature weights based on crop growth stages to ensure more targeted drought monitoring indicators for different growth stages, reducing misjudgments. The deep drought diagnosis subsystem uses a spatiotemporal convolutional network to simultaneously extract spatial anomalies (such as sudden drops in local soil moisture) and temporal evolution patterns (such as continuous drought trends) from a three-dimensional drought feature field, improving the ability to identify early signs of drought. Combined with farmland patch-level analysis, it avoids the problem of missed detection of local droughts caused by traditional large-scale mean statistics. The drought response subsystem accurately classifies drought levels based on the diagnostic results and reverse-locates key drought-causing factors (such as insufficient irrigation or poor soil water retention), providing a basis for precision agricultural management. It generates differentiated irrigation plans based on real-time IoT data, avoiding water waste caused by traditional uniform irrigation. Automatic correction of model thresholds through feedback from actual irrigation effects enables the system to continuously optimize and improve long-term prediction reliability.

[0059] In summary, this embodiment forms a closed-loop data flow: microwave remote sensing data fusion provides input, deep drought diagnosis analyzes patterns, and the drought response subsystem outputs decisions and provides feedback for optimization, ensuring the overall adaptability of the system; ultimately achieving an upgrade in agricultural drought prevention and control model from passive monitoring to active prediction, and from extensive management to precise regulation.

[0060] Example 2:

[0061] like Figure 2 As shown, based on Embodiment 1, the microwave remote sensing data fusion subsystem provided in this embodiment of the invention includes:

[0062] The microwave data acquisition and processing module is used to acquire C, L, and X band data from multi-band microwave sensors, record polarization scattering matrices and interferometric coherence information respectively, and align data from different frequency bands in time and space; perform radiometric calibration on C, L, and X band data to eliminate sensor gain differences; and correct geometric offsets between flight zones through a positioning system to ensure spatial consistency of data.

[0063] The adaptive feature extraction module is used to divide the monitoring area into key stages such as seedling stage, jointing stage, heading stage and maturity stage by combining historical agricultural data (such as sowing date and variety growth period), and to dynamically allocate weights according to the key stages.

[0064] Among them, during the seedling stage: the focus is on surface soil moisture (dominated by high-resolution C-band data) to suppress vegetation scattering interference; during the jointing to heading stage: vegetation structure parameters are enhanced (dominated by L-band penetration data) and crop height changes are monitored in conjunction with InSAR; during the maturity stage: the soil-vegetation mixed signal is balanced (X-band provides detailed texture) to avoid misjudgment of soil moisture caused by canopy shading.

[0065] The temporal registration module is used to align the C, L, and X band data period by period using the initial C, L, and X band data as a reference, through feature point matching (such as field edges and irrigation facilities), to eliminate the influence of UAV heading offset and attitude fluctuation; calculate the temporal coherence of the registered C, L, and X band data and remove outliers; and stack the corrected microwave features (such as backscattering coefficient and coherence coefficient) of each period in a time series to form a three-dimensional drought feature field.

[0066] In the above embodiments, this embodiment solves the gain difference and spatial offset problem between multi-band (C / L / X band) sensors through radiometric calibration and geometric correction, ensuring the comparability of data from different frequency bands in terms of radiation magnitude and spatial location, laying a physical foundation for subsequent fusion analysis. Based on a dynamic weight allocation mechanism for crop growth stages (seedling to maturity), precise matching of monitoring strategies and crop physiological characteristics is achieved; through complementary advantages of frequency bands (C band surface moisture, L band vegetation structure, X band texture details), mutual interference between soil and vegetation signals in a single growth stage is effectively suppressed. Using feature point matching and temporal coherence screening, the instability of the UAV platform and the influence of environmental noise are eliminated, constructing a continuously comparable three-dimensional drought feature field (space × time × microwave parameters) in the time dimension, providing a temporally stable data foundation for drought evolution analysis. The final output three-dimensional feature field integrates multi-band scattering characteristics (backscattering coefficient), interference information (coherence coefficient), and phenological stage characteristics, realizing a unified expression of multiple parameters such as soil moisture, vegetation structure, and surface texture in the spatiotemporal dimension, which significantly improves the accuracy and robustness of agricultural drought monitoring.

[0067] Example 3:

[0068] Based on Embodiment 2, the timing registration module provided in this embodiment of the invention includes:

[0069] The feature point matching submodule is used to identify corresponding feature points in the C, L and X band data of each period by utilizing natural features (such as field edges and irrigation facilities) in the monitoring area, calculate the spatial offset, and adjust the geometric position of the C, L and X band data period by period.

[0070] The temporal coherence calculation submodule is used to calculate the interferometric coherence at the same spatial location on the C, L, and X band data of each period after registration, and to evaluate the stability of data at different time points; if the coherence of C, L, and X band data of a certain period is significantly lower than the threshold, it is judged as an outlier and removed.

[0071] The microwave feature standardization submodule is used to extract key microwave features (such as backscattering coefficient and coherence coefficient) from the C, L and X band data of each stage after registration and outlier removal, and to normalize them according to the crop growth stage (seedling stage, jointing stage, heading stage and maturity stage).

[0072] The feature field construction submodule is used to stack the standardized microwave features of each period in chronological order to form a three-dimensional data field of space (two-dimensional) - time (one-dimensional) - features (multi-parameters);

[0073] Specifically: the spatial dimension maintains the consistency of the corrected geographic coordinates to ensure that data from the same location at different time points can be compared; the temporal dimension dynamically adjusts the weights according to the crop growth stage (e.g., C-band dominates during the seedling stage, and X-band supplements during the maturity stage) to reflect the changes in drought characteristics at different phenological stages; the characteristic dimension integrates multi-band microwave parameters (e.g., scattering intensity, coherence) to form a comprehensive drought index, avoiding misjudgments based on a single frequency band or time point.

[0074] In the above embodiments, this embodiment uses a feature point matching submodule to periodically correct the spatial offset of C, L, and X band data, ensuring that microwave remote sensing data acquired at different times are strictly aligned geometrically, eliminating the impact of unmanned aerial vehicle platform instability factors (such as heading deviation and attitude fluctuations) on the observation results. The temporal coherence calculation submodule filters valid observation periods based on interferometric stability, eliminating data anomalies caused by environmental interference (such as precipitation and strong winds), ensuring that subsequent analysis is based only on highly reliable time-series datasets. The microwave feature standardization submodule eliminates the inherent influence of phenological differences on microwave features (such as backscattering coefficients) through growth stage-dependent normalization processing, enabling comparative analysis of data from seedling to maturity under a unified standard. The feature field construction submodule generates a three-dimensional data field (space × time × feature), achieving the following: the spatial dimension maintains the corrected geographic coordinate framework, supporting pixel-level temporal change tracking; the temporal dimension highlights key parameters of each phenological stage through adaptive weighting of growth stages (dynamically dominated by C / L / X bands); and the feature dimension integrates multi-band scattering and interferometric characteristics to construct an interference-resistant composite drought index.

[0075] Example 4:

[0076] like Figure 3 As shown, based on Example 1, the deep drought diagnostic subsystem provided in this embodiment of the invention includes:

[0077] The structured parsing module is used to take the three-dimensional drought feature field as the input data of the spatiotemporal convolutional network. The spatial dimension of the three-dimensional drought feature field has been temporally registered to ensure the consistency of geographic coordinates, the temporal dimension is dynamically weighted according to the crop growth stage, and the feature dimension integrates multi-band scattering and interference parameters.

[0078] The spatial anomaly detection module is used to slide the analysis window on the farmland spatial grid, extract the microwave characteristics (such as backscattering intensity and coherence coefficient) statistics (mean, variance, range) of each pixel within the window, and quantify the feature dispersion of the local area; compare the feature statistics of adjacent farmland patches within the same growth stage, and if the features of a patch deviate from the threshold range of the surrounding area (such as a sudden drop in scattering intensity or a sudden increase in coherence), it is marked as a potential drought anomaly area; combine historical agricultural data (such as soil type and irrigation records) to eliminate interference from non-drought factors (such as crop lodging and field operation traces) and confirm drought-related abnormal patches;

[0079] The temporal pattern capture module is used to slice the three-dimensional drought feature field along the time axis and extract temporal feature curves (such as monthly changes in scattering coefficients and cumulative coherence) at the locations of anomalous patches in the spatial domain. It suppresses high-frequency noise through moving averages and highlights the trend changes in drought.

[0080] The stage inflection point detection module is used to analyze the slope changes and curvature extremes of time-series curves according to crop growth stages, identify drought acceleration or mitigation periods, establish statistical correlations between early drought characteristics and later evolution, and predict drought chain reaction pathways.

[0081] In the above embodiments, the structured analysis module ensures that the input data of the three-dimensional drought feature field (space × time × feature) possesses geographical consistency, growth stage adaptability, and multi-frequency band fusion, providing a standardized data foundation for subsequent analysis. The spatial anomaly detection module, through local statistical analysis and contextual verification, eliminates non-drought interference factors, accurately identifies drought-related farmland patch-level anomaly areas, and avoids misjudgments. The temporal pattern capture module extracts temporal feature curves at the locations of spatial anomaly patches, suppresses noise interference, highlights the long-term trend of drought, and enhances the temporal continuity of drought monitoring. The stage inflection point detection module, combined with crop growth stage division, captures key nodes of drought acceleration or mitigation and establishes a correlation model between early characteristics and later evolution to predict drought development paths.

[0082] Example 5:

[0083] Based on Example 4, the airspace anomaly detection module provided in this embodiment of the invention includes:

[0084] The window space constraint analysis submodule is used to set a rectangular analysis window of a fixed size on the farmland grid based on the spatial dimension registered by the three-dimensional drought feature field. The size of the rectangular analysis window is determined according to the crop planting density and the resolution of the microwave sensor.

[0085] The multi-feature synchronous extraction submodule is used to calculate three types of statistics in parallel within each rectangular analysis window: central trend, dispersion, and extreme value difference.

[0086] Among them, the central tendency is the arithmetic mean of the backscattering intensity of all pixels within the window; the dispersion is the standard deviation of the coherence coefficient within the window; and the extreme value difference is the difference between the largest and smallest eigenvalues ​​of the polarization scattering matrix.

[0087] The adaptive normalization submodule is used to perform stage-specific corrections for the three types of statistics—vegetative growth stage, reproductive growth stage, and maturity stage—by referencing the dynamic weights of crop growth stages over time.

[0088] The vegetative growth stage contributes to the dispersion of the coherence coefficient; the reproductive growth stage increases the weight of extreme differences in polarization characteristics; the maturity stage emphasizes the central trend changes in backscattering intensity.

[0089] The composite discrete index construction submodule is used to generate local discrete indices by linearly combining the corrected statistics; spatial consistency is tested by using the discrete indices of adjacent rectangular analysis windows, isolated rectangular analysis windows with abrupt changes in discrete indices are marked as suspicious areas, and the original discrete index is retained for areas that show gradient changes in three or more consecutive rectangular analysis windows. Combined with historical agricultural data from the spatial anomaly detection module, non-drought interference such as agricultural machinery operation trajectories is excluded.

[0090] Window dispersion index = (coherence standard deviation × growth stage weight) + (polarization range × frequency band coefficient) - (scattering mean stability factor); the frequency band coefficient comes from the multi-band interference parameters fused from the feature dimensions.

[0091] In the above embodiments, the spatial anomaly detection module achieves accurate identification of arid farmland areas through multi-level collaborative analysis; the window spatial constraint analysis submodule ensures that the analysis unit matches the crop planting pattern and sensor observation capabilities by dividing the window into gridded rectangular windows, thus solving the problem of spatial scale adaptation; the spatial consistency verification mechanism in the construction of the composite discrete index effectively distinguishes between real arid areas and isolated anomalies (such as agricultural machinery interference). The multi-feature synchronous extraction submodule combines three complementary features—backscattering intensity (central tendency), coherence coefficient (dispersion), and polarization characteristics (extreme value differences)—to construct a multi-dimensional drought characterization system; the adaptive normalization submodule solves the nonlinear interference of crop phenological changes on feature responses through dynamic weight adjustment of growth stages. The growth stage-specific correction strategy (emphasizing coherence during the vegetative stage, focusing on polarization differences during the reproductive stage, and depending on scattering intensity during the maturity stage) achieves time-series adaptive detection, and comparison with historical agricultural data excludes the influence of persistent non-drought factors. The final output local discrete index integrates spatial-temporal-feature three-dimensional information. Its mutation detection results can reflect the vegetation water stress caused by the decrease in soil moisture content, while effectively suppressing interference factors such as crop phenological changes and agricultural activities; thus realizing quantitative and localized monitoring of farmland drought.

[0092] Example 6:

[0093] Based on Example 5, the multi-feature synchronous extraction submodule provided in this embodiment of the invention includes:

[0094] The data slicing and window loading unit is used to extract the corresponding backscattering intensity matrix, coherence coefficient matrix and polarization scattering matrix from the spatial dimension for each rectangular analysis window based on the registered three-dimensional drought feature field. The arithmetic mean of the backscattering intensity values ​​of all pixels in the window is performed to generate a scalar that characterizes the average scattered energy of the region, reflecting the overall microwave reflectance characteristics of the vegetation canopy and soil background. The calculation result will be used as the benchmark term for the construction of the discrete index.

[0095] The discreteness calculation unit is used to synchronously analyze the standard deviation of the coherence coefficient within the window. The coherence coefficient characterizes the temporal stability of the radar echo signal, while the discreteness is related to the uniformity of vegetation cover. The calculation results are dynamically adjusted by weighting the growth stages to adapt to the sensitivity differences in coherence caused by crop physiological changes.

[0096] The extreme value difference calculation unit is used to decompose the eigenvalues ​​of the polarization scattering matrix, take the algebraic difference between the maximum and minimum eigenvalues, and reveal the anisotropic scattering characteristics of ground targets. It needs to integrate the frequency band coefficients generated by multi-band interferometric parameters. After the three types of statistical quantities are output, they are injected into the buffer queue of the adaptive normalization submodule. The central trend quantity is used to generate the scattering mean stability factor (to suppress background noise). The discreteness quantity and the extreme value difference quantity are respectively bound to the growth stage weights and frequency band coefficients to ensure the spatiotemporal consistency of each parameter in subsequent linear combination.

[0097] In the above embodiments, the multi-feature synchronous extraction submodule of this embodiment achieves refined characterization of farmland drought characteristics through multi-dimensional feature collaborative computation; the data slicing and window loading unit establish a spatial benchmark framework to ensure that backscattering intensity, coherence coefficient, and polarization scattering matrix maintain geometric consistency within a unified spatial unit; three types of statistics (central tendency / dispersion / extreme difference) constitute a complementary feature group: the central tendency reflects the overall reflectance characteristics of the region, the dispersion represents the spatial heterogeneity of vegetation, and the extreme difference captures the anisotropy of ground object scattering. The parallel computing architecture realizes the synchronous extraction of the three types of statistics, avoiding the feature temporal deviation in traditional serial processing; the eigenvalue decomposition of the polarization scattering matrix and the calculation of the standard deviation of the coherence coefficient share the same window data slice, reducing the overhead of repeated data loading. The output statistics directly match the input requirements of the adaptive normalization submodule: the central tendency generates a scattering mean stability factor for noise suppression, the dispersion is pre-bound with growth stage weights, and the extreme difference is pre-fused with frequency band coefficients; the cache queue mechanism ensures the spatiotemporal alignment of statistics, providing parameter consistency guarantees for the subsequent construction of composite discrete indices. The output features triples (mean / standard deviation / range) with clear physical meaning. Their synergistic effect is reflected in the triple verification of microwave scattering intensity distribution characteristics (central tendency), vegetation cover uniformity (dispersion), and land cover structure anisotropy (extreme value difference), providing multidimensional discrimination basis for drought detection.

[0098] Example 7:

[0099] Based on Example 5, the adaptive normalization submodule provided in this embodiment of the invention includes:

[0100] The crop growth stage determination unit is used to classify the current window's growth stage based on phenological data in the time dimension: vegetative growth stage, reproductive growth stage, or maturity stage. The determination result will serve as the benchmark for dynamic weight allocation.

[0101] The growth period correction unit is used for vegetative growth period correction, enhances the contribution weight of the standard deviation of the coherence coefficient, applies a gain coefficient greater than 1 to the dispersion quantity, and amplifies its proportion in the composite dispersion index; maintains the basic weight of polarization range: reduces the sensitivity of backscattering mean, as sparse vegetation cover leads to strong interference from soil background reflection.

[0102] Correction for reproductive growth period: The formation of flower and fruit organs significantly alters the dielectric properties of vegetation, elevating the weight of the product of frequency band coefficients and range to a dominant position; Maintaining the monitoring function of coherence dispersion: Maintaining standard weights as an auxiliary discrimination indicator; Controlling interference from scattering mean: Canopy closure reduces the contribution of soil reflection, but vegetation's own scattering changes gradually.

[0103] Maturity correction enhances the indicative value of backscattering mean: plant structure is stable, and water loss directly reflects a decrease in overall scattering energy, so the weight of the scattering mean stability factor is increased to the highest priority; weakens the influence of polarization characteristics: wilting tissue leads to degradation of anisotropic scattering characteristics, so an attenuation coefficient is applied to the range term; limits the applicability of coherence indicators: leaf abscission makes time-series coherence monitoring ineffective, so the weight of dispersion is reduced to the minimum.

[0104] Weight transition processing uses linear interpolation during the growth stage transition period to avoid weight jumps between adjacent stages, ensure the temporal continuity of the composite discrete index, and calibrate the interpolation parameters based on historical crop data to match the physiological transition characteristics of specific crop varieties.

[0105] The calibration result verification unit is used to verify the rationality of the inversion weights through spatial consistency. The synchronous growth stages of adjacent windows should present a coordinated weight distribution. Abnormal weight configuration will cause spatial discontinuity of the discrete index. The weight threshold is optimized by combining historical drought records to ensure that the calibration strategy at each stage matches the actual drought response.

[0106] In the above embodiments, this embodiment achieves the following: through dynamic adjustment based on growth stage perception, the contribution of the three types of statistics is always optimally matched with the crop's physiological state, ultimately realizing the precise drought detection capability of capturing canopy development abnormalities during the vegetative stage, identifying water stress in flowers and fruits during the reproductive stage, and monitoring overall plant decay during the maturity stage.

[0107] Example 8:

[0108] like Figure 4 As shown, based on Example 1, the drought response subsystem provided in this embodiment of the invention includes:

[0109] The spatiotemporal feature decoding module is used to decompose the three-dimensional drought feature field output by the spatiotemporal convolutional network into the spatial dimension of farmland patch anomaly distribution and the temporal dimension of drought evolution trend; based on the gradient response of the features of each layer of the spatiotemporal convolutional network, the contribution of polarization scattering matrix, coherence coefficient and backscattering intensity to the diagnostic results is quantified, and the core features that dominate drought discrimination are identified.

[0110] The dynamic drought classification module is used to classify regions into mild (local discrete anomalies), moderate (continuous patch anomalies), and severe (region-wide consistent anomalies) drought levels based on the spatial continuity of farmland patch anomaly intensity. It also combines a historical drought event database to match historical level labels of similar patterns with current drought evolution patterns (such as gradual water shortage and sudden water stress) to correct the initial classification results.

[0111] The drought-causing factor reverse tracing module is used in polarization-dominant areas. If the abnormal contribution of the polarization scattering matrix eigenvalues ​​exceeds the threshold, it is located to vegetation structure variations (such as fruit shrinkage, changes in leaf tilt angle) or soil surface fissures. In coherence-dominant areas, when the coherence coefficient is abnormally high, it is associated with uneven canopy cover or differences in root water absorption. A continuous decrease in temporal coherence indicates a degradation of soil water holding capacity. In scattering intensity-dominant areas, abrupt changes in the mean backscattering value are associated with a sharp drop in surface soil moisture content or interference from agricultural machinery operations.

[0112] The threshold adaptive correction module is used to compare the changes in the feature field after actual irrigation with the prediction results, adjust the drought sensitivity threshold of the spatiotemporal convolutional network for misjudged areas, and dynamically update the weights of the feature-factor association mapping based on the accuracy of drought-causing factor localization.

[0113] In the above embodiments, the output feature field of the spatiotemporal convolutional network directly determines the initial boundary of the hierarchical classification, and its spatial resolution and temporal depth affect the granularity of drought-causing factor localization. Irrigation verification data is fed back to the network threshold correction, forming a closed loop of "diagnosis-response-verification-optimization" to ensure that the system continuously adapts to changes in the farmland environment. The competitive relationship of the contribution of polarization, coherence, and scattering features (such as the increase in coherence weight during the vegetative growth period) is reflected in the priority difference of different drought-causing factors in reverse localization. By integrating the interpretability of deep learning features with cross-validation of multi-source data, a closed-loop source tracing from drought phenomenon to its cause is achieved, providing a causal decision-making basis for precision irrigation.

[0114] Figure 5 A block diagram of an exemplary electronic device suitable for implementing embodiments of the present invention is shown.

[0115] Electronic devices may include a central processing unit / microprocessor / main control chip, etc.; a storage medium coupled to the central processing unit / microprocessor / main control chip, etc., and storing computer-executable instructions therein for performing the steps of various methods of embodiments of the present invention when executed by a processor.

[0116] The central processing unit / microprocessor / main control chip, etc., may include, but are not limited to, one or more processors or microprocessors.

[0117] Storage media may include, but are not limited to, random access memory (RAM), read-only memory (ROM), flash memory, EPROM memory, EEPROM memory, registers, and computer storage media (such as hard disks, floppy disks, solid-state drives, removable disks, CD-ROMs, DVD-ROMs, Blu-ray discs, etc.).

[0118] In addition, the electronic device may include (but is not limited to) a data bus, an input / output bus / external bus / device bus, a display, and input / output devices (e.g., keyboard, mouse, speaker, etc.).

[0119] The central processing unit / microprocessor / main control chip, etc., can communicate with external devices via the I / O bus through a wired or wireless network (not shown).

[0120] The storage medium may also store at least one computer-executable instruction for performing the steps of various functions and / or methods in the embodiments described herein when run by a central processing unit / microprocessor / main control chip, etc.

[0121] In one embodiment, the at least one computer-executable instruction may also be compiled into or comprise a software product, wherein one or more computer-executable instructions are executed by a processor to perform the steps of the various functions and / or methods in the embodiments described herein.

[0122] Figure 6 A schematic diagram of a computer-readable storage medium according to an embodiment of the present invention is shown.

[0123] like Figure 6As shown, instructions, such as computer-readable instructions, are stored on a non-transitory computer-readable storage medium. When the computer-readable instructions are executed by a processor, the various methods described above can be performed. The non-transitory computer-readable storage medium includes, but is not limited to, volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and / or cache memory. Non-transitory non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc. For example, the non-transitory computer-readable storage medium can be connected to a computing device such as a computer, and then, when the computing device executes the computer-readable instructions stored on the computer-readable storage medium, the various methods described above can be performed.

[0124] In the several embodiments provided by this invention, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0125] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0126] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0127] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for executing all or part of the steps of the methods of the various embodiments of this invention through a computer device (which may be a personal computer, server, or network device, etc.). The aforementioned storage medium includes: USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, optical disks, and other media capable of storing program code.

[0128] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An agricultural drought monitoring and prediction system based on UAV microwave remote sensing and deep learning, characterized in that, Include: The microwave remote sensing data fusion subsystem is used to simultaneously acquire polarization scattering matrix and interferometric coherence data by using a multi-band microwave sensor mounted on an UAV; and dynamically adjusts feature weights based on crop growth stages. Geometric deviations between UAV flight strips are eliminated through temporal registration, and a three-dimensional drought feature field is constructed by fusing multi-temporal data. The deep drought diagnostic subsystem is used to input a three-dimensional drought feature field into a spatiotemporal convolutional network to extract farmland patch-level anomalies in the spatial domain and capture drought evolution patterns in the temporal domain. The drought response subsystem is used to classify drought levels based on the diagnostic results of the spatiotemporal convolutional network and reversely locate key drought-causing factors; it combines farmland IoT data to generate differentiated irrigation plans; and it compares the actual irrigation effect with the prediction results to automatically correct the drought judgment threshold of the spatiotemporal convolutional network. The deep drought diagnostic subsystem includes a structured parsing module and a spatial anomaly detection module; The structured parsing module is used to take the three-dimensional drought feature field as the input data of the spatiotemporal convolutional network. The spatial dimension of the three-dimensional drought feature field has been temporally registered to ensure the consistency of geographic coordinates, the temporal dimension is dynamically weighted according to the crop growth stage, and the feature dimension integrates multi-band scattering and interference parameters. The spatial anomaly detection module is used to slide the analysis window on the farmland spatial grid, extract the microwave feature statistics of each pixel within the window, and quantify the feature dispersion of the local area. The airspace anomaly detection module includes: The window space constraint analysis submodule is used to set a rectangular analysis window of a fixed size on the farmland grid based on the spatial dimension registered by the three-dimensional drought feature field. The size of the rectangular analysis window is determined according to the crop planting density and the resolution of the microwave sensor. The multi-feature synchronous extraction submodule is used to calculate three types of statistics in parallel within each rectangular analysis window: central trend, dispersion, and extreme value difference. The adaptive normalization submodule is used to perform stage-specific corrections for the three types of statistics—vegetative growth stage, reproductive growth stage, and maturity stage—by referencing the dynamic weights of crop growth stages over time. The composite discrete index construction submodule is used to generate local discrete indices by linearly combining the corrected statistics; spatial consistency is tested by using the discrete indices of adjacent rectangular analysis windows, isolated rectangular analysis windows with abrupt changes in discrete indices are marked as suspicious areas, and the original discrete index is retained for areas that show gradient changes in three or more consecutive rectangular analysis windows. Combined with historical agricultural data, interference from agricultural machinery operation trajectories is eliminated. The multi-feature synchronous extraction submodule includes: The data slicing and window loading unit is used to extract the corresponding backscattering intensity matrix, coherence coefficient matrix and polarization scattering matrix from the spatial dimension for each rectangular analysis window based on the registered three-dimensional drought feature field. Perform an arithmetic mean operation on the backscattering intensity values ​​of all pixels within the window to generate a central trend value; The discreteness calculation unit is used to simultaneously analyze the standard deviation of the coherence coefficient within the window. The extreme value difference calculation unit is used to decompose the eigenvalues ​​of the polarization scattering matrix and take the algebraic difference between the maximum and minimum eigenvalues. The central tendency is used to generate the scattering mean stability factor, while the dispersion and extreme value difference are bound to the growth stage weights and frequency band coefficients, respectively, to ensure the spatiotemporal consistency of each parameter in subsequent linear combinations.

2. The agricultural drought monitoring and prediction system based on UAV microwave remote sensing and deep learning as described in claim 1, characterized in that, The microwave remote sensing data fusion subsystem includes: The microwave data acquisition and processing module is used to acquire C, L and X band data from multi-band microwave sensors, record the polarization scattering matrix and interference coherence information respectively, and align the data of different frequency bands in time and space. Radiometric calibration was performed on C, L, and X band data; geometric offsets between flight zones were corrected using a positioning system; The adaptive feature extraction module is used to divide the monitoring area into key stages such as seedling stage, jointing stage, heading stage and maturity stage by combining historical agricultural data, and to dynamically allocate weights according to the key stages. The temporal registration module is used to align the C, L, and X band data period by period using the initial C, L, and X band data as a reference, and eliminate the influence of UAV heading deviation and attitude fluctuation. It calculates the temporal coherence of the registered C, L, and X band data and removes outliers. It stacks the corrected microwave features of each period in a time series to form a three-dimensional drought feature field.

3. The agricultural drought monitoring and prediction system based on UAV microwave remote sensing and deep learning as described in claim 2, characterized in that, In the adaptive feature extraction module, during the seedling stage, the focus is on surface soil moisture, with high-resolution C-band data dominating; from jointing to heading stage, vegetation structure parameters are enhanced, with L-band penetration data dominating, combined with InSAR monitoring of crop height changes; and during the maturity stage, the soil-vegetation mixed signal is balanced, with X-band providing detailed texture.

4. The agricultural drought monitoring and prediction system based on UAV microwave remote sensing and deep learning as described in claim 2, characterized in that, The timing registration module includes: The feature point matching submodule is used to identify corresponding feature points in the C, L and X band data of each period by utilizing the natural features in the monitoring area, calculating the spatial offset, and adjusting the geometric position of the C, L and X band data period by period. The temporal coherence calculation submodule is used to calculate the interferometric coherence at the same spatial location on the C, L, and X band data of each period after registration, and to evaluate the stability of data at different time points; if the coherence of C, L, and X band data of a certain period is significantly lower than the threshold, it is judged as an outlier and removed. The microwave feature standardization submodule is used to extract key microwave features from the C, L and X band data of each period after registration and outlier removal, and to normalize them according to the crop growth stage. The feature field construction submodule is used to stack the standardized microwave features from each period in chronological order to form a three-dimensional data field of space, time, and features.

5. The agricultural drought monitoring and prediction system based on UAV microwave remote sensing and deep learning as described in claim 4, characterized in that, in, The spatial dimension maintains the consistency of the corrected geographic coordinates to ensure that data from the same location at different time points can be compared; the temporal dimension dynamically adjusts the weights according to the crop growth stage to reflect the changes in drought characteristics at different phenological stages; the characteristic dimension integrates multi-band microwave parameters to form a comprehensive drought index.

6. The agricultural drought monitoring and prediction system based on UAV microwave remote sensing and deep learning as described in claim 1, characterized in that, The deep drought diagnostic subsystem also includes: The temporal pattern capture module is used to slice the three-dimensional drought feature field along the time axis, extract temporal feature curves at the locations of spatially anomalous patches, suppress high-frequency noise through moving averages, and highlight the trend changes of drought. The stage inflection point detection module is used to analyze the slope changes and curvature extremes of the time-series characteristic curves according to the crop growth stages, identify the drought acceleration period or the drought relief period; establish a statistical correlation between the drought characteristics in the early stage and the evolution in the later stage, and predict the drought chain reaction path.

7. The agricultural drought monitoring and prediction system based on UAV microwave remote sensing and deep learning as described in claim 6, characterized in that, The airspace anomaly detection module is used to compare the characteristic statistics of adjacent farmland patches within the same growth stage. If the characteristics of a patch deviate from the threshold range of the surrounding area, it is marked as a potential drought anomaly area. By combining historical agricultural data to eliminate interference from non-drought factors, drought-related abnormal patches are confirmed.

8. The agricultural drought monitoring and prediction system based on UAV microwave remote sensing and deep learning as described in claim 1, characterized in that, in, Extreme value difference: The difference between the largest and smallest eigenvalues ​​of the polarization scattering matrix; The vegetative growth stage contributes to the dispersion of the coherence coefficient; the reproductive growth stage enhances the weight of the extreme value differences of polarization characteristics. Maturity stage: Focuses on the central trend changes in backscattering intensity.

9. The agricultural drought monitoring and prediction system based on UAV microwave remote sensing and deep learning as described in claim 1, characterized in that, The drought response subsystem includes: The spatiotemporal feature decoding module is used to decompose the diagnostic results output by the spatiotemporal convolutional network into the spatial dimension of abnormal distribution of farmland patches and the temporal dimension of drought evolution trend. Based on the gradient response of the features of each layer of the spatiotemporal convolutional network, the module quantifies the contribution of polarization scattering matrix, coherence coefficient and backscattering intensity to the diagnostic results and identifies the core features that dominate drought discrimination. The dynamic drought level classification module is used to classify regions into mild, moderate and severe drought levels based on the spatial continuity of abnormal intensity of farmland patches; combined with the historical drought event database, it matches the historical level labels of similar patterns to the current drought evolution pattern to correct the preliminary classification results; The drought-causing factor reverse tracing module is used in polarization-dominant areas. If the abnormal contribution of the polarization scattering matrix eigenvalues ​​exceeds the threshold, it is located to vegetation structure variation or soil surface cracks. In coherence-dominant areas, when the coherence coefficient is abnormally high, it is associated with uneven canopy cover or differences in root water absorption. A continuous decrease in temporal coherence indicates soil water-holding capacity degradation. In scattering intensity-dominant areas, abrupt changes in the mean backscattering value are associated with a sharp drop in surface soil moisture content or interference from agricultural machinery operations. The threshold adaptive correction module is used to compare the changes in the feature field after actual irrigation with the prediction results, adjust the drought sensitivity threshold of the spatiotemporal convolutional network for misjudged areas, and dynamically update the weights of the feature-factor association mapping based on the accuracy of drought-causing factor localization.

Images