A soil moisture anomaly percentage grade prediction method and system based on multi-source remote sensing data

Abstract

The application discloses a soil moisture anomaly percentage grade prediction method and system based on multi-source remote sensing data. The method comprises the following steps: obtaining multi-source remote sensing data and geographic auxiliary data, and preprocessing and unifying to 1km resolution; calculating the multi-year synchronous average value based on passive microwave soil moisture data, obtaining soil moisture anomaly percentage SMAPI and dividing it into 10 levels, and resampling to 1km as a true value label; stacking multi-source features into a multi-channel input tensor; constructing an improved DeepLabv3+ model to directly map low-resolution levels to high-resolution levels in a semantic segmentation task, and adopting a compound loss function combining cross-entropy and Focal Loss for optimization; inputting the data to be predicted into the trained model, and directly outputting a 1km SMAPI level distribution map. The application skips the traditional downscaling and grading redundant process, realizes direct and accurate prediction of high-resolution drought levels, the error is controllable, and the extreme event recognition ability is strong.

Description

Technical Field

[0001] This invention relates to the fields of agricultural drought remote sensing monitoring and hydrological remote sensing inversion technology, specifically to a method and system for predicting soil moisture anomaly percentage levels based on multi-source remote sensing data, which is applicable to watershed-scale dynamic monitoring of soil moisture and operational applications of agricultural drought levels. Background Technology

[0002] Soil moisture is a key parameter for surface hydrological cycle and agricultural drought monitoring. Soil moisture anomaly percentage (SMAPI) can eliminate climatic background differences and directly reflect the abundance or scarcity of soil moisture compared to the same period in previous years. It is a core drought assessment indicator specified in the national standard "Meteorological Drought Grade". Accurately obtaining large-scale, high-resolution SMAPI grade distribution is of great significance for water resource management, agricultural drought resistance and disaster reduction, and ecological protection.

[0003] Existing methods for monitoring soil moisture drought levels based on remote sensing technology mainly suffer from the following technical bottlenecks:

[0004] (1) Passive microwave remote sensing products have low spatial resolution: Passive microwave soil moisture products, represented by satellites such as SMAP and AMSR-E, have all-weather monitoring capabilities, but their spatial resolution (about 9km-25km) is low and cannot meet the needs of fine-grained drought monitoring at the watershed scale.

[0005] (2) The traditional drought level inversion process is redundant and the error accumulates step by step: Existing technologies generally adopt a step-by-step process of "low-resolution soil moisture downscaling → continuous soil moisture numerical inversion → anomaly percentage calculation → level threshold classification". Each step will introduce calculation error and interpolation error, and the numerical regression error will be transmitted and amplified to the level classification stage, resulting in the distortion of the final result. At the same time, the model optimization objective is to minimize the continuous numerical fitting error, which does not match the core requirement of drought monitoring (level classification accuracy).

[0006] (3) Low accuracy of extreme drought and flood sample identification: The proportion of extreme drought and extreme flood events under natural conditions is extremely low. Existing models are trained using ordinary cross-entropy loss, which has a serious class imbalance problem. The recall rate of extreme events is insufficient and cannot meet the needs of drought emergency early warning.

[0007] In summary, existing technologies cannot achieve direct and accurate output of high-resolution SMAPI levels. There is an urgent need for a prediction method that skips redundant intermediate steps, has controllable errors, and provides spatial refinement, in order to achieve high-resolution, operational-grade drought level monitoring. Summary of the Invention

[0008] To address the problems of low resolution, redundant drought level determination process, and insufficient spatial prediction accuracy in existing microwave soil moisture products, this invention provides a method and system for predicting soil moisture anomaly percentage levels based on multi-source remote sensing data. It breaks away from the lengthy process of first downscaling and then classifying soil moisture, integrating soil moisture downscaling and SMAPI level classification into a pixel-level semantic segmentation task. Through multi-source remote sensing feature fusion, an improved DeepLabv3+ deep learning model, and optimization of class balance loss, it directly achieves accurate mapping from 9km low-resolution microwave levels to 1km high-resolution SMAPI levels, significantly improving the precision of monitoring and the reliability of classification, thus meeting operational monitoring needs.

[0009] To achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0010] A method for predicting the percentage anomaly of soil moisture based on multi-source remote sensing data includes the following steps:

[0011] Step (1) Multi-source data acquisition and preprocessing: Acquire multi-source remote sensing data and geographic auxiliary data. The multi-source remote sensing data includes passive microwave soil moisture data, vegetation index data, surface temperature data, evapotranspiration data, surface albedo data, and precipitation data. The geographic auxiliary data includes elevation data and soil attribute data. Perform spatiotemporal matching, projection transformation, cropping, resampling, and standardization preprocessing on the multi-source remote sensing data and geographic auxiliary data to unify them to a spatial resolution of 1km, and obtain preprocessed multi-source feature data.

[0012] Step (2) Soil moisture anomaly percentage calculation and grade label construction: Based on the passive microwave soil moisture data in the multi-source remote sensing data, calculate the average value of the same period over many years, and then calculate the soil moisture anomaly percentage SMAPI. Divide it into 10 discrete drought / wet grades according to the preset threshold to obtain grade labels with a resolution of 9km. Use bilinear interpolation to resample the grade labels of the 9km resolution to a resolution of 1km as the ground truth labels for model training.

[0013] Step (3) Multi-channel feature tensor synthesis: The multi-source remote sensing data preprocessed in step (1) are stacked and synthesized according to the pixel dimension. The synthesis dimension is H×W×C multi-channel input tensor, where H is the image height, W is the image width, and C is the number of feature channels.

[0014] Step (4) Improved DeepLabv3+ model construction and training: The soil moisture anomaly percentage level prediction is defined as a pixel-by-pixel multi-class semantic segmentation task. An improved DeepLabv3+ model adapted to soil moisture level prediction is constructed. The multi-channel input tensor synthesized in step (3) is used as input, and the level label with 1km resolution obtained in step (2) is used as the ground truth. The model parameters are optimized by using a class balance composite loss function that combines cross-entropy and Focal Loss to obtain the level prediction model after training.

[0015] Step (5) High-resolution grade prediction: Input the multi-channel feature tensor of the period to be predicted into the trained grade prediction model, and output a 1km high-resolution soil moisture anomaly percentage grade distribution map.

[0016] Furthermore, the specific method for spatiotemporal matching in step (1) is as follows: taking the 1-day time window of passive microwave soil moisture data as a benchmark, the MODIS product is synthesized using the average value within the window, and the GPM precipitation data is calculated using the cumulative value within the window, so as to achieve time dimension alignment of all data.

[0017] Furthermore, the formula for calculating the soil moisture anomaly percentage (SMAPI) in step (2) is as follows:

[0018] ;

[0019] ;

[0020] In the formula: , where d represents the average soil moisture for the same period over many years in the yth ten-day period and yth year, and n is the number of years in the statistical period; denoted as SMAP soil moisture value for the same period d in year i; SM represents the measured soil moisture value for the current period; SMAPI represents the soil moisture anomaly percentage.

[0021] Furthermore, the 10-level SMAPI classification rules in step (2) are as follows: Level 1 is SMAPI > 200% (extremely waterlogged), Level 2 is 100% < SMAPI ≤ 200% (severely waterlogged), Level 3 is 80% < SMAPI ≤ 100% (slightly waterlogged), Level 4 is 50% < SMAPI ≤ 80% (normally moist), Level 5 is 20% < SMAPI ≤ 50% (normal), Level 6 is 0% < SMAPI ≤ 20% (normally dry), Level 7 is -20% < SMAPI ≤ 0% (mildly dry), Level 8 is -50% < SMAPI ≤ -20% (moderately dry), Level 9 is -80% < SMAPI ≤ -50% (severely dry), and Level 10 is SMAPI ≤ -80% (extremely dry).

[0022] Furthermore, the number of feature channels C=12 mentioned in step (3) specifically includes the Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), Land Surface Temperature (LST), Actual Evapotranspiration (ET), Potential Evapotranspiration (PET), Albedo, Precipitation Data, Elevation Data (DEM), Slope Data, Soil Sand Content, Soil Clay Content, and Land Cover Type.

[0023] Furthermore, the improved DeepLabv3+ model described in step (4) includes an encoder unit, a multi-scale dilated spatial pyramid pooling unit, a cross-feature skip connection unit, a decoder unit, and a classification output unit, wherein:

[0024] The encoder unit uses a ResNet50 with the fully connected layers removed as the backbone network, performs 4 downsampling operations on the 12-channel input features, and outputs deep semantic features at 1 / 16 of the original resolution.

[0025] The multi-scale ASPP unit includes one 1×1 standard convolutional branch, four 3×3 dilated convolutional branches with dilation rates of 4, 8, 12 and 16 respectively, and one global average pooling branch. The output features of each branch are concatenated and then dimensionality reduced by 1×1 convolution to obtain multi-scale fused features.

[0026] The cross-feature skip connection unit concatenates the 1 / 4 resolution shallow detail features output from the second layer of the encoder with the upsampled ASPP features after 1×1 convolution dimensionality reduction.

[0027] The decoder unit performs two 3×3 convolutions on the spliced ​​features to extract and fuse the features, and then restores them to the original resolution of 1km through bilinear upsampling.

[0028] The classification output unit uses a 1×1 convolution to map the decoder output features into a 10-channel feature map, and outputs 10 levels of classification probability per pixel through the Softmax activation function. The level corresponding to the maximum probability is taken as the final prediction result.

[0029] Furthermore, the calculation formula for the category-balanced composite loss function mentioned in step (4) is as follows:

[0030] ;

[0031] In the formula: For standard cross-entropy loss, For Focal Loss, To predict the probability of the corresponding true class for the model. To balance the weighting factor of the proportion of samples in different drought / humidity levels, γ is the focusing parameter for easy and difficult samples.

[0032] A soil moisture anomaly percentage level prediction device based on multi-source remote sensing data, used to implement the method described above, includes:

[0033] The data preprocessing module is used to acquire multi-source remote sensing data and geographic auxiliary data. The multi-source remote sensing data includes passive microwave soil moisture data, vegetation index data, surface temperature data, evapotranspiration data, surface albedo data, and precipitation data. The geographic auxiliary data includes elevation data and soil attribute data. The module performs spatiotemporal matching, projection transformation, cropping, resampling, and standardization preprocessing on the multi-source remote sensing data and geographic auxiliary data to unify them to a spatial resolution of 1km, thereby obtaining preprocessed multi-source feature data.

[0034] The label construction module is used to calculate the average value of the same period over many years based on the passive microwave soil moisture data in the multi-source remote sensing data, and then calculate the soil moisture anomaly percentage (SMAPI). It is then divided into 10 discrete drought / wet level labels according to a preset threshold to obtain level labels with a resolution of 9km. The level labels with a resolution of 9km are resampled to a resolution of 1km using bilinear interpolation and used as ground truth labels for model training.

[0035] The feature synthesis module is used to stack and synthesize the preprocessed multi-source remote sensing data according to the pixel dimension. The synthesis dimension is a multi-channel input tensor of H×W×C, where H is the image height, W is the image width, and C is the number of feature channels.

[0036] The model training module is used to define the soil moisture anomaly percentage level prediction as a pixel-by-pixel multi-class semantic segmentation task, construct an improved DeepLabv3+ model adapted to soil moisture level prediction, take the multi-channel input tensor synthesized in step (3) as input, take the level label with 1km resolution obtained in step (2) as the true value, and use the class balance composite loss function that integrates cross-entropy and Focal Loss to optimize the model parameters, and obtain the trained level prediction model.

[0037] The grade prediction module is used to input the multi-channel feature tensor of the time period to be predicted into the trained grade prediction model and output a 1km high-resolution soil moisture anomaly percentage grade distribution map.

[0038] A soil moisture anomaly percentage prediction system based on multi-source remote sensing data includes:

[0039] Computer-readable storage media and processors;

[0040] The computer-readable storage medium is used to store executable instructions;

[0041] The processor is used to read executable instructions stored in the computer-readable storage medium and execute the soil moisture anomaly percentage level prediction method based on multi-source remote sensing data as described above.

[0042] A non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method for predicting soil moisture anomaly percentage levels based on multi-source remote sensing data as described above.

[0043] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects:

[0044] (1) Streamlined process and controllable error: It breaks the traditional lengthy process of first downscaling soil moisture, then calculating anomalies, and finally dividing thresholds. The model merges downscaling and drought level into a pixel-level semantic segmentation task and directly outputs the classification level, avoiding the error transmission of intermediate calculation links. The process is simplified, the error is reduced, and the needs of drought monitoring are met.

[0045] (2) Improved data resolution: The low-resolution microwave drought level is refined to a spatial resolution of 1km. By using the improved ASPP module in the DeepLabv3+ architecture, the rainfall, topography, and vegetation hydrological feedback mechanisms at different spatial scales are effectively captured, the true spatial heterogeneous distribution of drought on the land surface is restored, and the needs of farmland-scale soil moisture monitoring are met.

[0046] (3) Considering the imbalance of extreme event samples: Focal Loss is introduced as a special composite loss function, which is specifically optimized for extreme drought and extreme flood samples with extremely low occurrence frequency, significantly improving the recall rate and stability of disaster early warning.

[0047] (4) Strong operational adaptability: The input data are all publicly available remote sensing products, which can realize rapid mapping of drought levels over a wide range and multiple time phases, and are suitable for scenarios such as agricultural drought relief and hydrological monitoring. Attached Figure Description

[0048] Figure 1 This is a flowchart of a method for predicting the percentage of soil moisture anomalies based on multi-source remote sensing data, provided by an embodiment of the present invention.

[0049] Figure 2 This is a schematic diagram of the improved DeepLabv3+ model structure according to an embodiment of the present invention.

[0050] Figure 3 This is a schematic diagram of the SMAPI 10-level classification standard according to an embodiment of the present invention. Detailed Implementation

[0051] To clearly illustrate the solutions in this invention, preferred embodiments are given below in conjunction with the accompanying drawings. The following description is merely exemplary and not intended to limit the application or use of this disclosure.

[0052] Please see Figure 1 This invention provides a method for predicting the percentage anomaly of soil moisture based on multi-source remote sensing data, comprising the following steps:

[0053] Step (1) Multi-source data acquisition and preprocessing: Acquire multi-source time-series remote sensing and geographic auxiliary data. For the spatiotemporal resolution differences of different data sources, carry out spatiotemporal matching, projection normalization, resampling, outlier removal and standardization preprocessing to unify the data spatial resolution to 1km, time reference and projection coordinate system, and obtain preprocessed multi-source feature data. The multi-source remote sensing data includes passive microwave soil moisture data, vegetation index data, surface temperature data, evapotranspiration data, surface albedo data and precipitation data. The geographic auxiliary data includes elevation data and soil attribute data.

[0054] Step (2) Soil moisture anomaly percentage calculation and grade label construction: Based on the passive microwave soil moisture data, calculate the average value of the same period over many years, calculate the phenological scale soil moisture anomaly percentage SMAPI, and divide it into discrete 10 levels of drought / wetness according to the preset threshold to obtain grade labels with a resolution of 9km; use bilinear interpolation to resample the 9km grade labels to a resolution of 1km as ground truth labels for model training.

[0055] Step (3) Multi-channel feature tensor synthesis: The multi-source remote sensing data preprocessed in step (1) are stacked and synthesized according to the pixel dimension. The synthesis dimension is H×W×C multi-channel input tensor, where H is the image height, W is the image width, and C is the number of feature channels.

[0056] Step (4) Improved DeepLabv3+ model construction and training: The soil moisture anomaly percentage level prediction is defined as a pixel-by-pixel multi-class semantic segmentation task. An improved DeepLabv3+ model adapted to soil moisture level prediction is constructed. The multi-channel input tensor synthesized in step (3) is used as input, and the level label with 1km resolution obtained in step (2) is used as the ground truth. The model parameters are optimized by using a class balance composite loss function that combines cross-entropy and Focal Loss to obtain the level prediction model after training.

[0057] Step (5) High-resolution grade prediction: Input the multi-channel feature tensor of the period to be predicted into the trained grade prediction model, and directly output a 1km high-resolution soil moisture anomaly percentage grade distribution map.

[0058] The multi-source data acquisition in step (1) includes:

[0059] Passive microwave soil moisture data: NASA SMAP Level 3 soil moisture product (SPL3SMP) was used, with a spatial resolution of 9 km and a temporal resolution of 1 day.

[0060] Optical remote sensing data: MODIS products were used, including NDVI (MOD13A1, 500m, 16d), EVI (MOD13A1, 500m, 16d), LST (MOD11A2, 500m, 8d), ET / PET (MOD16A2), GPP (MOD17A2H, 500m, 8d), surface albedo (MCD43A3, 500m, 16d), and land cover type data (MCD12Q1, 500m, 1a).

[0061] Meteorological data: GPM IMERG Final daily precipitation product (0.1°);

[0062] Geographic auxiliary data: SRTM DEM elevation data (30m), soil property data from the World Soil Database HWSD (sand content, clay content, 1km).

[0063] The specific method of preprocessing in step (1) is as follows:

[0064] Spatiotemporal matching: Based on the 1-day time resolution of SMAP data, the MODIS products are synthesized using in-window average values, and the GPM precipitation data are calculated using in-window cumulative values ​​to achieve temporal alignment of all data.

[0065] Spatial matching: All data were uniformly projected to the WGS84 UTM coordinate system and clipped according to the boundary of the study area. Bilinear interpolation was used for data with a resolution higher than 1km, and inverse distance weighted interpolation was used for data with a resolution lower than 1km. All data were uniformly resampled to a resolution of 1km.

[0066] Standardization processing: Z-score standardization was applied to NDVI, EVI, LST, ET, PET, GPP, Albedo surface albedo, GPM IMERGFinal daily precipitation products, DEM elevation data, and soil property data as continuous features of the World Soil Database HWSD. Unique thermal coding was used for land cover types to eliminate dimensional differences.

[0067] In step (2), the soil moisture anomaly percentage (SMAPI) is calculated based on SMAP 9km soil moisture data. The average soil moisture for the same period over many years is calculated, and then the pixel-by-pixel soil moisture anomaly percentage is calculated. The formula is as follows:

[0068]

[0069]

[0070] In the formula: , where d represents the average soil moisture for the same period over many years in the yth ten-day period and yth year, and n is the number of years in the statistical period; denoted as SMAP soil moisture value for the same period d in year i; SM represents the measured soil moisture value for the current period; SMAPI represents the soil moisture anomaly percentage.

[0071] In step (2), the grading labels are divided into 10 levels of arid / humidity according to a preset threshold, such as... Figure 3 As shown, the specific classification rules are as follows: Level 1 is SMAPI > 200% (extremely waterlogged), Level 2 is 100% < SMAPI ≤ 200% (severely waterlogged), Level 3 is 80% < SMAPI ≤ 100% (slightly waterlogged), Level 4 is 50% < SMAPI ≤ 80% (normally moist), Level 5 is 20% < SMAPI ≤ 50% (normal), Level 6 is 0% < SMAPI ≤ 20% (normally dry), Level 7 is -20% < SMAPI ≤ 0% (mildly dry), Level 8 is -50% < SMAPI ≤ -20% (moderately dry), Level 9 is -80% < SMAPI ≤ -50% (severely dry), and Level 10 is SMAPI ≤ -80% (extremely dry).

[0072] In step (2), the SMAPI level labels with a resolution of 9km are resampled to 1km using nearest neighbor interpolation and used as ground truth labels for model training, ensuring that the spatial resolution of the labels is consistent with that of the input features.

[0073] In step (3), the multi-source remote sensing multi-channel feature tensor synthesis synthesizes multi-channel feature tensors by stacking the preprocessed multi-source remote sensing data according to the pixel dimension. The total number of channels is C=12. The specific channels include NDVI, EVI, LST, ET, PET, GPP, Albedo, precipitation, DEM, soil sand content, soil clay content, and land cover. The final input tensor dimension is H×W×C, where H is the number of image rows, W is the number of image columns, and C is the number of feature channels.

[0074] The improved DeepLabv3+ model described in step (4) consists of an encoder unit, a multi-scale ASPP unit, a cross-layer skip connection unit, a decoder unit, and a classification output unit, as follows: Figure 2 As shown, the specific structural optimizations are as follows:

[0075] Encoder unit: A ResNet50 with fully connected layers removed is used as the backbone network. The input tensor is downsampled four times to output deep hydrological semantic features at 1 / 16 of the original resolution, thereby improving the ability to extract complex surface features.

[0076] Multi-scale ASPP unit: It sets up one 1×1 standard convolution branch, four 3×3 dilated convolution branches with dilation rates of 4, 8, 12 and 16 respectively, and one global average pooling branch to capture multi-scale hydrological response features. The features of each branch are concatenated and then reduced in dimensionality by 1×1 convolution to achieve multi-scale feature fusion.

[0077] Cross-layer skip connection unit: Extracts 1 / 4 resolution shallow terrain and vegetation detail features from the second layer output of the encoder, reduces the dimension by 1×1 convolution, and then splices them with the upsampled ASPP fusion feature channel to make up for the loss of details caused by downsampling;

[0078] Decoder unit: The spliced ​​features are refined and fused by two 3×3 convolutions, and then restored to the original input resolution of 1km by bilinear upsampling to ensure the output spatial accuracy;

[0079] Classification output unit: The features are mapped to a 10-channel feature map using a 1×1 convolution, and the pixel-by-pixel classification probability is output after passing through the Softmax activation function. The level corresponding to the highest probability is taken as the final prediction result.

[0080] The category-balanced composite loss function described in step (4) takes into account both overall classification accuracy and extreme sample recognition ability. The calculation formula is as follows:

[0081]

[0082] In the formula: The standard cross-entropy loss is used to ensure overall classification convergence; Focal Loss focuses on samples from extreme droughts and floods that are difficult to distinguish. α is the model's predicted probability of the true level; α is the class weight factor, which balances the distribution of samples at each level; γ is the focusing parameter, which strengthens the training weight for low-probability extreme samples.

[0083] In step (4), the model training adopts a spatiotemporal hierarchical sampling strategy: the training set (2018-2022, accounting for 80%), the validation set and the test set (2023, each accounting for 10%) are divided by year to avoid spatiotemporal data leakage; the AdamW optimizer is used, the initial learning rate is set to 1e-4, the batch size is 16, the maximum number of iterations is 100, and the early stopping strategy (Patience=10) is enabled to prevent the model from overfitting.

[0084] This embodiment uses Henan Province as the study area, with data covering the period from 2018 to 2023. The specific implementation process is as follows: Figure 1 As shown, the specific implementation steps are as follows:

[0085] Step 1: Multi-source data acquisition and preprocessing

[0086] Step 1.1: Data Acquisition. Acquire multi-source remote sensing and geographic auxiliary data for the study area, including SMAP soil moisture data from 2018 to 2023, MODIS13A3 NDVI, MODIS13A2 EVI, MOD11A1 LST, MOD16A2 ET, MOD16A2 PET, MCD43A3 Albedo, MCD12Q1 land cover type, GPM precipitation data, DEM elevation data, HWSD soil property data (sand content ratio, clay content ratio), and land cover data.

[0087] Step 1.2: Time Matching Processing. Based on the SMAP daily time window, the MODIS 8d / 16d products are synthesized using the average value within the window, and the GPM daily precipitation data are calculated using the cumulative value within the window, thus aligning the time dimensions of all data.

[0088] Step 1.3: Spatial matching. All data were uniformly projected to the WGS84 UTM coordinate system, clipped to the study area boundary, and resampled to a 1km resolution;

[0089] Step 1.4 Standardization Processing. Continuous features including NDVI, EVI, LST, ET, PET, GPP, Albedo surface albedo, GPM IMERG Final daily precipitation products, DEM elevation data, and soil property data are standardized using the World Soil Database (HWSD) using Z-score standardization. Land cover types are encoded using unique thermal coding to eliminate dimensional differences.

[0090] Step 2: Calculate the soil moisture anomaly percentage and construct discretized level true value labels.

[0091] Step 2.1: Calculate the multi-year average soil moisture at the pixel-by-pixel climatological scale based on SMAP soil moisture data from 2018 to 2023, and calculate the soil moisture anomaly percentage (SMAPI).

[0092]

[0093]

[0094] In the formula: , where d represents the average soil moisture for the same period over many years in the yth ten-day period and yth year, and n is the number of years in the statistical period; denoted as SMAP soil moisture value for the same period d in year i; SM represents the measured soil moisture value for the current period; SMAPI represents the soil moisture anomaly percentage.

[0095] Step 2.2: Divide the water into 10 levels according to the preset thresholds: Level 1 is SMAPI > 200% (extremely waterlogged), Level 2 is 100% < SMAPI ≤ 200% (severely waterlogged), Level 3 is 80% < SMAPI ≤ 100% (slightly waterlogged), Level 4 is 50% < SMAPI ≤ 80% (normally wet), Level 5 is 20% < SMAPI ≤ 50% (normal), Level 6 is 0% < SMAPI ≤ 20% (normally dry), Level 7 is -20% < SMAPI ≤ 0% (mildly dry), Level 8 is -50% < SMAPI ≤ -20% (moderately dry), Level 9 is -80% < SMAPI ≤ -50% (severely dry), and Level 10 is SMAPI ≤ -80% (extremely dry).

[0096] Step 2.3: Use the nearest neighbor interpolation method to resample the soil moisture level to 1km and construct the ground value label for model training;

[0097] Step 3: Multichannel feature tensor synthesis.

[0098] Step 3.1: Stack the preprocessed multi-source data NDVI, EVI, LST, ET, PET, Albedo, precipitation, DEM, sand content, clay content, and land cover into a multi-channel input tensor of H×W×12 dimensions;

[0099] Step 4: Improve the DeepLabv3+ deep learning prediction model.

[0100] Step 4.1: The encoder uses a ResNet50 with the fully connected layers removed as the backbone network, performs 4 downsampling operations on the 13-channel input, and outputs deep features at 1 / 16 resolution.

[0101] Step 4.2: The ASPP module is configured with one 1×1 convolution, four 3×3 dilated convolutions with dilation rates of 4 / 8 / 12 / 16, and one global average pooling. After the branch features are concatenated, they are reduced to 256 channels by a 1×1 convolution.

[0102] Step 4.3: Skip connection: The 1 / 4 resolution features output from the second layer of the backbone network are reduced to 48 channels by 1×1 convolution and then concatenated with ASPP features that have been upsampled by 4 times.

[0103] Step 4.4: The decoder extracts features by splicing features and performing 2 layers of 3×3 convolution, then upsamples them by 4 times to restore the original resolution of 500m;

[0104] Step 4.5: The classification output is mapped to 10 channels using a 1×1 convolution, and the classification probability is output via Softmax.

[0105] Step 4.6: Introduce FocalLoss joint optimization. To address the classification bias caused by the extremely limited sample size due to extreme drought / flood disasters in nature, a loss function is constructed, with the specific formula as follows:

[0106]

[0107] In the formula: For standard cross-entropy loss, For FocalLoss, To predict the probability of the corresponding true class for the model. To balance the weighting factor of the proportion of samples in different drought / humidity levels, γ is the focusing parameter for easy and difficult samples.

[0108] Step 4.7: Use spatiotemporal stratified sampling to divide the data into training, validation, and test sets. The data from 2018 to 2022 is used as the training set (accounting for 80%), and the data from 2023 is divided into the validation and test sets in a 1:1 ratio (each accounting for 10%).

[0109] Step 4.8: Train the model using the AdamW optimizer, setting the initial learning rate to 1e-4, batch size to 16, and number of iterations to 100. Use an early stopping strategy (Patience=10) to prevent overfitting.

[0110] Step 5: High-resolution grading prediction.

[0111] Step 5.1: Input the multi-channel features of the period to be predicted in 2023 into the trained model, and directly output a spatial distribution map of soil moisture anomaly percentage at 1km resolution.

[0112] Compared with the prior art, the present invention has the following features and effects:

[0113] First, the process is streamlined and the error is controllable: This invention breaks away from the lengthy process of "soil moisture downscaling → continuous numerical inversion → anomaly calculation → grade classification" in the traditional method. It integrates downscaling and grade prediction into a pixel-by-pixel semantic segmentation task, directly outputting SMAPI grades at a resolution of 1km. This avoids the accumulation and propagation of errors in intermediate steps and improves the reliability of the prediction results.

[0114] Second, the spatial resolution is significantly improved: by leveraging the multi-scale ASPP module and skip connection structure in the improved DeepLabv3+ model, multi-source features such as terrain, vegetation, and precipitation are effectively integrated, refining the 9km low-resolution microwave drought level into a 1km high-resolution level map, meeting the needs of refined drought monitoring at the farmland and watershed scales.

[0115] Third, it has strong ability to identify extreme drought and flood samples: In response to the problem that the proportion of extreme drought and flood samples under natural conditions is extremely low, a class balance composite loss function that combines cross-entropy and Focal Loss is introduced, which significantly improves the recall rate and classification stability of extreme events and enhances emergency early warning capabilities.

[0116] Fourth, it has good operational adaptability: the input data are all publicly available remote sensing products, and after model training, it can achieve rapid hierarchical mapping of large areas and multiple time phases, which is suitable for operational scenarios such as agricultural drought relief and hydrological monitoring.

[0117] Another aspect of the present invention provides a soil moisture anomaly percentage level prediction device based on multi-source remote sensing data, used to implement the method described above, comprising:

[0118] The data preprocessing module is used to acquire multi-source remote sensing data and geographic auxiliary data. The multi-source remote sensing data includes passive microwave soil moisture data, vegetation index data, surface temperature data, evapotranspiration data, surface albedo data, and precipitation data. The geographic auxiliary data includes elevation data and soil attribute data. The module performs spatiotemporal matching, projection transformation, cropping, resampling, and standardization preprocessing on the multi-source remote sensing data and geographic auxiliary data to unify them to a spatial resolution of 1km, thereby obtaining preprocessed multi-source feature data.

[0119] The label construction module is used to calculate the average value of the same period over many years based on the passive microwave soil moisture data in the multi-source remote sensing data, and then calculate the soil moisture anomaly percentage (SMAPI). It is then divided into 10 discrete drought / wet level labels according to a preset threshold to obtain level labels with a resolution of 9km. The level labels with a resolution of 9km are resampled to a resolution of 1km using bilinear interpolation and used as ground truth labels for model training.

[0120] The feature synthesis module is used to stack and synthesize the preprocessed multi-source remote sensing data according to the pixel dimension. The synthesis dimension is a multi-channel input tensor of H×W×C, where H is the image height, W is the image width, and C is the number of feature channels.

[0121] The model training module is used to define the soil moisture anomaly percentage level prediction as a pixel-by-pixel multi-class semantic segmentation task, construct an improved DeepLabv3+ model adapted to soil moisture level prediction, take the multi-channel input tensor synthesized in step (3) as input, take the level label with 1km resolution obtained in step (2) as the true value, and use the class balance composite loss function that integrates cross-entropy and Focal Loss to optimize the model parameters, and obtain the trained level prediction model.

[0122] The grade prediction module is used to input the multi-channel feature tensor of the time period to be predicted into the trained grade prediction model and output a 1km high-resolution soil moisture anomaly percentage grade distribution map.

[0123] Another aspect of the present invention provides a soil moisture anomaly percentage prediction system based on multi-source remote sensing data, comprising: a computer-readable storage medium and a processor;

[0124] The computer-readable storage medium is used to store executable instructions;

[0125] The processor is used to read executable instructions stored in the computer-readable storage medium and execute the method for predicting the percentage anomaly of soil moisture based on multi-source remote sensing data.

[0126] In another aspect, the present invention provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned method for predicting the percentage anomaly of soil moisture based on multi-source remote sensing data.

[0127] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0128] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0129] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0130] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0131] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for predicting the percentage anomaly of soil moisture based on multi-source remote sensing data, characterized in that: Includes the following steps: Step (1) Multi-source data acquisition and preprocessing: Acquire multi-source remote sensing data and geographic auxiliary data. The multi-source remote sensing data includes passive microwave soil moisture data, vegetation index data, surface temperature data, evapotranspiration data, surface albedo data, and precipitation data. The geographic auxiliary data includes elevation data and soil attribute data. Perform spatiotemporal matching, projection transformation, cropping, resampling, and standardization preprocessing on the multi-source remote sensing data and geographic auxiliary data to unify them to a spatial resolution of 1km, and obtain preprocessed multi-source feature data. Step (2) Soil moisture anomaly percentage calculation and grade label construction: Based on the passive microwave soil moisture data in the multi-source remote sensing data, calculate the average value of the same period over many years, and then calculate the soil moisture anomaly percentage SMAPI. Divide it into 10 discrete drought / wet grades according to the preset threshold to obtain grade labels with a resolution of 9km. Use bilinear interpolation to resample the grade labels of the 9km resolution to a resolution of 1km as the ground truth labels for model training. Step (3) Multi-channel feature tensor synthesis: The multi-source remote sensing data preprocessed in step (1) are stacked and synthesized according to the pixel dimension. The synthesis dimension is H×W×C multi-channel input tensor, where H is the image height, W is the image width, and C is the number of feature channels. Step (4) Improved DeepLabv3+ model construction and training: The soil moisture anomaly percentage level prediction is defined as a pixel-by-pixel multi-class semantic segmentation task. An improved DeepLabv3+ model adapted to soil moisture level prediction is constructed. The multi-channel input tensor synthesized in step (3) is used as input, and the level label with 1km resolution obtained in step (2) is used as the ground truth. The model parameters are optimized by using a class balance composite loss function that combines cross-entropy and Focal Loss to obtain the level prediction model after training. Step (5) High-resolution grade prediction: Input the multi-channel feature tensor of the period to be predicted into the trained grade prediction model, and output a 1km high-resolution soil moisture anomaly percentage grade distribution map.

2. The method according to claim 1, characterized in that, The specific method for spatiotemporal matching in step (1) is as follows: taking the 1-day time window of passive microwave soil moisture data as a benchmark, the MODIS product is synthesized using the average value within the window, and the GPM precipitation data is calculated using the cumulative value within the window, so as to achieve time dimension alignment of all data.

3. The method as described in claim 1, characterized in that, The formula for calculating the soil moisture anomaly percentage (SMAPI) in step (2) is: ； ； In the formula: , where d represents the average soil moisture for the same period over many years in the yth ten-day period and yth year, and n is the number of years in the statistical period; denoted as SMAP soil moisture value for the same period d in year i; SM represents the measured soil moisture value for the current period; SMAPI represents the soil moisture anomaly percentage.

4. The method as described in claim 1, characterized in that, The 10-level SMAPI classification rules in step (2) are as follows: Level 1 is SMAPI > 200% (extremely waterlogged), Level 2 is 100% < SMAPI ≤ 200% (severely waterlogged), Level 3 is 80% < SMAPI ≤ 100% (slightly waterlogged), Level 4 is 50% < SMAPI ≤ 80% (normally moist), Level 5 is 20% < SMAPI ≤ 50% (normal), Level 6 is 0% < SMAPI ≤ 20% (normally dry), Level 7 is -20% < SMAPI ≤ 0% (mildly dry), Level 8 is -50% < SMAPI ≤ -20% (moderately dry), Level 9 is -80% < SMAPI ≤ -50% (severely dry), and Level 10 is SMAPI ≤ -80% (extremely dry).

5. The method as described in claim 1, characterized in that, The number of feature channels C=12 mentioned in step (3) includes the following channels in order: Normalized Difference Vegetation Index (NDVI), Enhanced Difference Vegetation Index (EVI), Land Surface Temperature (LST), Actual Evapotranspiration (ET), Potential Evapotranspiration (PET), Albedo, Precipitation Data, Elevation Data (DEM), Slope Data, Soil Sand Content, Soil Clay Content, and Land Cover Type.

6. The method as described in claim 1, characterized in that, The improved DeepLabv3+ model described in step (4) includes an encoder unit, a multi-scale dilated spatial pyramid pooling unit, a cross-feature skip connection unit, a decoder unit, and a classification output unit, wherein: The encoder unit uses a ResNet50 with the fully connected layers removed as the backbone network, performs 4 downsampling operations on the 12-channel input features, and outputs deep semantic features at 1 / 16 of the original resolution. The multi-scale ASPP unit includes one 1×1 standard convolutional branch, four 3×3 dilated convolutional branches with dilation rates of 4, 8, 12 and 16 respectively, and one global average pooling branch. The output features of each branch are concatenated and then dimensionality reduced by 1×1 convolution to obtain multi-scale fused features. The cross-feature skip connection unit concatenates the 1 / 4 resolution shallow detail features output from the second layer of the encoder with the upsampled ASPP features after 1×1 convolution dimensionality reduction. The decoder unit performs two 3×3 convolutions on the spliced ​​features to extract and fuse the features, and then restores them to the original resolution of 1km through bilinear upsampling. The classification output unit uses a 1×1 convolution to map the decoder output features into a 10-channel feature map, and outputs 10 levels of classification probability per pixel through the Softmax activation function. The level corresponding to the maximum probability is taken as the final prediction result.

7. The method as described in claim 1, characterized in that, The formula for calculating the category-balanced composite loss function mentioned in step (4) is as follows: ； In the formula: For standard cross-entropy loss, For Focal Loss, To predict the probability of the corresponding true class for the model. To balance the weighting factor of the proportion of samples in different drought / humidity levels, γ is the focusing parameter for easy and difficult samples.

8. A soil moisture anomaly percentage prediction device based on multi-source remote sensing data, used to implement the method described in any one of claims 1-7, characterized in that, include: The data preprocessing module is used to acquire multi-source remote sensing data and geographic auxiliary data. The multi-source remote sensing data includes passive microwave soil moisture data, vegetation index data, surface temperature data, evapotranspiration data, surface albedo data, and precipitation data. The geographic auxiliary data includes elevation data and soil attribute data. The module performs spatiotemporal matching, projection transformation, cropping, resampling, and standardization preprocessing on the multi-source remote sensing data and geographic auxiliary data to unify them to a spatial resolution of 1km, thereby obtaining preprocessed multi-source feature data. The label construction module is used to calculate the average value of the same period over many years based on the passive microwave soil moisture data in the multi-source remote sensing data, and then calculate the soil moisture anomaly percentage (SMAPI). It is then divided into 10 discrete drought / wet level labels according to a preset threshold to obtain level labels with a resolution of 9km. The level labels with a resolution of 9km are resampled to a resolution of 1km using bilinear interpolation and used as ground truth labels for model training. The feature synthesis module is used to stack and synthesize the preprocessed multi-source remote sensing data according to the pixel dimension. The synthesis dimension is a multi-channel input tensor of H×W×C, where H is the image height, W is the image width, and C is the number of feature channels. The model training module is used to define the soil moisture anomaly percentage level prediction as a pixel-by-pixel multi-class semantic segmentation task, construct an improved DeepLabv3+ model adapted to soil moisture level prediction, take the multi-channel input tensor synthesized in step (3) as input, take the level label with 1km resolution obtained in step (2) as the true value, and use the class balance composite loss function that integrates cross-entropy and FocalLoss to optimize the model parameters, and obtain the level prediction model after training. The grade prediction module is used to input the multi-channel feature tensor of the time period to be predicted into the trained grade prediction model and output a 1km high-resolution soil moisture anomaly percentage grade distribution map.

9. A soil moisture anomaly percentage prediction system based on multi-source remote sensing data, characterized in that, include: Computer-readable storage media and processors; The computer-readable storage medium is used to store executable instructions; The processor is used to read executable instructions stored in the computer-readable storage medium and execute the soil moisture anomaly percentage level prediction method based on multi-source remote sensing data as described in any one of claims 1-8.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method for predicting the percentage anomaly of soil moisture based on multi-source remote sensing data as described in any one of claims 1-8.

Images