Method for predicting biomass of field crops in cloud-sheltered area of satellite image in rainy season

Abstract

The invention discloses a method for predicting biomass of field crops in a cloud layer shielding area of a satellite image in a rainy season. The method comprises the following steps: acquiring key growth period data of field crops in a research area; satellite image farmland extraction and crop biomass data inversion are carried out; after possible disaster-affected farmland is removed, satellite images of an actual prediction area shielded by a cloud layer and a reference area not shielded by the cloud layer and respective crop biomass data are obtained, and after comparison, a crop biomass distribution diagram of a to-be-predicted date is obtained, and crop biomass prediction of the cloud layer shielded area is realized. By adopting a space-air-ground integrated monitoring technology, under multi-dimensional cooperation of a satellite, an unmanned aerial vehicle and a ground sensing network, accurate inversion of satellite-scale crop biomass data is realized, the problem of lack of satellite-scale data due to cloud layer shielding in a key growth period of crops can be solved, and the accuracy of crop biomass inversion is improved. Satellite scale crop growth vigor continuous monitoring is realized.

Description

Technical Field

[0001] The invention relates to a method for predicting the biomass of field crops, and in particular to a method for predicting the biomass of field crops in a cloud-blocked area of ​​satellite images in rainy seasons. Background Art

[0002] Satellites can cover a wide area and are suitable for monitoring large-scale farmland. Based on satellite remote sensing, crop growth status analysis and evaluation can be achieved on a large scale. Accurate crop growth prediction can help optimize agricultural management measures such as irrigation, fertilization, and pest control. This can not only increase crop yields, but also reduce resource waste and environmental pollution. When using remote sensing technology to monitor crop growth, optical satellite data (such as Landsat, Sentinel-2, etc.) is one of the most commonly used data sources. However, these data are easily affected by clouds and cloud shadows, especially in cloudy or rainy areas. Cloud obstruction can cause missing or incomplete data, thus affecting the accurate and continuous prediction of crop growth. Therefore, it is particularly important to develop methods that can still effectively predict crop growth in the case of cloud obstruction. By developing crop growth prediction methods suitable for cloud-obstructed areas, the accuracy and reliability of predictions can be significantly improved, enabling farmers, agricultural managers, and policymakers to make decisions based on accurate information.

[0003] In recent years, with the development of remote sensing technology, machine learning algorithms, and big data processing capabilities, researchers have begun to explore new methods to solve the problem of cloud occlusion. For example, Synthetic Aperture Radar (SAR) data is not affected by clouds and can provide all-weather observation data. Although SAR data has obvious advantages in solving the problem of cloud occlusion, its geometric distortion, speckle noise, low contrast, polarization complexity, cost, and technical barriers cannot be ignored. For example, compared with optical data, especially in flat and uniform areas, the contrast of SAR data is usually low, and low contrast may make it difficult to distinguish different types of land objects, thereby affecting crop type identification and growth indicator estimation.

[0004] In addition, multi-source data fusion and time series analysis methods also provide new ways to solve this problem. Combine data from different sensors (such as optical satellites, SAR, drones, ground observation stations, etc.) to provide more comprehensive and accurate information. By fusing data from different sources, the advantages of each data source can be complemented and the limitations of a single data source can be overcome. Time series analysis is a statistical method used to analyze data points that change over time and extract useful information and patterns from them. In agricultural monitoring, time series analysis can help fill data gaps caused by cloud cover and capture dynamic changes in crop growth. However, existing time series analysis methods are more used to predict the morphological evolution and movement of remote sensing satellite cloud images to effectively monitor and quickly predict disastrous weather conditions such as thunderstorms, strong winds, and heavy rains. These methods are not applicable when used for farmland crop prediction because they do not take into account the particularity of farmland crops and the abnormal growth or even death of crops that may be caused by precipitation. Summary of the invention

[0005] In order to solve the problems existing in the background technology, the present invention provides a method for predicting field crop biomass in cloud-blocked areas of satellite images in rainy seasons. The present invention is inspired by multi-source data fusion and time series analysis methods, and proposes a simple and fast method for predicting crop biomass in cloud-blocked areas of satellite images in rainy seasons to cope with the challenges brought by climate change and improve the adaptability and resilience of agricultural monitoring systems.

[0006] The technical solution adopted by the present invention is:

[0007] The method for predicting crop biomass in a cloud-blocked area of ​​satellite images in rainy seasons of the present invention comprises:

[0008] S1: Construct a farmland image segmentation model, obtain historical satellite images of the key growth period of field crops in the study area that are not blocked by clouds, and train the farmland image segmentation model to obtain a trained image segmentation model.

[0009] S2: Obtain satellite images to be predicted during the critical growth period of field crops in the study area, as well as drone images and ground sampling data; input the satellite images to be predicted into the trained farmland image segmentation model and output the farmland mask after processing; process the satellite images to be predicted based on the farmland mask and combined with the elevation information to obtain the satellite images to be predicted that only contain the farmland area and its elevation.

[0010] S3: Construct a crop biomass inversion model at different growth stages based on the UAV images and ground sampling data in step S2, so as to obtain crop biomass data by inverting the crop biomass inversion model at different growth stages, the UAV images and the satellite images to be predicted that only contain the farmland area and its elevation.

[0011] S4: The date blocked by clouds in the satellite image to be predicted that only contains the farmland area and its elevation is taken as the date to be predicted. Based on the precipitation data in the preset time period before the date to be predicted and the elevation data of the farmland area, the farmland that may be affected in the study area is determined and removed to obtain the satellite image of the actual prediction area blocked by clouds and the reference area not blocked by clouds.

[0012] S5: Perform the inversion operation in step S3 on the satellite images of the actual prediction area blocked by clouds within a preset time period before the prediction date and the reference area not blocked by clouds within the prediction date and the preset time period before the prediction date to obtain their respective crop biomass data, and then obtain the crop biomass data of the actual prediction area blocked by clouds on the prediction date after comparison, so as to obtain the crop biomass distribution map of the prediction date and realize the crop biomass prediction of the cloud-blocked area.

[0013] If cloud-free satellite images are obtained for a later date, the predicted crop biomass value of the actual predicted area on the predicted date should be immediately corrected based on the later data.

[0014] The step S1 is specifically as follows:

[0015] S11: Replace the ResNet module in the backbone layer of the Pyramid Attention Network (PAN) with a lightweight MobileNetV1 network to build a farmland image segmentation model.

[0016] S12: For each historical satellite image that is not blocked by clouds during the critical growth period of field crops in the study area, it is uniformly cropped into several square cropped satellite images of the same size. The labels of the farmland coverage area and other areas are marked in each cropped satellite image and divided into training set and validation set according to the preset ratio.

[0017] S13: Binary cross entropy loss and Dice coefficient loss are combined into a hybrid loss function to construct a farmland image segmentation model. The training set and the validation set are input into the farmland image segmentation model for training until the hybrid loss function converges to obtain a trained image segmentation model.

[0018] In step S2, the satellite images are panchromatic and multispectral images of the high-resolution series satellites; the drone images are RGB and multispectral images; the ground sampling data are the biomass data of field crops at each preset sampling point in a typical field in the study area; the key growth period of field crops includes no less than two different growth periods, and each growth period is sorted according to the phenological period. All images and ground sampling data are classified by date and stored in the database and are called historical date images; the ground sampling data can be collected by selecting typical fields in the study area, and the biomass data is specifically the fresh weight data of the aboveground part of wheat.

[0019] In the step S2, farmland extraction is performed on the satellite image to be predicted according to the farmland mask, and then the average elevation information of each farmland in the satellite image to be predicted is obtained in combination with the satellite digital elevation model (DEM), so as to obtain a satellite image to be predicted that only contains the farmland area and its elevation.

[0020] In the step S3, firstly, various vegetation indexes and visible light image texture features at various preset sampling points are obtained according to the drone image in step S2. For each preset sampling point, the various vegetation indexes at the preset sampling point and the biomass data of field crops in the ground sampling data are subjected to correlation regression analysis and processing, and several vegetation indexes with correlations higher than a preset threshold are screened out. According to the various screened vegetation indexes and visible light image texture features and using an integrated learning regression algorithm, an inversion model for the biomass of crops in different growth periods is constructed. The inversion model for the biomass of crops in different growth periods obtains the various different growth indexes in the key growth period of field crops at the drone scale according to the drone image inversion. The biomass distribution of field crops in the key growth period is taken as the first crop biomass data, and then the first crop biomass data is interpolated and upscaled to match the resolution of the satellite image, and the upscaled first crop biomass data is used as the true value of the satellite-scale biomass inversion, and the satellite image to be predicted containing only the farmland area and its elevation in step S2 and the true value of the satellite-scale biomass inversion are subjected to the same operation as the biomass data of field crops in the UAV image and ground sampling data in step S3, so as to invert the biomass distribution of field crops in different growth periods in the key growth period of field crops at the satellite scale as the final crop biomass data.

[0021] In the step S4, precipitation data on the predicted date and the preset time period before it in the study area is obtained, including precipitation amount, precipitation time and precipitation duration, and the disaster value of each farmland in the study area is determined based on the elevation data of the farmland area. The farmland whose disaster value exceeds the disaster threshold value is regarded as the possible disaster-stricken farmland, and the possible disaster-stricken farmland is cropped and removed from the satellite image to be predicted that only contains the farmland area and its elevation. The remaining area blocked by clouds in the study area is used as the actual prediction area, and the area not blocked by clouds is used as the reference area.

[0022] The damage value R of the farmland is as follows:

[0023]

[0024] Among them, α1, α2, α3 and α4 are the weights of the first, second, third and fourth factors, respectively, which are determined according to historical waterlogging events; β1 and β2 are the first and second nonlinear indexes, respectively, which are used to simulate the nonlinear effects of soil permeability coefficient and terrain on the water accumulation cutoff; P is the cumulative precipitation in the study area on the date to be predicted and the preset time period before it; T is the precipitation duration in the study area on the date to be predicted and the preset time period before it; K is the soil permeability coefficient of the farmland in the study area, which is determined according to the soil type of the farmland; G represents the terrain influence coefficient, which is determined according to the elevation data of the farmland area. The higher the elevation, the lower the terrain influence coefficient, and the less likely it is to be waterlogged. The critical value of the disaster value R of the farmland and the size of each parameter are determined through historical waterlogging events, and then the waterlogging situation of each field in the study area is judged based on the generated empirical formula.

[0025] In step S5, for each pixel in the satellite image, the crop biomass data obtained from the satellite image of the reference area that is not blocked by clouds during the preset time period before the predicted date are arranged in order of growth period to obtain a first biomass array A. i [a1, a2, …, a n ], where a n , …, a2 and a1 represent the crop biomass data of the to-be-predicted date, …, n-1 days before the to-be-predicted date and n days before the to-be-predicted date of the i-th pixel in the satellite image respectively; the crop biomass data obtained from the satellite image of the actual prediction area blocked by clouds within the preset time period before the to-be-predicted date are arranged in sequence according to the growth period to obtain the second biomass array B i[ b1,b2,…,b n-1 ], where b n-1, …, b2 and b1 represent the crop biomass data of the ith pixel in the satellite image 1 day before the predicted date, …, n-1 days before the predicted date and n days before the predicted date respectively; the second biomass array of each pixel in the satellite image is compared with the first n-1 elements in the first biomass array. If the first n-1 elements in one of the first biomass arrays are exactly the same as the second biomass array, the nth element of the first biomass array is used as the nth element of the second biomass array, that is, as the crop biomass data of the actual predicted area blocked by clouds on the predicted date. The crop biomass data of the actual predicted area blocked by clouds on the predicted date are fused with the actual predicted area to obtain the crop biomass distribution map of the predicted date.

[0026] The beneficial effects of the present invention are:

[0027] The present invention adopts integrated air-ground-space monitoring technology, and realizes the accurate inversion of satellite-scale crop biomass data with the multi-dimensional cooperation of satellites, drones and ground sensor networks. It can solve the problem of lack of satellite-scale data due to cloud cover during the critical growth period of crops, and realize continuous monitoring of crop growth at the satellite scale. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 is a flow chart of the method of the present invention;

[0029] Figure 2 This is an example of satellite imagery during the cloud cover period;

[0030] Figure 3 This is the structure diagram of the farmland image segmentation model;

[0031] Figure 4 It is the label map of satellite image classification results;

[0032] Figure 5 is the extracted farmland area map;

[0033] Figure 6 is the relationship diagram between wheat biomass and vegetation index at seedling and tillering stages, where: Figure 6 (a) is the relationship between a certain day in the seedling stage and the vegetation index. Figure 6 (b) is the relationship diagram between a certain day in the tillering period and the vegetation index;

[0034] Figure 7 This is the inversion result diagram of the first biomass data of wheat seedling and tillering stages. Figure 7 (a) is the inversion result of the first biomass data of wheat seedling stage. Figure 7 (b) is the inversion result diagram of the first biomass data of wheat tillering stage;

[0035] Figure 8This is a schematic diagram of the statistical distribution of AGB estimation accuracy of each level of the Stacking model during the test phase, where: Figure 8 (a) is the seedling model R 2 Schematic diagram of the statistical distribution of AGB estimation accuracy of each level of models during the testing phase. Figure 8 (b) is a diagram showing the statistical distribution of the AGB estimation accuracy of the seedling model RMSE at each level of the model during the test phase. Figure 8 (c) is the tillering stage model R 2 Schematic diagram of the statistical distribution of AGB estimation accuracy of each level of models during the testing phase. Figure 8 (d) is a schematic diagram of the statistical distribution of the AGB estimation accuracy of the tillering period model RMSE at each level of the model during the testing phase. DETAILED DESCRIPTION

[0036] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0037] like Figure 1 and Figure 2 As shown, the present invention studies the key growth period of wheat in a certain area in a certain month during specific implementation. The satellite image cloud coverage in the study area in a certain month is high, and this period is the key growth period of crops. For example, wheat is in the tillering-booting stage. The lack of crop growth monitoring during this period will lead to the lack of key data in the monitoring of this growth cycle. The method for predicting crop biomass in the cloud-blocked area of ​​satellite images in rainy seasons of the present invention is as follows:

[0038] S1: Construct a farmland image segmentation model, obtain historical satellite images of the key growth period of field crops in the study area that are not blocked by clouds, and train the farmland image segmentation model to obtain a trained image segmentation model, as follows:

[0039] S11: Replace the ResNet module in the backbone layer of the pyramid attention network PAN with a lightweight MobileNetV1 network to build a farmland image segmentation model.

[0040] S12: For each historical satellite image of the key growth period of field crops in the study area that is not blocked by clouds, it is evenly cropped into several square cropped satellite images of the same size, and the labels of the farmland coverage area and other areas are marked in each cropped satellite image and divided into training set and validation set according to the preset ratio; in the specific implementation, the original historical satellite image is cropped, and the pixel size of the cropped image is 512×512, and then it is labeled, the field is named farm, the pixel value is 1, and other coverage type areas are labeled other, and the pixel value is 255. When dividing, the training set, validation set and test set can be divided according to 7:2:1 to obtain the data set for the semantic segmentation model.

[0041] S13: The binary cross entropy loss and the Dice coefficient loss are combined into a hybrid loss function to construct a farmland image segmentation model. The training set and the validation set are input into the farmland image segmentation model for training until the hybrid loss function converges to obtain a trained image segmentation model.

[0042] like Figure 3 The model shown in the figure is a farmland image segmentation model. By comparing the performance of the model of this method with the classic semantic segmentation model UNet and other networks on the data set, the results are shown in Table 1. The evaluation index in the table is the training result of the 50th round. It can be seen that the model constructed by PAN using MobileNet2 as the backbone network performs better than the model constructed by using Resnet as the backbone network. And considering the training time and prediction accuracy, the model selected by the present invention is also the best performing model. The satellite images are classified using the trained model. The results are shown in Figure 4 According to the classification results, farmland is extracted from satellite images. The extraction results are shown in Figure 5 The farmland label map is superimposed on the satellite image DEM to extract the average elevation data of each field.

[0043] Table 1

[0044]

[0045]

[0046] S2: Obtain satellite images to be predicted during the critical growth period of field crops in the study area, as well as drone images and ground sampling data; input the satellite images to be predicted into the trained farmland image segmentation model and output the farmland mask after processing; process the satellite images to be predicted based on the farmland mask and combined with the elevation information to obtain the satellite images to be predicted that only contain the farmland area and its elevation.

[0047] Satellite images are panchromatic and multispectral images of the Gaofen series satellites; drone images are RGB and multispectral images; ground sampling data are biomass data of field crops at each preset sampling point in a typical field in the study area; the key growth period of field crops includes no less than two different growth periods, and each growth period is sorted according to the phenological period. All images and ground sampling data are classified by date and stored in the database and are called historical date images; ground sampling data can be collected from typical fields in the study area, and biomass data is specifically fresh weight data of the aboveground part of wheat.

[0048] The farmland is extracted from the satellite image to be predicted according to the farmland mask, and then the average elevation information of each farmland in the satellite image to be predicted is obtained in combination with the satellite digital elevation model DEM, so as to obtain the satellite image to be predicted that only contains the farmland area and its elevation.

[0049] S3: Construct a crop biomass inversion model at different growth stages based on the UAV images and ground sampling data in step S2, so as to obtain crop biomass data by inverting the crop biomass inversion model at different growth stages, the UAV images and the satellite images to be predicted that only contain the farmland area and its elevation.

[0050] First, various vegetation indices and visible light image texture features at various preset sampling points are obtained according to the drone image in step S2. For each preset sampling point, the various vegetation indices at the preset sampling points and the biomass data of field crops in the ground sampling data are subjected to correlation regression analysis, and several vegetation indices with correlations higher than a preset threshold are screened out. According to the screened vegetation indices and visible light image texture features, an integrated learning regression algorithm is used to construct a crop biomass inversion model at different growth stages. The integrated learning regression algorithm specifically adopts the Stacking model, and a total of two layers of learners are designed. Among them, the primary learner uses a total of 5 machine learning algorithms, namely Gaussian regression, support vector regression, random forest, multivariate linear regression and Cubist regression. The secondary learner selects the linear ridge regression algorithm. The biomass data set is input into the Stacking model for training according to the five-fold crossover method. A total of 200 rounds of training are conducted, and the data set is re-divided in each round of training to avoid repeated training. The crop biomass inversion model at different growth stages obtains the field crop biomass distribution of each different growth stage in the key growth period of field crops at the drone scale according to the drone image inversion as the first crop biomass data, such as Figure 7 (a) and Figure 7 As shown in (b), the first biomass data inversion results of wheat seedling stage and wheat tillering stage can be obtained; then the first crop biomass data is interpolated and upscaled to match the resolution of the satellite image, and the upscaled first crop biomass data is used as the true value of the satellite-scale biomass inversion, and the satellite image to be predicted that only contains the farmland area and its elevation in step S2 and the true value of the satellite-scale biomass inversion are subjected to the same operation as the biomass data of field crops in the drone image and ground sampling data in step S3, thereby inverting the biomass distribution of field crops in different growth periods in the key growth period of field crops at the satellite scale as the final crop biomass data.

[0051] The spectral vegetation index includes normalized difference vegetation index, improved simple ratio vegetation index, green normalized difference vegetation index, green wave wide dynamic vegetation index, improved soil adjustment vegetation index, red edge simple ratio vegetation index, nonlinear vegetation index and structure insensitive pigment index. The texture features of visible light images are 8: angular second moment, contrast, homogeneity, correlation, entropy, contrast, cluster prominence and cluster shadow. Figure 6 (a) and Figure 6 (b) shows the normalized vegetation index of wheat at the seedling stage and the tillering stage.

[0052] S4: The date blocked by clouds in the satellite image to be predicted that only contains the farmland area and its elevation is taken as the date to be predicted. Based on the precipitation data in the preset time period before the date to be predicted and the elevation data of the farmland area, the farmland that may be affected in the study area is determined and removed to obtain the satellite image of the actual prediction area blocked by clouds and the reference area not blocked by clouds.

[0053] Obtain precipitation data on the predicted date and the preset time period before it in the study area, including precipitation amount, precipitation time and precipitation duration, and determine the disaster value of each farmland in the study area based on the elevation data of the farmland area. The farmland with a disaster value exceeding the critical value is regarded as the possible disaster-stricken farmland, and the possible disaster-stricken farmland is cropped and removed from the satellite image to be predicted that only contains the farmland area and its elevation. The remaining area blocked by clouds in the study area is taken as the actual prediction area, and the area not blocked by clouds is taken as the reference area.

[0054] The damage value R of farmland is as follows:

[0055]

[0056] Among them, α1, α2, α3 and α4 are the weights of the first, second, third and fourth factors, respectively, which are determined according to historical waterlogging events; β1 and β2 are the first and second nonlinear indexes, respectively, which are used to simulate the nonlinear effects of soil permeability coefficient and terrain on the water accumulation cutoff; P is the cumulative precipitation in the study area on the date to be predicted and the preset time period before it; T is the precipitation duration in the study area on the date to be predicted and the preset time period before it; K is the soil permeability coefficient of the farmland in the study area, which is determined according to the soil type of the farmland; G represents the terrain influence coefficient, which is determined according to the elevation data of the farmland area. The higher the elevation, the lower the terrain influence coefficient, and the less likely it is to be waterlogged. The critical value of the disaster value R of the farmland and the size of each parameter are determined through historical waterlogging events, and then the waterlogging situation of each field in the study area is judged based on the generated empirical formula.

[0057] S5: Perform the inversion operation in step S3 on the satellite images of the actual prediction area blocked by clouds within a preset time period before the prediction date and the reference area not blocked by clouds within the prediction date and the preset time period before the prediction date to obtain their respective crop biomass data, and then obtain the crop biomass data of the actual prediction area blocked by clouds on the prediction date after comparison, so as to obtain the crop biomass distribution map of the prediction date and realize the crop biomass prediction of the cloud-blocked area.

[0058] For each pixel in the satellite image, the crop biomass data obtained from the satellite images of the reference area that is not blocked by clouds during the predicted date and the preset time period before the predicted date are arranged in the order of growth period to obtain the first biomass array A i [a1, a2, …, a n ], where a n , …, a2 and a1 represent the crop biomass data of the to-be-predicted date, …, n-1 days before the to-be-predicted date and n days before the to-be-predicted date of the i-th pixel in the satellite image respectively; the crop biomass data obtained from the satellite image of the actual prediction area blocked by clouds within the preset time period before the to-be-predicted date are arranged in sequence according to the growth period to obtain the second biomass array B i[ b1,b2,…,b n-1 ], where b n-1 , …, b2 and b1 represent the crop biomass data of the ith pixel in the satellite image 1 day before the predicted date, …, n-1 days before the predicted date and n days before the predicted date respectively; the second biomass array of each pixel in the satellite image is compared with the first n-1 elements in the first biomass array. If the first n-1 elements in one of the first biomass arrays are exactly the same as the second biomass array, the nth element of the first biomass array is used as the nth element of the second biomass array, that is, as the crop biomass data of the actual predicted area blocked by clouds on the predicted date. The crop biomass data of the actual predicted area blocked by clouds on the predicted date are fused with the actual predicted area to obtain the crop biomass distribution map of the predicted date.

[0059] If cloud-free satellite images are obtained for a later date, the predicted crop biomass value of the actual predicted area on the predicted date should be immediately corrected based on the later data.

[0060] As shown in Table 2, the accuracy evaluation of the first biomass prediction using the Stacking model (the average of 200 rounds of training results) is listed. It can be seen that the performance of the integrated learning model is better than that of a single machine learning algorithm. Figure 8 (a) Figure 8 (b) Figure 8 (c) and Figure 8 As shown in (d), the superior performance of the integrated learning model is more intuitively reflected. The violin shape of the model of the present invention is better than that of a single machine learning model, that is, the model is more stable.

[0061] Table 2

[0062]

[0063] The above description of the embodiments is to facilitate the understanding and application of the present invention by those skilled in the art. It is obvious that those skilled in the art can easily make various modifications to the above embodiments and apply the general principles described herein to other embodiments without creative work. Therefore, the present invention is not limited to the above embodiments, and improvements and modifications made by those skilled in the art based on the disclosure of the present invention should be within the scope of protection of the present invention.

Claims

1. A method for predicting crop biomass in cloud-blocked areas of satellite images in rainy seasons, characterized in that: include: S1: Build a farmland image segmentation model, obtain historical satellite images of the key growth period of field crops in the study area that are not blocked by clouds, and train the farmland image segmentation model to obtain a trained image segmentation model; S2: Obtain satellite images, drone images and ground sampling data for the key growth period of field crops in the study area; input the satellite images to be predicted into the trained farmland image segmentation model and output the farmland mask after processing; process the satellite images to be predicted based on the farmland mask and combined with the elevation information to obtain the satellite images to be predicted that only contain the farmland area and its elevation; S3: constructing a crop biomass inversion model at different growth stages according to the UAV image and ground sampling data in step S2, thereby obtaining crop biomass data by inverting the crop biomass inversion model at different growth stages, the UAV image, and the satellite image to be predicted that only includes the farmland area and its elevation; S4: The date blocked by clouds in the satellite image to be predicted that only contains the farmland area and its elevation is taken as the date to be predicted, and the farmland that may be affected in the study area is determined based on the precipitation data in the preset time period before the date to be predicted and the elevation data of the farmland area, and then removed to obtain the satellite image of the actual prediction area blocked by clouds and the reference area not blocked by clouds; S5: Perform the inversion operation in step S3 on the satellite images of the actual prediction area blocked by clouds within a preset time period before the prediction date and the reference area not blocked by clouds within the prediction date and the preset time period before the prediction date to obtain their respective crop biomass data, and then obtain the crop biomass data of the actual prediction area blocked by clouds on the prediction date after comparison, so as to obtain the crop biomass distribution map of the prediction date and realize the crop biomass prediction of the cloud-blocked area.

2. The method for predicting crop biomass in cloud-blocked areas of satellite images in rainy seasons according to claim 1, characterized in that: The step S1 is specifically as follows: S11: Replace the ResNet module in the backbone layer of the pyramid attention network PAN with a lightweight MobileNetV1 network to build a farmland image segmentation model; S12: For each historical satellite image that is not blocked by clouds during the key growth period of field crops in the study area, it is uniformly cropped into a number of square cropped satellite images of the same size. The labels of the farmland coverage area and other areas are marked in each cropped satellite image and divided into training set and validation set according to the preset ratio; S13: The binary cross entropy loss and the Dice coefficient loss are combined into a hybrid loss function to construct a farmland image segmentation model. The training set and the validation set are input into the farmland image segmentation model for training until the hybrid loss function converges to obtain a trained image segmentation model.

3. The method for predicting crop biomass in cloud-blocked areas of satellite images in rainy seasons according to claim 1, characterized in that: In step S2, the satellite images are panchromatic and multispectral images of the high-resolution series satellites; the drone images are visible light RGB and multispectral images; the ground sampling data are the biomass data of field crops at each preset sampling point in the study area; and the key growth period of field crops includes no less than two different growth periods.

4. The method for predicting crop biomass in cloud-blocked areas of satellite images in rainy seasons according to claim 1, characterized in that: In step S2, farmland extraction is performed on the satellite image to be predicted according to the farmland mask, and then the average elevation information of each farmland in the satellite image to be predicted is obtained in combination with the satellite digital elevation model DEM, so as to obtain the satellite image to be predicted that only contains the farmland area and its elevation.

5. The method for predicting crop biomass in cloud-blocked areas of satellite images in rainy seasons according to claim 3, characterized in that: In the step S3, firstly, various vegetation indexes and visible light image texture features at various preset sampling points are obtained according to the drone image in step S2. For each preset sampling point, the various vegetation indexes at the preset sampling point and the biomass data of field crops in the ground sampling data are subjected to correlation regression analysis and processing, and several vegetation indexes with correlations higher than a preset threshold are screened out. According to the various screened vegetation indexes and visible light image texture features and using an integrated learning regression algorithm, an inversion model for the biomass of crops in different growth periods is constructed. The inversion model for the biomass of crops in different growth periods obtains the various different growth indexes in the key growth period of field crops at the drone scale according to the drone image inversion. The biomass distribution of field crops in the key growth period is taken as the first crop biomass data, and then the first crop biomass data is interpolated and upscaled to match the resolution of the satellite image, and the upscaled first crop biomass data is used as the true value of the satellite-scale biomass inversion, and the satellite image to be predicted containing only the farmland area and its elevation in step S2 and the true value of the satellite-scale biomass inversion are subjected to the same operation as the biomass data of field crops in the UAV image and ground sampling data in step S3, so as to invert the biomass distribution of field crops in different growth periods in the key growth period of field crops at the satellite scale as the final crop biomass data.

6. The method for predicting crop biomass in cloud-blocked areas of satellite images in rainy seasons according to claim 3, characterized in that: In the step S4, precipitation data on the predicted date and the preset time period before it in the study area is obtained, including precipitation amount, precipitation time and precipitation duration, and the disaster value of each farmland in the study area is determined based on the elevation data of the farmland area. The farmland whose disaster value exceeds the disaster threshold value is regarded as the possible disaster-stricken farmland, and the possible disaster-stricken farmland is cropped and removed from the satellite image to be predicted that only contains the farmland area and its elevation. The remaining area blocked by clouds in the study area is used as the actual prediction area, and the area not blocked by clouds is used as the reference area.

7. The method for predicting crop biomass in cloud-blocked areas of satellite images in rainy seasons according to claim 3, characterized in that: The damage value R of the farmland is as follows: Among them, α1, α2, α3 and α4 are the weights of the first, second, third and fourth factors respectively; β1 and β2 are the first and second nonlinear indexes respectively; P is the cumulative precipitation in the study area during the predicted date and the preset time period before it; T is the precipitation duration in the study area during the predicted date and the preset time period before it; K is the soil permeability coefficient of the farmland in the study area; G represents the terrain influence coefficient, which is determined according to the elevation data of the farmland area.

8. The method for predicting crop biomass in cloud-blocked areas of satellite images in rainy seasons according to claim 3, characterized in that: In step S5, for each pixel in the satellite image, the crop biomass data obtained from the satellite image of the reference area that is not blocked by clouds during the preset time period before the predicted date are arranged in order of growth period to obtain a first biomass array A. i [a1, a2, …, a n ], where a n , …, a2 and a1 represent the crop biomass data of the to-be-predicted date, …, n-1 days before the to-be-predicted date and n days before the to-be-predicted date of the i-th pixel in the satellite image respectively; the crop biomass data obtained from the satellite image of the actual prediction area blocked by clouds within the preset time period before the to-be-predicted date are arranged in sequence according to the growth period to obtain the second biomass array B i[ b1,b2,…,b n-1 ], where b n-1 , …, b2 and b1 represent the crop biomass data of the ith pixel in the satellite image 1 day before the predicted date, …, n-1 days before the predicted date and n days before the predicted date respectively; the second biomass array of each pixel in the satellite image is compared with the first n-1 elements in the first biomass array. If the first n-1 elements in one of the first biomass arrays are exactly the same as the second biomass array, the nth element of the first biomass array is used as the nth element of the second biomass array, that is, as the crop biomass data of the actual predicted area blocked by clouds on the predicted date. The crop biomass data of the actual predicted area blocked by clouds on the predicted date are fused with the actual predicted area to obtain the crop biomass distribution map of the predicted date.

Images