A vegetation distribution detection method and system based on radar imaging

Abstract

The present invention provides a vegetation distribution detection method and system based on radar imaging, which generates vegetation coding features by fusing vegetation reflectivity sequence data and vegetation coverage sequence data in different radar detection cycles, and encoding and representing them respectively. Subsequently, the vegetation coding features are fused with the corresponding sequence data and loaded into the corresponding vegetation distribution detection network for vegetation distribution detection, which can not only generate the accurate vegetation reflectivity of the target vegetation area at the target cycle observation node, but also determine the growth trend parameters of the target vegetation. By combining the vegetation reflectivity and the growth trend parameters, the vegetation distribution state parameters of the target vegetation area at the target cycle observation node can be comprehensively and accurately determined, which improves the accuracy and reliability of the detection compared with the prior art.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a vegetation distribution detection method and system based on radar imaging. Background Art

[0002] Vegetation distribution detection (such as bamboo vegetation distribution detection) plays a vital role in many fields such as ecology, agricultural management, and environmental monitoring. Traditional vegetation distribution detection methods often rely on ground field surveys or remote sensing image analysis, but these methods have many limitations. Ground field surveys are time-consuming and laborious, and are limited by geographical and weather conditions, making it difficult to achieve large-scale, high-frequency monitoring. Although remote sensing image analysis can provide vegetation information over a large range, its accuracy and timeliness are also limited by factors such as image resolution and cloud cover.

[0003] With the development of radar technology, radar imaging technology has gradually become an important means of vegetation distribution detection. Radar can penetrate clouds and achieve all-weather and all-day monitoring, and its high-resolution imaging capability enables the detailed features of vegetation to be clearly presented. However, the existing technology still has some shortcomings. For example, the data within a single radar detection cycle may be interfered by various factors, resulting in inaccurate detection results; at the same time, there is a lack of effective analysis and prediction of vegetation changes in different time periods. Summary of the invention

[0004] In view of the above-mentioned problems, in combination with the first aspect of the present invention, an embodiment of the present invention provides a vegetation distribution detection method based on radar imaging, the method comprising:

[0005] Acquire first vegetation reflectivity sequence data and first vegetation coverage sequence data of a target vegetation area within a first radar detection cycle, and acquire second vegetation reflectivity sequence data and second vegetation coverage sequence data of a control vegetation area within a second radar detection cycle, wherein the control vegetation area and the target vegetation area are vegetation areas with the same ecological conditions, and the first radar detection cycle and the second radar detection cycle are different detection cycles;

[0006] The first vegetation reflectance sequence data and the second vegetation reflectance sequence data are fused and encoded to generate a first vegetation coding feature, and the first vegetation coverage sequence data and the second vegetation coverage sequence data are fused and encoded to generate a second vegetation coding feature;

[0007] The first vegetation coding feature and the first vegetation coverage sequence data are fused and loaded into a first vegetation distribution detection network for vegetation distribution detection to generate a target vegetation reflectivity of the target vegetation area at a target period observation node;

[0008] fusing the second vegetation coding feature and the first vegetation reflectance sequence data, loading them into a second vegetation distribution detection network for vegetation distribution detection, and determining the target vegetation growth trend parameter of the target vegetation area at the target period observation node according to the detection result of the second vegetation distribution detection network;

[0009] The target vegetation distribution state parameters of the target vegetation area at the target period observation node are determined according to the target vegetation reflectivity and the target vegetation growth trend parameters.

[0010] On the other hand, an embodiment of the present invention also provides a vegetation distribution detection system based on radar imaging, including a processor and a machine-readable storage medium, wherein the machine-readable storage medium is connected to the processor, the machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0011] Based on the above aspects, the embodiment of the present application generates vegetation coding features by fusing vegetation reflectivity sequence data and vegetation coverage sequence data within different radar detection cycles, and encoding and representing them respectively. Subsequently, the vegetation coding features are fused with the corresponding sequence data and loaded into the corresponding vegetation distribution detection network for vegetation distribution detection. This can not only generate the accurate vegetation reflectivity of the target vegetation area at the target period observation node, but also determine the growth trend parameters of the target vegetation. By combining vegetation reflectivity and growth trend parameters, the vegetation distribution state parameters of the target vegetation area at the target period observation node can be comprehensively and accurately determined, which improves the accuracy and reliability of detection compared to the prior art. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 It is a schematic diagram of the execution flow of the vegetation distribution detection method based on radar imaging provided by an embodiment of the present invention.

[0013] Figure 2 It is a schematic diagram of the hardware architecture of a vegetation distribution detection system based on radar imaging provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0014] The present invention will be described in detail below with reference to the accompanying drawings. Figure 1 It is a flow chart of a vegetation distribution detection method based on radar imaging provided by an embodiment of the present invention. The vegetation distribution detection method based on radar imaging is introduced in detail below.

[0015] Step S110, obtaining the first vegetation reflectivity sequence data and the first vegetation coverage sequence data of the target vegetation area within the first radar detection cycle, and obtaining the second vegetation reflectivity sequence data and the second vegetation coverage sequence data of the control vegetation area within the second radar detection cycle, the control vegetation area and the target vegetation area are vegetation areas with the same ecological conditions, and the first radar detection cycle and the second radar detection cycle are different detection cycles.

[0016] In an exemplary scenario, the server is monitoring a large agricultural planting area, in which there is a specific target vegetation area, such as a bamboo field planted with square bamboo, with an area of ​​1,000 square meters. The server sets a first radar detection cycle, which can be from March 1 to May 31 of a certain year, a total of 3 months. During this first radar detection cycle, the server continuously collects data through radar equipment deployed near the bamboo field. The radar equipment can detect the reflectivity of the square bamboo vegetation, and each detection will obtain a data point. After continuous detection in this cycle, the first vegetation reflectivity sequence data is formed. For example, the reflectivity may be detected to be 0.3 on March 1, 0.32 on March 2, 0.31 on March 3, and so on, thus forming a reflectivity sequence data arranged in chronological order within this time period.

[0017] At the same time, the server can also determine the vegetation coverage based on radar data or in combination with other image analysis technologies. Vegetation coverage refers to the percentage of the vertical projection area of ​​vegetation (including leaves, stems, and branches) on the ground to the total area of ​​the statistical area. For this bamboo field, a vegetation coverage data will be obtained every day during these three months. These data are arranged in chronological order to form the first vegetation coverage sequence data. For example, the vegetation coverage is 10% on March 1, and as the bamboo grows, it may reach 80% by May 31.

[0018] In addition, in this large agricultural planting area, there is a control vegetation area, which is also planted with square bamboo, with an area of ​​800 square meters, and has the same ecological conditions as the target vegetation area, which means that they have similar soil types, altitudes, surrounding water sources, and similar climate influence ranges. The server sets a second radar detection cycle for the control vegetation area, which is from April 1 to June 30 of the same year, also for 3 months. Similarly, during this cycle, the server collects the second vegetation reflectivity sequence data and the second vegetation coverage sequence data of the control vegetation area through the corresponding equipment. For example, on April 1, the vegetation reflectivity of the control area was 0.33, and the vegetation coverage was 12%. As time goes by, new data is continuously generated and formed into a sequence. Since the first radar detection cycle and the second radar detection cycle are different detection cycles, this enables the server to obtain data from the target vegetation area and the control vegetation area from different time spans, providing a more comprehensive basis for subsequent analysis.

[0019] Step S120, the first vegetation reflectance sequence data and the second vegetation reflectance sequence data are fused and encoded to generate a first vegetation coding feature, and the first vegetation coverage sequence data and the second vegetation coverage sequence data are fused and encoded to generate a second vegetation coding feature.

[0020] In this embodiment, after the first vegetation reflectance sequence data of the target vegetation area and the second vegetation reflectance sequence data of the reference vegetation area are acquired, the fusion and encoding operations are started to generate the first vegetation coding feature.

[0021] First, the server splices the first vegetation reflectivity sequence data and the second vegetation reflectivity sequence data to generate the first fusion space. Assume that the first vegetation reflectivity sequence data has 100 data points (corresponding to 100 detections of the target vegetation area in the first radar detection cycle), and the second vegetation reflectivity sequence data has 90 data points (corresponding to the number of detections of the control vegetation area in the second radar detection cycle). The server splices the two sequence data together in a certain order, for example, arranging the first vegetation reflectivity sequence data first, and then arranging the second vegetation reflectivity sequence data, so as to construct a new spatial structure, namely the first fusion space.

[0022] Next, the server performs filtering operations on the first fusion space through a variety of filter groups to generate multiple first filtering result features of the same dimension. Assume that the server has three different filters, namely low-pass filter, high-pass filter and band-pass filter. For the data in the first fusion space, the low-pass filter may filter out the high-frequency noise information and retain the overall trend information of the low frequency to obtain a filtered result feature; the high-pass filter is the opposite, it highlights the high-frequency change part in the data and obtains another different filtering result feature; the band-pass filter screens out the data features of the middle frequency band and also obtains a corresponding filtering result feature. Although these three filtering result features are obtained through different filtering methods, they have the same dimension, which is convenient for subsequent operations.

[0023] Then, the server integrates the members of the same spatial point in each first filtering result feature to obtain multiple first combination features. For example, for the first data point in the first fusion space, the corresponding value in the low-pass filtering result feature, the corresponding value in the high-pass filtering result feature, and the corresponding value in the band-pass filtering result feature are integrated together to form a new combination feature; the same operation is performed for the second data point, and so on, so that multiple first combination features are obtained.

[0024] Finally, the server encodes the multiple first combination features to generate the first vegetation coding feature. The server loads the multiple first combination features into the first recurrent neural network for time series feature mapping, and generates the first potential features corresponding to each first combination feature. For example, for the first first combination feature, after being processed by the first recurrent neural network, it is mapped into a new potential feature representation according to its front and back relationship in the time series, and this potential feature can better reflect the significance of the combination feature in the entire time series. Then, the server assigns the focus attention coefficient to each first potential feature and generates the focus attention coefficient corresponding to each first potential feature. This is like giving each potential feature a weight to indicate its importance in the whole. For example, if a potential feature is very critical in the trend prediction of the overall data, it may be assigned a higher focus attention coefficient. Finally, the first potential features are fused and calculated based on these focus attention coefficients to generate the first vegetation coding feature.

[0025] For the first vegetation coverage sequence data and the second vegetation coverage sequence data, the server uses similar steps to generate the second vegetation coding feature. First, the two are spliced ​​to form a fusion space, and then the filtering operation is performed to obtain the filtering result feature, and then the members of the same spatial point are integrated to obtain the combined feature, and finally the second vegetation coding feature is generated through the operation of the recurrent neural network and the focusing attention coefficient.

[0026] Step S130, fusing the first vegetation coding feature and the first vegetation coverage sequence data, and loading them into a first vegetation distribution detection network for vegetation distribution detection, to generate a target vegetation reflectivity of the target vegetation area at a target period observation node.

[0027] In this embodiment, after the first vegetation coding feature is generated and the first vegetation coverage sequence data has been acquired, they are fused and loaded into the first vegetation distribution detection network for vegetation distribution detection to generate the target vegetation reflectivity of the target vegetation area at the target period observation node.

[0028] First, the server obtains the first ecological condition characteristics corresponding to the target vegetation area. For the target vegetation area of ​​the bamboo field mentioned earlier, the first ecological condition characteristics may include information such as soil nutrient content, soil moisture, surrounding temperature, and light intensity. For example, the specific content values ​​of nutrients such as nitrogen, phosphorus, and potassium in the soil, the current soil moisture is 30% (relative humidity value obtained by some measurement method), the surrounding average temperature is 20 degrees Celsius, and the light intensity is 1000 lux.

[0029] Then, the server encodes the first ecological condition feature to generate a third vegetation coded feature. This encoding process may involve converting the numerical values ​​of each ecological condition into a vector representation that can be processed by the computer according to certain rules. For example, the numerical value of the soil nutrient content is normalized and mapped to a specific vector space, and the temperature, light intensity, etc. are also processed similarly, and then these vectors are combined to form the third vegetation coded feature.

[0030] Next, the server fuses the third vegetation coding feature, the first vegetation coding feature and the first vegetation coverage sequence data. This fusion process may be to combine the three data structures according to certain rules, such as splicing them in dimensions or merging them by some weighted summation method.

[0031] Finally, the server loads the fused data into the first vegetation distribution detection network for vegetation distribution detection. The first vegetation distribution detection network is a pre-trained neural network model that can predict the reflectivity of vegetation based on the input feature data. Assume that this network has multiple hidden layers, and the data is calculated and transmitted through layers of neurons in the network. For example, the input data first enters the first hidden layer, and the neurons process the data according to the preset weights and activation functions, and then pass the processed results to the next hidden layer. After multiple such operations, the target vegetation reflectivity of the target vegetation area at the target period observation node is finally obtained in the output layer. For example, the target period observation node is May 15th. After calculation by the first vegetation distribution detection network, the target vegetation reflectivity of this day is 0.4.

[0032] Step S140, the second vegetation coding feature and the first vegetation reflectance sequence data are fused and loaded into a second vegetation distribution detection network for vegetation distribution detection, and the target vegetation growth trend parameters of the target vegetation area at the target period observation node are determined according to the detection results of the second vegetation distribution detection network.

[0033] In this embodiment, after generating the second vegetation coding feature and acquiring the first vegetation reflectance sequence data, they are fused and loaded into the second vegetation distribution detection network for vegetation distribution detection, and the target vegetation growth trend parameters of the target vegetation area at the target period observation node are determined based on the detection results.

[0034] First, the server obtains the second ecological condition feature corresponding to the target vegetation area. For the target vegetation area of ​​the bamboo field, the second ecological condition feature may contain some information that is not completely the same as the first ecological condition feature, such as wind direction, wind speed, and the influence of other crops in the surrounding area. For example, the wind direction is southeast wind, the wind speed is 2 meters per second, there are a small number of other weed crops in the surrounding area, and the coverage area of ​​these weed crops accounts for 5%.

[0035] Then, the server encodes the second ecological condition feature to generate a fourth vegetation coded feature. This encoding process is similar to the previous encoding of the first ecological condition feature, converting information such as wind direction, wind speed, and the impact of other surrounding crops into a suitable vector representation, and then combining them into the fourth vegetation coded feature.

[0036] Next, the server aggregates the fourth vegetation coded feature, the second vegetation coded feature, and the first vegetation reflectance sequence data to generate input feature data. This aggregation process may be to combine these data together according to a certain logic, for example, firstly perform some form of splicing on the second vegetation coded feature and the first vegetation reflectance sequence data, and then add the fourth vegetation coded feature to the splicing result in a weighted manner, thereby forming a complete input feature data structure.

[0037] Afterwards, the server loads the input feature data into the second vegetation distribution detection network. The second vegetation distribution detection network is also a trained neural network model, which has its own mapping unit. When the input feature data enters the network, the input feature data is mapped according to the mapping unit to generate a category mapping result. This category mapping result represents the probability scores of multiple reference vegetation growth matching models. For example, assuming that there are three reference vegetation growth matching models, namely model A, model B and model C, after network mapping, the probability score of model A is 0.3, the probability score of model B is 0.5, and the probability score of model C is 0.2.

[0038] Then, the server determines the target vegetation growth matching model from the multiple reference vegetation growth matching models according to the category mapping result. In the above example, since the probability score of model B is the highest, model B is determined as the target vegetation growth matching model.

[0039] Finally, the server loads the initial node of the periodic observation of the first radar detection cycle, the target periodic observation node, and each periodic observation node between the initial node of the periodic observation and the target periodic observation node into the target vegetation growth matching model. For example, the initial node of the periodic observation of the first radar detection cycle is March 1, the target periodic observation node is May 15, and there are multiple periodic observation nodes such as March 15 and April 1 in the middle. The data of these nodes are sequentially input into model B, and model B generates multiple target vegetation growth parameter sequences according to its own algorithm. For example, a sequence of plant height growth parameters of square bamboo, a sequence of leaf number growth parameters of square bamboo, etc. may be generated. Based on the fusion results of these multiple target vegetation growth parameter sequences, the target vegetation growth trend parameters of the target vegetation area at the target periodic observation node are generated. For example, by weighted summation or other fusion algorithms of plant height growth parameter sequences and leaf number growth parameter sequences, the target vegetation growth trend parameters are obtained, which can be used to indicate whether the overall growth trend of square bamboo on May 15 is fast growth or slow growth.

[0040] Step S150, determining the target vegetation distribution state parameters of the target vegetation area at the target period observation node according to the target vegetation reflectivity and the target vegetation growth trend parameters.

[0041] In this embodiment, after the target vegetation reflectivity and the target vegetation growth trend parameter are obtained, the target vegetation distribution state parameter of the target vegetation area at the target period observation node is determined based on these two parameters.

[0042] Assuming that the reflectivity of the target vegetation is 0.4, the growth trend parameter of the target vegetation indicates that the square bamboo is in a rapid growth stage at the target period observation node (May 15). The server determines the target vegetation distribution state parameter according to the pre-set rules and algorithms. If the reflectivity of the target vegetation is between 0.3 and 0.5, and the growth trend is rapid growth, the server may combine some historical data and empirical models to determine the vegetation distribution state of the target vegetation area. For example, according to historical data, when the reflectivity is in this range and the growth is rapid, the distribution of vegetation may be relatively uniform, and the density of vegetation is also within a suitable range, then the server can determine that the target vegetation distribution state parameter is "uniform and appropriate density". This target vegetation distribution state parameter can be further used for agricultural production decision-making, such as whether irrigation, fertilization, or pest control operations are required. If the target vegetation distribution state parameter shows that the vegetation distribution is uneven or the density is too high or too low, corresponding measures can be taken to improve the growth of vegetation. In this way, the vegetation status of the target vegetation area can be comprehensively analyzed, and data can be acquired and processed from different angles.

[0043] Based on the above steps, the embodiment of the present application generates vegetation coding features by fusing vegetation reflectivity sequence data and vegetation coverage sequence data within different radar detection cycles, and encoding and representing them respectively. Subsequently, the vegetation coding features are fused with the corresponding sequence data and loaded into the corresponding vegetation distribution detection network for vegetation distribution detection. This can not only generate the accurate vegetation reflectivity of the target vegetation area at the target period observation node, but also determine the growth trend parameters of the target vegetation. By combining vegetation reflectivity and growth trend parameters, the vegetation distribution state parameters of the target vegetation area at the target period observation node can be comprehensively and accurately determined, which improves the accuracy and reliability of detection compared to the prior art.

[0044] In a possible implementation, step S120 includes:

[0045] Step S121: splicing the first vegetation reflectance sequence data and the second vegetation reflectance sequence data to generate a first fusion space.

[0046] Step S122: performing filtering operations on the first fusion space respectively through a diversified filter group to generate a plurality of first filtering result features of the same dimension.

[0047] Step S123: Integrate the members of the same spatial point in each of the first filtering result features to obtain a plurality of first combined features, and encode the plurality of the first combined features to generate a first vegetation coded feature.

[0048] In a possible implementation, step S123 includes:

[0049] Step S1231: Load the plurality of first combined features into a first recurrent neural network for time series feature mapping to generate first potential features corresponding to each of the first combined features.

[0050] Step S1232, assigning a focus attention coefficient to each of the first potential features to generate a focus attention coefficient corresponding to each of the first potential features, performing a fusion calculation on each of the first potential features according to the focus attention coefficient to generate a first vegetation coding feature.

[0051] In a possible implementation manner, the reference vegetation area includes at least two, and the second radar detection cycles corresponding to the respective reference vegetation areas are different detection cycles. Step S121 includes:

[0052] Step S1211, splicing the second vegetation reflectivity sequence data corresponding to each of the reference vegetation areas according to the periodic order of each of the second radar detection periods to generate a candidate fusion space.

[0053] Step S1212: determining the embedding spatial point of the first vegetation reflectivity sequence data in the candidate fusion space according to the first radar detection cycle and the cycle order of each of the second radar detection cycles.

[0054] Step S1213: embedding the first vegetation reflectance sequence data into the candidate fusion space according to the embedding space point to generate a first fusion space.

[0055] In this embodiment, for the process of encoding and representing the generation of the first vegetation coding feature after fusing the first vegetation reflectivity sequence data and the second vegetation reflectivity sequence data, it is assumed that the target vegetation area is a specific square bamboo planting area, and the first radar detection period is set from the 10th day to the 50th day after sowing. During this first radar detection period, the server continuously obtains the first vegetation reflectivity sequence data through the radar detection device. For example, the reflectivity on the 10th day is 0.2, and on the 11th day is 0.21. As the square bamboo grows and is affected by environmental factors, a series of data arranged in chronological order is obtained until the 50th day.

[0056] At the same time, there are at least two control vegetation areas, both of which are bamboo planting areas and have the same ecological conditions as the target vegetation area. Assume that the second radar detection period of one of the control vegetation areas is from the 15th day to the 55th day after sowing, and the second radar detection period of the other control vegetation area is from the 20th day to the 60th day after sowing. The server obtains the second vegetation reflectivity sequence data of each control vegetation area within its respective second radar detection period.

[0057] First, the first fusion space is generated. The server splices the second vegetation reflectivity sequence data corresponding to each control vegetation area according to the periodic order of each second radar detection cycle. For example, the reflectivity data of the first control vegetation area is arranged in chronological order, and then the reflectivity data of the second control vegetation area is arranged, so that the candidate fusion space is generated. Then, based on the first radar detection cycle and the periodic order of each second radar detection cycle, the embedding space point of the first vegetation reflectivity sequence data in the candidate fusion space is determined. For example, the appropriate embedding point is determined by comparing the relative time position of the start and end of the cycle. Finally, based on this embedding space point, the first vegetation reflectivity sequence data is embedded into the candidate fusion space, thereby generating the first fusion space.

[0058] Next, the first fusion space is filtered by a variety of filter groups. Assume that the filter group includes a low-pass filter, a band-pass filter, and a high-pass filter. Taking the low-pass filter as an example, it can smooth out the high-frequency fluctuations in the first fusion space data, retain the overall low-frequency trend, and obtain a low-pass filter result feature. The band-pass filter screens out the data features of the middle frequency band, and the high-pass filter highlights the high-frequency change part, and obtains the respective filter result features, and these first filter result features have the same dimension.

[0059] Afterwards, the members of the same spatial point in each first filtering result feature are integrated to obtain multiple first combination features. For each spatial point in the first fusion space, its corresponding values ​​in the low-pass, band-pass and high-pass filtering result features are integrated to form multiple first combination features.

[0060] Then, the plurality of first combination features are encoded to represent the generation of the first vegetation coding feature. The server loads the plurality of first combination features into the first recurrent neural network for time series feature mapping. For each first combination feature, the first recurrent neural network performs mapping according to the order relationship in the time series to generate the first potential feature corresponding to each first combination feature. For example, the early data and late data of a first combination feature in the time series are converted into potential features that can reflect the law of change of its time series after passing through the recurrent neural network.

[0061] Then, each first potential feature is assigned a focus attention coefficient. This process is based on factors such as the importance of the data and its relevance to the target. For example, the first potential feature corresponding to the key growth stage of square bamboo may be assigned a higher focus attention coefficient, while the first potential feature corresponding to relatively minor factors such as environmental fluctuations may be assigned a lower coefficient. Based on these focus attention coefficients, each first potential feature is fused and calculated to finally generate the first vegetation coding feature. This first vegetation coding feature integrates the vegetation reflectance information of the target vegetation area and the control vegetation area. After multi-step processing, it can more effectively reflect the characteristics of the vegetation and provide data support for subsequent operations such as vegetation distribution detection.

[0062] In terms of the training step of the first vegetation distribution detection network, it is assumed that the first template vegetation area obtained by the server is a specific soybean planting area. The third radar detection cycle is set to a specific stage of soybean growth, such as from the 20th day to the 60th day after sowing, and the fourth radar detection cycle is another related stage, such as from the 30th day to the 70th day after sowing.

[0063] The server obtains the third vegetation reflectivity sequence data and the third vegetation coverage sequence data of the first template vegetation area in the third radar detection cycle, and obtains the fourth vegetation reflectivity sequence data of the second template vegetation area in the fourth radar detection cycle. Here, the second template vegetation area and the first template vegetation area are soybean planting areas with the same ecological conditions.

[0064] When obtaining the third vegetation reflectivity sequence data, it is possible to first obtain the reference vegetation reflectivity sequence data and reference vegetation coverage sequence data of the first template vegetation area within a longer reference radar detection period (such as from the 10th day to the 80th day after sowing). Then, the third vegetation reflectivity sequence data is generated by performing roaming collection in the reference vegetation reflectivity sequence data through the first roaming window. For example, the roaming unit parameter of the first roaming window is set to 5 days, and data is collected once every 5 days to obtain the third vegetation reflectivity sequence data of a suitable scale. For the third vegetation coverage sequence data, it is generated in a similar roaming collection manner in the reference vegetation coverage sequence data through the second roaming window, and its continuous detection period is ensured to be consistent with the requirements.

[0065] When determining the first template vegetation area, the template ecological condition can be determined from multiple reference ecological conditions, and the vegetation area updated in accordance with the template ecological condition can be output as the first template vegetation area; or the vegetation areas updated by multiple reference ecological conditions can be output as reference vegetation areas corresponding to each reference ecological condition, and each reference vegetation area can be extracted respectively, and the first template vegetation area can be generated based on the extraction result. Finally, the first vegetation distribution detection network is optimized based on the third vegetation reflectance sequence data, the third vegetation coverage sequence data, and the fourth vegetation reflectance sequence data, so that the network can better adapt to the analysis of vegetation reflectance and coverage data, and improve the accuracy of vegetation distribution detection in the target vegetation area.

[0066] For the training steps of the second vegetation distribution detection network. Taking the soybean planting area as an example, the server obtains the third vegetation coverage sequence data and the fifth vegetation reflectivity sequence data of the first template vegetation area in the third radar detection cycle, and obtains the fourth vegetation coverage sequence data of the second template vegetation area in the fourth radar detection cycle. The various data acquisition methods and cycle settings here are similar to the relevant concepts in the training of the first vegetation distribution detection network, but the focus is on the vegetation coverage and reflectivity data in different cycles. Through these data, the second vegetation distribution detection network is optimized according to a certain algorithm, so that it can accurately perform vegetation distribution detection according to the input data and determine the vegetation growth trend parameters.

[0067] In the process of fusing the first vegetation coding feature and the first vegetation coverage sequence data and loading them into the first vegetation distribution detection network for vegetation distribution detection to generate the target vegetation reflectance of the target vegetation area at the target period observation node. Assume that the target period observation node of the target vegetation area (square bamboo planting area) is the 30th day after sowing. The server obtains the first ecological condition feature corresponding to the target vegetation area, such as soil moisture of 40%, soil fertility (nitrogen, phosphorus, potassium content, etc.), ambient air temperature of 25 degrees Celsius, and light duration of 12 hours. The first ecological condition feature is encoded to generate the third vegetation coding feature. This process involves converting the values ​​of each ecological condition into a vector form according to certain rules and combining them. Then the third vegetation coding feature, the first vegetation coding feature and the first vegetation coverage sequence data are fused. The fusion method can be to splice them in the data structure or add them according to weights. Finally, the fused data is loaded into the first vegetation distribution detection network. After being processed by the internal computing layer, the network finally outputs the target vegetation reflectance of the target vegetation area on the 30th day after sowing, for example, 0.3, according to the pre-trained model parameters.

[0068] After the second vegetation coding feature and the first vegetation reflectivity sequence data are fused, they are loaded into the second vegetation distribution detection network for vegetation distribution detection, and the target vegetation growth trend parameters of the target vegetation area at the target period observation node are determined according to the detection results of the second vegetation distribution detection network. Still taking the target period observation node of the square bamboo planting area as the 30th day after sowing as an example. The server obtains the second ecological condition feature corresponding to the target vegetation area, such as a wind speed of 2 meters per second, a southerly wind direction, and the distribution density of other surrounding plants (such as weeds). The second ecological condition feature is encoded to represent the generation of a fourth vegetation coding feature. The fourth vegetation coding feature, the second vegetation coding feature, and the first vegetation reflectivity sequence data are aggregated, for example, the second vegetation coding feature and the first vegetation reflectivity sequence data are first combined according to a certain logic, and then the fourth vegetation coding feature is added to form a complete input feature data. The input feature data is loaded into the second vegetation distribution detection network, and the mapping unit in the second vegetation distribution detection network maps the input feature data to obtain the probability scores of multiple reference vegetation growth matching models. Assuming that the probability scores of the three reference vegetation growth matching models are 0.3, 0.5 and 0.2 respectively, the model with the highest probability score (here is the model with a probability score of 0.5) is selected as the target vegetation growth matching model. The data of the initial node of the periodic observation of the first radar detection cycle (such as the 10th day after sowing), the target periodic observation node (the 30th day after sowing), and the periodic observation nodes between the initial node of the periodic observation and the target periodic observation node (such as the 15th day, the 20th day, etc.) are loaded into the target vegetation growth matching model respectively. The model generates multiple target vegetation growth parameter sequences according to its own algorithm, such as the plant height growth parameter sequence of square bamboo, the leaf number growth parameter sequence, etc. Finally, based on the fusion results of these multiple target vegetation growth parameter sequences, for example, by weighted summation, the target vegetation growth trend parameter of the target vegetation area on the 30th day after sowing is generated. This parameter can indicate whether the growth of square bamboo in this period is fast growing, slow growing or stagnant.

[0069] Finally, the target vegetation distribution state parameters of the target vegetation area at the target period observation node are determined based on the target vegetation reflectivity and the target vegetation growth trend parameters. For example, when the target vegetation reflectivity is 0.3 and the target vegetation growth trend parameters indicate that the square bamboo grows slowly, the server may conclude that the target vegetation distribution state parameters are slightly low vegetation density and sparse distribution based on pre-set rules and historical data. This conclusion can provide a basis for subsequent agricultural management decisions, such as whether replanting or adjustment of planting spacing is needed.

[0070] In a possible implementation, step S130 includes:

[0071] Step S131, obtaining a first ecological condition characteristic corresponding to the target vegetation area.

[0072] Step S132: Encode the first ecological condition feature to generate a third vegetation coding feature.

[0073] Step S133, fusing the third vegetation coding feature, the first vegetation coding feature and the first vegetation coverage sequence data and loading them into a first vegetation distribution detection network for vegetation distribution detection, and generating a target vegetation reflectivity of the target vegetation area at a target period observation node.

[0074] In this embodiment, it is further assumed that the target vegetation area is a specific square bamboo planting field. When the target vegetation reflectivity is generated, the server first obtains the first ecological condition feature corresponding to the target vegetation area. For this square bamboo planting field, the first ecological condition feature contains data from multiple aspects. In terms of soil, the server obtains that the soil humidity is 30%, and this humidity value is accurately measured by a soil moisture sensor. The sensor is distributed in different locations of the square bamboo field, and its average value is taken as a representative. The fertility of the soil is also taken into consideration. For example, the nitrogen content in the soil is 0.15 grams per kilogram of soil, the phosphorus content is 0.08 grams per kilogram of soil, and the potassium content is 0.1 grams per kilogram of soil. These values ​​are obtained by a soil composition analysis instrument. In terms of meteorological conditions, the server obtains the surrounding temperature data. The current average temperature is 15 degrees Celsius. This data comes from the temperature sensor installed at the meteorological station near the square bamboo field. Light conditions are equally important. The server obtains the current light intensity of 800 lux, which is measured by a light intensity sensor. The sensor can accurately sense the light intensity at different times and record it.

[0075] Next, the server encodes the first ecological condition feature to generate the third vegetation coding feature. For the data of 30% soil moisture, the server maps it to a specific numerical range according to the pre-set coding rules. Assume that the mapped value is 0.3 (within the numerical range of 0-1). The nitrogen, phosphorus, and potassium content of the soil are also converted according to their respective coding rules. For example, a nitrogen content of 0.15 grams per kilogram of soil may be encoded as 0.15 (within a specific coding range set according to soil fertility). The temperature of 15 degrees Celsius and the light intensity of 800 lux are also converted to corresponding values ​​according to their respective coding rules. Then, these encoded values ​​are combined in a certain order, such as soil moisture first, then nitrogen content, phosphorus content, potassium content, temperature, light intensity, etc., to form a vector, which is the third vegetation coding feature.

[0076] Finally, the server fuses the third vegetation coding feature, the first vegetation coding feature and the first vegetation coverage sequence data and loads them into the first vegetation distribution detection network for vegetation distribution detection, generating the target vegetation reflectivity of the target vegetation area at the target period observation node. Assume that the first vegetation coding feature is a multidimensional vector obtained through a series of previous processing, and the first vegetation coverage sequence data is a numerical sequence of bamboo vegetation coverage recorded in chronological order in the previous first radar detection cycle, such as 10% coverage on the 10th day and 15% coverage on the 15th day. The server operates the third vegetation coding feature, the first vegetation coding feature and the first vegetation coverage sequence data in a specific fusion mode. For example, the three data structures may be spliced ​​in dimension to form a new input data structure containing more comprehensive information. Then the fused data is loaded into the first vegetation distribution detection network. The first vegetation distribution detection network is a neural network model trained with a large amount of data, which contains multiple neuron layers. When the input data enters the network, the data is in the first layer of neurons, and each neuron linearly combines the input data according to the pre-learned weights and performs nonlinear transformation through the activation function. For example, a neuron performs a weighted summation on some dimensions of the input data, and then processes it through the ReLU activation function to obtain the output result. This output result is then passed to the next layer of neurons for similar operations. After processing by multiple layers of neurons, the target vegetation reflectance of the target vegetation area at the target period observation node is finally obtained in the output layer of the network. Assuming that the target period observation node is when the square bamboo grows to the 30th day, after calculation by the first vegetation distribution detection network, the target vegetation reflectance of this day is 0.35. This target vegetation reflectance can reflect the vegetation characteristics of the square bamboo at this specific target period observation node, and provide an important basis for the subsequent evaluation of the growth status of the square bamboo, agricultural management decisions, etc.

[0077] In a possible implementation, step S140 includes:

[0078] Step S141, obtaining a second ecological condition characteristic corresponding to the target vegetation area.

[0079] Step S142: Encode the second ecological condition feature to generate a fourth vegetation coding feature.

[0080] Step S143: Aggregate the fourth vegetation coding feature, the second vegetation coding feature and the first vegetation reflectivity sequence data to generate input feature data.

[0081] Step S144, loading the input feature data into a second vegetation distribution detection network, mapping the input feature data according to a mapping unit of the second vegetation distribution detection network, generating a category mapping result, wherein the category mapping result represents the probability scores of multiple reference vegetation growth matching models.

[0082] Step S145, determining a target vegetation growth matching model from a plurality of reference vegetation growth matching models according to the category mapping result, and loading the periodic observation initial node of the first radar detection period, the target periodic observation node, and each periodic observation node between the periodic observation initial node and the target periodic observation node into the target vegetation growth matching model, respectively, to generate a plurality of target vegetation growth parameter sequences.

[0083] Step S146: generating target vegetation growth trend parameters of the target vegetation area at the target period observation node based on the fusion results of the plurality of target vegetation growth parameter sequences.

[0084] In the present embodiment, for the target vegetation area of ​​the square bamboo planting field mentioned above, the second ecological condition feature includes data information of different aspects from before. For example, the server obtains the wind speed data of the area, and the wind speed is 3 meters per second. This data is accurately measured by an anemometer installed near the square bamboo field. The anemometer can continuously record the change of wind speed and transmit the data to the server. Wind direction information is also taken into consideration. The current wind direction is northwest wind, which is accurately measured by a wind direction sensor. In addition, about the influence of other plants around, the server detects the types and distribution of surrounding weeds, and finds that the main weed species is barnyard grass, and its distribution density in the square bamboo field is 5 strains per square meter. This data is obtained by combining image recognition technology with field sampling survey. Image recognition technology can analyze the aerial image of the square bamboo field, identify the types of weeds and estimate their approximate distribution density. Field sampling survey can accurately count specific areas, and the two are combined to obtain accurate weed distribution density data. There is also the air permeability of the soil. It has been measured that the air permeability of the soil is 5 ml per square centimeter per minute. This data is obtained by a special soil permeability measuring instrument. The instrument is inserted into the soil at different depths for measurement, and the average value is taken as the representative value.

[0085] Next, the server encodes the second ecological condition feature to generate a fourth vegetation coding feature. For the data of wind speed of 3 meters per second, according to the pre-set coding rules for wind speed, it is converted into a specific value, assuming that the converted value is 0.3 (within the coding interval set according to the wind speed range). The northwest wind may be encoded as a specific number or vector according to the coding rules of wind direction. For example, a three-dimensional vector can be used to represent different wind directions. The vector corresponding to the northwest wind is [- 0.5, 0.5, 0.707] (this is based on the pre-set wind direction vector representation rules). The weed distribution density of 5 plants per square meter may be converted to 0.05 according to its coding rules (within the coding interval set according to the weed density). The soil permeability of 5 milliliters per square centimeter per minute is also converted into a corresponding value according to its coding rules, assuming that it is 0.5 (within the coding interval set according to the soil permeability). These encoded values ​​are then combined in a certain order, such as wind speed, wind direction, weed distribution density, soil permeability, etc., to form a new vector, which is the fourth vegetation coding feature.

[0086] Afterwards, the server aggregates the fourth vegetation coding feature, the second vegetation coding feature and the first vegetation reflectivity sequence data to generate input feature data. Assume that the second vegetation coding feature is a data of a specific structure obtained by previous processing, such as a multidimensional vector, which contains feature information obtained from other related vegetation data processing. The first vegetation reflectivity sequence data is a numerical sequence of square bamboo vegetation reflectivity recorded in chronological order during the first radar detection cycle, such as the reflectivity on the 5th day is 0.2, the reflectivity on the 10th day is 0.22, etc. The server operates the fourth vegetation coding feature, the second vegetation coding feature and the first vegetation reflectivity sequence data according to a specific aggregation method. It may be that the second vegetation coding feature and the first vegetation reflectivity sequence data are first merged in some form, such as splicing in dimension or weighted summing according to certain weights. Then the fourth vegetation coding feature is added to the merged result in a specific way, such as vector addition of the fourth vegetation coding feature and the previous merged result or combination again according to weights, thereby generating a complete input feature data structure.

[0087] After that, the server loads the input feature data into the second vegetation distribution detection network. The second vegetation distribution detection network is a complex neural network model, which contains multiple neuron layers and a dedicated mapping unit. When the input feature data enters the network, the data first enters the first layer of neurons. In this layer, each neuron performs a linear combination of the input data according to the pre-learned weights. For example, a neuron performs a weighted summation of some dimensions of the input feature data. The calculation method is to multiply each dimension of the input feature data by the corresponding weight and then add them. Then, a nonlinear transformation is performed through an activation function, such as using a ReLU activation function, and the result after the weighted summation is processed. If the result is less than 0, 0 is output, and if it is greater than 0, the result itself is output. After being processed by the first layer of neurons, the output result is passed to the next layer of neurons for similar operations. In this way, after being processed layer by layer by multiple layers of neurons, the data reaches the mapping unit. The input feature data is mapped according to the mapping unit of the second vegetation distribution detection network to generate a category mapping result. This category mapping result represents the probability scores of multiple reference vegetation growth matching models. Assume that there are three reference vegetation growth matching models, namely model A, model B and model C. After being processed by the mapping unit, the probability score of model A is 0.3, the probability score of model B is 0.5, and the probability score of model C is 0.2.

[0088] Then, the server determines the target vegetation growth matching model from the multiple reference vegetation growth matching models according to the category mapping result. In the above example, since the probability score of model B is the highest, model B is determined as the target vegetation growth matching model.

[0089] Finally, the server loads the initial node of the periodic observation of the first radar detection cycle, the target periodic observation node, and each periodic observation node between the initial node of the periodic observation and the target periodic observation node into the target vegetation growth matching model. Assume that the initial node of the periodic observation of the first radar detection cycle is the first day of the growth of the square bamboo, the target periodic observation node is the 30th day, and the intermediate periodic observation nodes are the 5th day, the 10th day, the 15th day, the 20th day, and the 25th day. The data of these nodes are loaded into model B respectively. Model B processes the data of each periodic observation node according to its own algorithm structure and predefined parameters to generate multiple target vegetation growth parameter sequences. For example, a sequence of square bamboo plant height growth parameters may be generated, with a plant height of 5 cm on the first day, 8 cm on the 5th day, 12 cm on the 10th day, and so on until the 30th day; a sequence of square bamboo leaf number growth parameters may also be generated, with 1 leaf on the first day, 3 leaves on the 5th day, 5 leaves on the 10th day, and so on. Based on the fusion results of these multiple target vegetation growth parameter sequences, the target vegetation growth trend parameters of the target vegetation area at the target period observation node are generated. For example, for the plant height growth parameter sequence and the leaf number growth parameter sequence, the weighted summation method can be used for fusion. Assuming that the weight of the plant height growth parameter is 0.6 and the weight of the leaf number growth parameter is 0.4, the plant height and leaf number corresponding to each period observation node are calculated according to the weights and then summed to obtain a comprehensive numerical sequence. The value of this numerical sequence at the target period observation node (30th day) is the target vegetation growth trend parameter. This parameter can be used to indicate whether the overall growth trend of square bamboo on the 30th day is rapid growth, slow growth or stagnation, etc., and provides an important basis for subsequent square bamboo planting management decisions, such as whether it is necessary to adjust the amount of fertilizer, irrigation, or carry out pest and disease control operations.

[0090] In a possible implementation manner, the training step of the first vegetation distribution detection network includes:

[0091] Step A110, obtaining the third vegetation reflectivity sequence data and the third vegetation coverage sequence data of the first template vegetation area within the third radar detection cycle, obtaining the fourth vegetation reflectivity sequence data of the second template vegetation area within the fourth radar detection cycle, the second template vegetation area and the first template vegetation area are vegetation areas with the same ecological conditions, the third radar detection cycle and the fourth radar detection cycle are different detection cycles, the continuous detection period of the third vegetation reflectivity sequence data has a continuous unit of the periodic distance between the target periodic observation node and the periodic observation initial node of the first radar detection cycle, and the continuous detection period of the third vegetation coverage sequence data and the continuous detection period of the fourth vegetation reflectivity sequence data both have a continuous unit of the continuous detection period of the first vegetation reflectivity sequence data.

[0092] Step A120: Optimize the first vegetation distribution detection network according to the third vegetation reflectance sequence data, the third vegetation coverage sequence data, and the fourth vegetation reflectance sequence data.

[0093] In a possible implementation, step A110 includes:

[0094] Step A111, obtain the reference vegetation reflectivity sequence data and the reference vegetation coverage sequence data of the first template vegetation area within the reference radar detection cycle, the duration unit of the continuous detection cycle of the reference vegetation reflectivity sequence data is greater than the cycle distance between the target period observation node and the period observation initial node, and the duration unit of the continuous detection cycle of the reference vegetation coverage sequence data is greater than the continuous detection cycle of the first vegetation reflectivity sequence data.

[0095] Step A112, performing walking collection in the reference vegetation reflectance sequence data according to the first walking window to generate the third vegetation reflectance sequence data, wherein the walking unit parameter of the first walking window is smaller than the sequence scale of the reference vegetation reflectance sequence data.

[0096] Step A113, performing wandering collection in the reference vegetation coverage sequence data according to the second wandering window to generate the third vegetation coverage sequence data, wherein the wandering unit parameter of the second wandering window is smaller than the sequence scale of the reference vegetation coverage sequence data.

[0097] Before step A110, the method further includes:

[0098] Step A101, determining a template ecological condition from a plurality of reference ecological conditions, and outputting a vegetation area updated by the template ecological condition as the first template vegetation area.

[0099] Alternatively, in step A102, the vegetation areas updated by the multiple reference ecological conditions are output as reference vegetation areas corresponding to the reference ecological conditions, and each of the reference vegetation areas is extracted, and the first template vegetation area is generated based on the extraction results.

[0100] In this embodiment, first, regarding the acquisition of the first template vegetation area and related data, it is assumed that the server is processing a vegetation monitoring project in a large agricultural area, where there are vegetation areas with multiple different ecological conditions. When determining the first template vegetation area from multiple reference ecological conditions, there are two possible ways.

[0101] One way is to determine the template ecological condition from multiple reference ecological conditions, and output the vegetation area that meets the template ecological condition update as the first template vegetation area. For example, among many ecological conditions, the server determines a specific template ecological condition, which includes soil fertility (nitrogen content is 0.1-0.2 grams per kilogram of soil, phosphorus content is 0.05-0.1 grams per kilogram of soil, potassium content is 0.1-0.15 grams per kilogram of soil), average temperature between 15-20 degrees Celsius, annual precipitation between 500-600 mm, and specific soil type (such as loam). The server searches the entire agricultural area and determines the vegetation area that meets these conditions as the first template vegetation area, assuming here that it is a large area of ​​square bamboo planting area.

[0102] Another way is to output the vegetation areas updated by multiple reference ecological conditions as reference vegetation areas corresponding to each reference ecological condition, extract each reference vegetation area, and generate a first template vegetation area based on the extraction result. For example, there are three reference ecological conditions: the first is an area with high soil fertility (nitrogen content of 0.2-0.3 grams per kilogram of soil, phosphorus content of 0.1-0.15 grams per kilogram of soil, and potassium content of 0.15-0.2 grams per kilogram of soil) and low temperature (10-15 degrees Celsius), and the corresponding reference vegetation area is a square bamboo planting area; the second is an area with medium soil fertility (nitrogen content of 0.15-0.2 grams per kilogram of soil, phosphorus content of 0.08-0.12 grams per kilogram of soil, and potassium content of 0.12-0.18 grams per kilogram of soil) and moderate temperature (15-20 degrees Celsius), and the corresponding reference vegetation area is a soybean planting area; the third is an area with low soil fertility (nitrogen content of 0.1-0.15 grams per kilogram of soil, phosphorus content of 0.05-0.1 grams per kilogram of soil, potassium content of 0.1-0.2 grams per kilogram of soil) and moderate temperature (15-20 degrees Celsius). 0.15 grams) and a high temperature (20 - 25 degrees Celsius), the corresponding reference vegetation area is a sorghum planting area. The server extracts these three reference vegetation areas, which may involve the collection and collation of vegetation growth characteristics, soil-related data, meteorological data and other data, and then generates a first template vegetation area based on these extraction results. This first template vegetation area may be a virtual vegetation area that integrates different characteristics or a combination of some areas selected from these reference vegetation areas.

[0103] After the first template vegetation area is determined, the relevant vegetation reflectivity and vegetation coverage sequence data are obtained. First, the reference vegetation reflectivity sequence data and the reference vegetation coverage sequence data of the first template vegetation area within the reference radar detection cycle are obtained. Assuming that this reference radar detection cycle is a relatively long period of time, for example, from the first day to the 100th day after the square bamboo is planted, the duration unit of this cycle is greater than the cycle distance between the target cycle observation node and the initial node of the cycle observation. For the reference vegetation reflectivity sequence data, during this 100-day reference radar detection cycle, the radar equipment continuously detects the first template vegetation area (square bamboo planting area) to obtain a series of vegetation reflectivity values ​​arranged in chronological order. For example, the reflectivity on the first day is 0.15 and on the second day is 0.16. With the growth of square bamboo and changes in environmental factors (such as leaf growth, water absorption, etc.), the reflectivity value fluctuates continuously until a complete reference vegetation reflectivity sequence data is obtained on the 100th day. Similarly, the reference vegetation coverage sequence data was also acquired during this longer reference radar detection period. The coverage was calculated using image analysis technology combined with radar data, starting from an extremely low coverage (assumed to be 1%) on the first day when the square bamboo was planted, and gradually increasing with the growth of the square bamboo, reaching a higher coverage (assumed to be 80%) on the 100th day.

[0104] Then, according to the first wandering window, wandering collection is performed in the reference vegetation reflectivity sequence data to generate the third vegetation reflectivity sequence data. Assume that the wandering unit parameter of the first wandering window is 5 days, that is, data is collected once every 5 days. Starting from the first day of the reference radar detection cycle, a reflectivity data is collected on the first day, and then the next data is collected on the 6th day, and then the data is collected on the 11th day, and so on, until the data point that meets the continuous detection cycle requirements of the third vegetation reflectivity sequence data is collected. The continuous unit of the continuous detection cycle of the third vegetation reflectivity sequence data is the periodic distance between the target periodic observation node and the periodic observation initial node of the first radar detection cycle. Assuming that the target periodic observation node is the 50th day of the growth of square bamboo, and the periodic observation initial node of the first radar detection cycle is the 10th day, then this periodic distance is 40 days, so the collected third vegetation reflectivity sequence data is the reflectivity data corresponding to the 40 days collected according to the wandering window in the reference vegetation reflectivity sequence data.

[0105] For the generation of the third vegetation coverage sequence data, wandering collection is performed in the reference vegetation coverage sequence data according to the second wandering window. Assuming that the wandering unit parameter of the second wandering window is 3 days, starting from the 1st day of the reference radar detection cycle, the coverage data is collected at intervals of 3 days. For example, the coverage data is collected on the 1st day, then the next data is collected on the 4th day, and then the data is collected on the 7th day, and so on until the data points that meet the continuous detection cycle requirements of the third vegetation coverage sequence data are collected. Here, the continuous detection cycle of the third vegetation coverage sequence data and the continuous detection cycle of the fourth vegetation reflectivity sequence data are both continuous units of the continuous detection cycle of the first vegetation reflectivity sequence data. Assuming that the continuous detection cycle of the first vegetation reflectivity sequence data is 30 days, the collected third vegetation coverage sequence data and fourth vegetation reflectivity sequence data must meet the 30-day continuous unit requirement.

[0106] At the same time, obtain the fourth vegetation reflectivity sequence data of the second template vegetation area within the fourth radar detection cycle. The second template vegetation area and the first template vegetation area are vegetation areas with the same ecological conditions. It is assumed here that the second template vegetation area is also a square bamboo planting area. The fourth radar detection cycle and the third radar detection cycle are different detection cycles. For example, the fourth radar detection cycle is from the 20th day to the 70th day of the growth of the square bamboo. During this period, the radar equipment continuously detects the second template vegetation area to obtain the fourth vegetation reflectivity sequence data. For example, the reflectivity on the 20th day is 0.2, and on the 21st day it is 0.21, forming a complete sequence data in chronological order.

[0107] Finally, the first vegetation distribution detection network is optimized based on the third vegetation reflectance sequence data, the third vegetation coverage sequence data and the fourth vegetation reflectance sequence data. The first vegetation distribution detection network is a neural network model, which may include components such as multiple neuron layers, weight matrices and activation functions. During the optimization process, the third vegetation reflectance sequence data, the third vegetation coverage sequence data and the fourth vegetation reflectance sequence data are provided as input data to the first vegetation distribution detection network. The first vegetation distribution detection network performs forward propagation calculations based on these input data. The calculation process is that the data is linearly combined in each neuron layer according to the weights of the neurons, and then nonlinearly transformed through the activation function, and gradually transferred from the input layer to the output layer to obtain the prediction results. For example, for each data point in the third vegetation reflectance sequence data, when passing through the first layer of neurons, it is multiplied by the corresponding neuron weight and summed, and then a new output is obtained through the ReLU activation function, and this output is then passed to the next layer of neurons to continue similar operations.

[0108] At the same time, the first vegetation distribution detection network has a loss function, which is used to measure the difference between the network prediction result and the actual result. For example, the mean square error (MSE) can be used as the loss function to calculate the error between the predicted vegetation distribution and the actual vegetation distribution. After obtaining the prediction result and calculating the loss, the weight of the network is adjusted by the back propagation algorithm. The back propagation algorithm adjusts the weight layer by layer from the output layer in reverse according to the gradient of the loss function to the weight, so that the value of the loss function gradually decreases. For example, if the gradient of the loss function with respect to a certain weight is positive, then the weight is reduced; if the gradient is negative, the weight is increased. Through multiple iterations of this process, the weights are continuously adjusted, so that the difference between the prediction result and the actual result when the network processes these input data (the third vegetation reflectance sequence data, the third vegetation coverage sequence data, and the fourth vegetation reflectance sequence data) becomes smaller and smaller, thereby optimizing the first vegetation distribution detection network, enabling it to perform vegetation distribution detection more accurately, and improving the accuracy and reliability of vegetation distribution detection in the target vegetation area in practical applications.

[0109] In a possible implementation manner, the training step of the second vegetation distribution detection network includes:

[0110] Step B110, obtaining the third vegetation coverage sequence data and the fifth vegetation reflectivity sequence data of the first template vegetation area within the third radar detection cycle, obtaining the fourth vegetation coverage sequence data of the second template vegetation area within the fourth radar detection cycle, the second template vegetation area and the first template vegetation area are vegetation areas with the same ecological conditions, the third radar detection cycle and the fourth radar detection cycle are different detection cycles, the continuous detection period of the third vegetation coverage sequence data has a continuous unit of the periodic distance between the target periodic observation node and the periodic observation initial node of the first radar detection cycle, and the continuous detection period of the fifth vegetation reflectivity sequence data and the continuous detection period of the fourth vegetation coverage sequence data both have a continuous unit of the continuous detection period of the first vegetation reflectivity sequence data.

[0111] Step B120, optimizing the second vegetation distribution detection network according to the third vegetation coverage sequence data, the fifth vegetation reflectance sequence data and the fourth vegetation coverage sequence data.

[0112] In this embodiment, for the first template vegetation area, assuming that it is a square bamboo planting area, the server obtains the third vegetation coverage sequence data and the fifth vegetation reflectivity sequence data of the first template vegetation area within the third radar detection cycle. The third radar detection cycle is set according to a specific rule, and the continuous unit of its continuous detection cycle is the period distance between the target period observation node and the period observation initial node of the first radar detection cycle. For example, the target period observation node is the 60th day of the growth of the square bamboo, and the period observation initial node of the first radar detection cycle is the 20th day, then this period distance is 40 days. During the 40-day third radar detection cycle, the server obtains the third vegetation coverage sequence data through relevant equipment and technology. The server determines the vegetation coverage by combining radar data with image analysis technology. Starting from the 20th day, the vegetation coverage on the first day may be 30%. As the square bamboo grows, a vegetation coverage value will be obtained every day, such as 32% on the 21st day, 33% on the 22nd day, etc., so that the third vegetation coverage sequence data arranged in chronological order is formed.

[0113] At the same time, during this third radar detection cycle, the server also obtains the fifth vegetation reflectivity sequence data. The radar equipment detects the square bamboo planting area. The vegetation reflectivity detected on the 20th day is 0.25. After that, there will be a new reflectivity value every day, which may change due to the growth status of the square bamboo (such as leaf thickness, density, water content, etc.) and environmental factors (such as light, humidity, etc.). For example, the reflectivity on the 21st day is 0.26, and on the 22nd day it is 0.255, etc., thus forming the fifth vegetation reflectivity sequence data.

[0114] In addition, the server obtains the fourth vegetation coverage sequence data of the second template vegetation area within the fourth radar detection cycle. The second template vegetation area is also a square bamboo planting area and has the same ecological conditions as the first template vegetation area. The fourth radar detection cycle and the third radar detection cycle are different detection cycles, and the continuous unit of the continuous detection cycle of the fourth vegetation coverage sequence data is the continuous detection cycle of the first vegetation reflectivity sequence data. Assuming that the continuous detection cycle of the first vegetation reflectivity sequence data is 30 days, the fourth radar detection cycle is also set according to this 30-day unit. For example, the fourth radar detection cycle is from the 30th day to the 60th day of the growth of the square bamboo. During this period, the server obtains the vegetation coverage data every day. The vegetation coverage on the 30th day is 40%, and the vegetation coverage on the 31st day is 41%, etc., to form the fourth vegetation coverage sequence data.

[0115] Next, the second vegetation distribution detection network is optimized based on the third vegetation coverage sequence data, the fifth vegetation reflectance sequence data and the fourth vegetation coverage sequence data. The second vegetation distribution detection network is a model built based on a neural network, which contains multiple neuron layers, weight matrices, activation functions and other components.

[0116] The server provides the third vegetation coverage sequence data, the fifth vegetation reflectance sequence data, and the fourth vegetation coverage sequence data as input to the second vegetation distribution detection network. The network starts forward propagation calculations. When the data enters the first layer of neurons in the network, each neuron linearly combines the input data according to the pre-set weights. For example, for a data point in the third vegetation coverage sequence data, the corresponding data point in the fifth vegetation reflectance sequence data, and the related data point in the fourth vegetation coverage sequence data, the neurons will multiply these data by their respective corresponding weights, and then add the product results. After that, a nonlinear transformation is performed through an activation function (such as the ReLU function) to obtain a new output result. This output result will be passed to the next layer of neurons, and similar linear combination and nonlinear transformation operations will continue, and so on, layer by layer until it reaches the output layer of the network to obtain a prediction result.

[0117] At the same time, the second vegetation distribution detection network defines a loss function, such as a mean square error (MSE) function, which is used to measure the difference between the network prediction results and the actual results. The error between the predicted vegetation distribution related results and the actual vegetation distribution related results is calculated. According to this loss value, the weight of the network is adjusted by the back propagation algorithm. The back propagation algorithm adjusts the weights layer by layer from the output layer in reverse according to the gradient of the loss function to the weight. If the gradient of the loss function with respect to a certain weight is positive, then the value of the weight is reduced; if the gradient is negative, the value of the weight is increased. Through multiple iterations of this process, the weights are continuously adjusted, so that when the network processes these input data (the third vegetation coverage sequence data, the fifth vegetation reflectance sequence data, and the fourth vegetation coverage sequence data), the difference between the predicted results and the actual results is gradually reduced, thereby optimizing the second vegetation distribution detection network so that it can more accurately determine the vegetation growth trend parameters and other related results based on the input data in the actual vegetation distribution detection task.

[0118] Figure 2 The hardware structure of the vegetation distribution detection system 100 based on radar imaging for implementing the above-mentioned vegetation distribution detection method based on radar imaging provided by an embodiment of the present invention is shown as follows: Figure 2 As shown, the vegetation distribution detection system 100 based on radar imaging may include a processor 110 , a machine-readable storage medium 120 , a bus 130 , and a communication unit 140 .

[0119] The machine-readable storage medium 120 may store data and / or instructions. In some embodiments, the machine-readable storage medium 120 may store data acquired from an external terminal. In some embodiments, the machine-readable storage medium 120 may store data and / or instructions that the radar imaging-based vegetation distribution detection system 100 uses to execute or use to complete the exemplary method described in the present invention.

[0120] During the specific implementation process, one or more processors 110 execute computer executable instructions stored in the machine-readable storage medium 120, so that the processor 110 can execute the vegetation distribution detection method based on radar imaging in the above method embodiment. The processor 110, the machine-readable storage medium 120 and the communication unit 140 are connected through the bus 130, and the processor 110 can be used to control the sending and receiving actions of the communication unit 140.

[0121] The specific implementation process of the processor 110 can refer to the various method embodiments executed by the above-mentioned vegetation distribution detection system 100 based on radar imaging. The implementation principles and technical effects are similar and will not be repeated in this embodiment.

[0122] In addition, an embodiment of the present invention further provides a readable storage medium, in which computer executable instructions are preset. When a processor executes the computer executable instructions, the above vegetation distribution detection method based on radar imaging is implemented.

[0123] It should be noted that in order to simplify the description of the present invention and thus help understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present invention, various features are sometimes combined into one embodiment, drawing or description thereof.

Claims

1. A vegetation distribution detection method based on radar imaging, characterized in that: The method comprises: Acquire first vegetation reflectivity sequence data and first vegetation coverage sequence data of a target vegetation area within a first radar detection cycle, and acquire second vegetation reflectivity sequence data and second vegetation coverage sequence data of a control vegetation area within a second radar detection cycle, wherein the control vegetation area and the target vegetation area are vegetation areas with the same ecological conditions, and the first radar detection cycle and the second radar detection cycle are different detection cycles; The first vegetation reflectance sequence data and the second vegetation reflectance sequence data are fused and encoded to generate a first vegetation coding feature, and the first vegetation coverage sequence data and the second vegetation coverage sequence data are fused and encoded to generate a second vegetation coding feature; The first vegetation coding feature and the first vegetation coverage sequence data are fused and loaded into a first vegetation distribution detection network for vegetation distribution detection to generate a target vegetation reflectivity of the target vegetation area at a target period observation node; fusing the second vegetation coding feature and the first vegetation reflectance sequence data, loading them into a second vegetation distribution detection network for vegetation distribution detection, and determining the target vegetation growth trend parameter of the target vegetation area at the target period observation node according to the detection result of the second vegetation distribution detection network; Determining target vegetation distribution state parameters of the target vegetation area at the target period observation node according to the target vegetation reflectivity and the target vegetation growth trend parameter; The fusing of the first vegetation coding feature and the first vegetation coverage sequence data and loading them into a first vegetation distribution detection network for vegetation distribution detection to generate a target vegetation reflectivity of the target vegetation area at a target period observation node includes: Acquire a first ecological condition characteristic corresponding to the target vegetation area; Encoding the first ecological condition feature to generate a third vegetation coding feature; The third vegetation coding feature, the first vegetation coding feature and the first vegetation coverage sequence data are fused and loaded into a first vegetation distribution detection network for vegetation distribution detection to generate a target vegetation reflectivity of the target vegetation area at a target period observation node; The fusing of the second vegetation coding feature and the first vegetation reflectance sequence data and loading them into a second vegetation distribution detection network for vegetation distribution detection, and determining the target vegetation growth trend parameter of the target vegetation area at the target period observation node according to the detection result of the second vegetation distribution detection network, comprises: Acquire a second ecological condition characteristic corresponding to the target vegetation area; Encoding the second ecological condition feature to generate a fourth vegetation coding feature; Aggregating the fourth vegetation coding feature, the second vegetation coding feature and the first vegetation reflectance sequence data to generate input feature data; Loading the input feature data into a second vegetation distribution detection network, mapping the input feature data according to a mapping unit of the second vegetation distribution detection network, and generating a category mapping result, wherein the category mapping result represents a probability score of a plurality of reference vegetation growth matching models; Determine a target vegetation growth matching model from the multiple reference vegetation growth matching models according to the category mapping result, load the periodic observation initial node of the first radar detection period, the target periodic observation node, and each periodic observation node between the periodic observation initial node and the target periodic observation node into the target vegetation growth matching model, and generate multiple target vegetation growth parameter sequences; Based on the fusion results of the plurality of target vegetation growth parameter sequences, target vegetation growth trend parameters of the target vegetation area at the target period observation node are generated.

2. The vegetation distribution detection method based on radar imaging according to claim 1, characterized in that: The step of fusing the first vegetation reflectance sequence data and the second vegetation reflectance sequence data and encoding them to generate a first vegetation coding feature includes: splicing the first vegetation reflectance sequence data and the second vegetation reflectance sequence data to generate a first fusion space; Performing filtering operations on the first fusion space respectively through a diversified filter group to generate a plurality of first filtering result features of the same dimension; The members of the same spatial point in each of the first filtering result features are respectively integrated to obtain a plurality of first combined features, and the plurality of first combined features are encoded to generate a first vegetation coded feature.

3. The vegetation distribution detection method based on radar imaging according to claim 2 is characterized in that: The step of encoding the plurality of the first combined features to generate a first vegetation coded feature comprises: Loading the plurality of first combined features into a first recurrent neural network for time series feature mapping to generate first potential features corresponding to each of the first combined features; Focusing attention coefficients are allocated to each of the first potential features to generate focusing attention coefficients corresponding to each of the first potential features, and each of the first potential features is fused and calculated based on the focusing attention coefficients to generate a first vegetation coding feature.

4. The vegetation distribution detection method based on radar imaging according to claim 2 is characterized in that: The control vegetation areas include at least two, the second radar detection periods corresponding to the respective control vegetation areas are different detection periods, and the first vegetation reflectivity sequence data and the second vegetation reflectivity sequence data are spliced ​​to generate a first fusion space, including: splicing the second vegetation reflectivity sequence data corresponding to each of the reference vegetation areas according to the periodic order of each of the second radar detection periods to generate a candidate fusion space; Determining an embedding spatial point of the first vegetation reflectivity sequence data in the candidate fusion space according to the first radar detection cycle and the cycle order of each of the second radar detection cycles; The first vegetation reflectance sequence data is embedded into the candidate fusion space according to the embedding space point to generate a first fusion space.

5. The vegetation distribution detection method based on radar imaging according to claim 1, characterized in that: The training step of the first vegetation distribution detection network includes: Obtain third vegetation reflectivity sequence data and third vegetation coverage sequence data of the first template vegetation area within a third radar detection cycle, obtain fourth vegetation reflectivity sequence data of the second template vegetation area within a fourth radar detection cycle, the second template vegetation area and the first template vegetation area are vegetation areas with the same ecological conditions, the third radar detection cycle and the fourth radar detection cycle are different detection cycles, the continuous detection unit of the continuous detection cycle of the third vegetation reflectivity sequence data is the periodic distance between the target periodic observation node and the periodic observation initial node of the first radar detection cycle, and the continuous detection cycle of the third vegetation coverage sequence data and the continuous detection cycle of the fourth vegetation reflectivity sequence data are both continuous detection units of the first vegetation reflectivity sequence data; The first vegetation distribution detection network is optimized according to the third vegetation reflectance sequence data, the third vegetation coverage sequence data and the fourth vegetation reflectance sequence data.

6. The vegetation distribution detection method based on radar imaging according to claim 5 is characterized in that: The obtaining of the third vegetation reflectivity sequence data and the third vegetation coverage sequence data of the first template vegetation area within the third radar detection cycle includes: Acquire reference vegetation reflectivity sequence data and reference vegetation coverage sequence data of the first template vegetation area within a reference radar detection cycle, wherein the continuous detection unit of the reference vegetation reflectivity sequence data is greater than the cycle distance between the target periodic observation node and the periodic observation initial node, and the continuous detection unit of the reference vegetation coverage sequence data is greater than the continuous detection cycle of the first vegetation reflectivity sequence data; Performing wandering collection in the reference vegetation reflectance sequence data according to a first wandering window to generate the third vegetation reflectance sequence data, wherein a wandering unit parameter of the first wandering window is smaller than a sequence scale of the reference vegetation reflectance sequence data; The third vegetation coverage sequence data is generated by performing wandering collection in the reference vegetation coverage sequence data according to the second wandering window, and the wandering unit parameter of the second wandering window is smaller than the sequence scale of the reference vegetation coverage sequence data.

7. The vegetation distribution detection method based on radar imaging according to claim 1, characterized in that: The training step of the second vegetation distribution detection network includes: Obtain the third vegetation coverage sequence data and the fifth vegetation reflectivity sequence data of the first template vegetation area within the third radar detection cycle, and obtain the fourth vegetation coverage sequence data of the second template vegetation area within the fourth radar detection cycle, wherein the second template vegetation area and the first template vegetation area are vegetation areas with the same ecological conditions, the third radar detection cycle and the fourth radar detection cycle are different detection cycles, the continuous detection unit of the continuous detection cycle of the third vegetation coverage sequence data is the periodic distance between the target periodic observation node and the periodic observation initial node of the first radar detection cycle, and the continuous detection unit of the fifth vegetation reflectivity sequence data and the continuous detection cycle of the fourth vegetation coverage sequence data are both the continuous detection units of the first vegetation reflectivity sequence data; The second vegetation distribution detection network is optimized according to the third vegetation coverage sequence data, the fifth vegetation reflectance sequence data and the fourth vegetation coverage sequence data.

8. A vegetation distribution detection system based on radar imaging, characterized in that: The vegetation distribution detection system based on radar imaging includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the vegetation distribution detection method based on radar imaging as described in any one of claims 1 to 7.

Images