Field stubble coverage rate online detection system based on image recognition

Abstract

The invention discloses a field stubble coverage rate on-line detection system based on image recognition, and relates to the technical field of agricultural remote sensing detection, and the system can perform real-time observation on field stubble coverage conditions in arid and semi-arid cultivated land environments by synchronously obtaining visible light, near-infrared and surface temperature images with high spatial resolution, thereby improving the field stubble coverage rate on-line detection. The analysis units are grouped through an unsupervised clustering method, the units with similar features are classified into the same earth surface state type, the earth surface state type obtained through clustering is converted into discrete stubble coverage state grades, and then the stubble coverage rate of the target farmland is calculated in combination with set coverage proportion intervals corresponding to all the grades. According to the invention, through a processing structure of multi-source image one-time acquisition, analysis unit-level processing, unsupervised clustering identification and hierarchical coverage rate statistics, rapid identification and quantitative statistics of the coverage rate of the field stubbles under arid and semi-arid cultivated land conditions are realized.

Description

Technical Field

[0001] This invention relates to the field of agricultural remote sensing detection technology, specifically an online detection system for field stubble coverage based on image recognition. Background Technology

[0002] During the tillage process before crop sowing, the state of crop residue cover directly affects soil surface conditions, especially soil moisture retention. For arid and semi-arid farmlands in northern regions, where soil moisture is limited, rainfall is scarce, and evaporation is high, the level of crop residue cover directly affects soil moisture retention efficiency and surface water evaporation rate. Therefore, obtaining accurate crop residue cover information is crucial for assessing soil moisture status.

[0003] Existing stubble identification technologies mostly rely on a single spectral band or a fixed combination of spectral indices, such as judging stubble based on the linear characteristics of the shortwave infrared (SWIR) band. However, for arid and semi-arid farmland in northern China, these methods have limitations. The stubble moisture content, tillage time, surface exposure, atmospheric humidity, and light conditions vary greatly among different plots, resulting in highly variable spectral responses for stubble. At the same time, stubble is easily mixed with background materials such as bare soil, making it difficult for a single spectral index to reliably distinguish between different types of stubble.

[0004] Furthermore, this spectral-based, refined pixel-level coverage calculation relies on high-resolution, pixel-by-pixel processing of the entire farmland, requiring substantial computational resources and time. This computational load and processing complexity make it difficult for the method to operate in real-time under online observation conditions such as aerial photography or drones, failing to meet the need for rapid detection of stubble coverage and limiting its applicability in online detection and rapid decision-making applications. Summary of the Invention

[0005] (a) Technical problems to be solved This invention provides an online detection system for field stubble coverage based on image recognition, which can quickly observe and calculate stubble coverage in arid and semi-arid farmland.

[0006] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: an online detection system for field stubble coverage based on image recognition, applicable to arid and semi-arid farmland, including: The acquisition module is used to simultaneously acquire high spatial resolution visible light images, near-infrared remote sensing images and surface temperature images of the target farmland. After performing geometric correction and radiometric consistency processing on the acquired images, the target farmland is divided into several analysis units. The surface feature extraction module is used to calculate, within each analysis unit, spectral features reflecting surface moisture based on the visible light and near-infrared images, and structural features characterizing surface roughness and shading properties; calculate and analyze the mean and dispersion of surface temperature based on the surface temperature image to describe the thermal response characteristics of the surface; and form a state feature vector describing the overall surface state of the analysis unit after normalizing various features. The surface state clustering module is used to cluster the analysis units using the state feature vectors corresponding to each analysis unit as input and an unsupervised clustering method. During the clustering process, the analysis units with similar feature distributions are grouped into the same category by calculating the similarity between the analysis units in terms of spectral features, structural features, and thermal state features. Each category obtained by clustering represents a type of surface state caused by differences in stubble cover under the current observation conditions. The mapping module maps the different land surface conditions into multiple stubble cover status levels based on the differences in spectral response, land surface temperature, and structural characteristics of the different land surface condition types. The stubble coverage calculation module is used to calculate the stubble coverage rate of the target farmland by statistically analyzing the area ratio of each stubble coverage status level corresponding to the analysis unit in the target farmland, and combining the set stubble coverage ratio range corresponding to each stubble coverage status level.

[0007] In some feasible embodiments, the acquisition module acquires high spatial resolution multi-source images of the target farmland area under the same observation conditions. The multi-source images include visible light images, near-infrared remote sensing images, and land surface temperature images. Each type of image records corresponding spatial location information and covers the same target farmland spatial range. The acquisition module performs geometric correction processing on the visible light images, near-infrared remote sensing images, and land surface temperature images respectively, unifying the various images to the same spatial reference coordinate system. Based on the processed images, the acquisition module divides the target farmland area into several non-overlapping analysis units according to a preset spatial scale; each analysis unit corresponds to the same spatial range in the visible light image, near-infrared remote sensing image, and surface temperature image.

[0008] In some feasible embodiments, the surface feature extraction module takes the set of analysis units output by the acquisition module as input and performs feature extraction processing on the multi-source images within each analysis unit; the extraction steps are as follows: For the visible light and near-infrared images within each analysis unit, the surface feature extraction module reads the reflectance ratio of the red light band and near-infrared band of each pixel and uses it as the spectral response value of each pixel; it then summarizes all the spectral response values ​​within the analysis unit to generate spectral features that reflect the overall moisture-related response of the analysis unit. Based on the spatial distribution of pixels within the analysis unit, the surface feature extraction module performs spatial neighborhood analysis on the pixel grayscale values. First, it calculates the difference in grayscale values ​​between each pixel and its surrounding neighboring pixels to quantify the local grayscale gradient. Then, it statistically summarizes the local grayscale gradient and neighborhood consistency index within the entire analysis unit to form the structural features of the analysis unit, which are used to reflect the surface roughness and occlusion characteristics. The surface feature extraction module reads the pixel values ​​of the surface temperature image of the analysis unit, calculates the mean value of the temperature pixels within the analysis unit to reflect the overall heat level, and calculates the dispersion of the temperature pixels to quantify the differences in local temperature distribution. The calculated mean value and dispersion are combined to form the thermal response feature of the analysis unit.

[0009] In some feasible embodiments, the surface feature extraction module normalizes the spectral features, structural features, and thermal response features extracted in each analysis unit to unify the numerical scale and dimensions. Then, the three types of features of the analysis unit are combined in a fixed order to form a state feature vector. The surface feature extraction module generates a state feature vector for each analysis unit and maintains the correspondence with the spatial location of the analysis unit.

[0010] In some feasible embodiments, the surface state clustering module takes the state feature vector of each analysis unit output by the surface feature extraction module as input, and before performing clustering processing, organizes the state feature vectors of all analysis units in spatial order to form a feature matrix; wherein, each row of the feature matrix represents the state feature vector of an analysis unit, and each column of the matrix represents a feature dimension. For each pair of analysis units in the feature matrix, the similarity between the corresponding state feature vectors is calculated using the vector similarity method. The similarity is then used as the input for unsupervised clustering in the surface state clustering module to quantify the closeness of each analysis unit in the three types of features.

[0011] In some feasible embodiments, the surface state clustering module first delineates the area during the clustering process. There are 10 categories, each category is represented by a corresponding centroid vector. This indicates that during the clustering process, for each analysis unit... The corresponding state feature vector They were assigned to the category that minimized the squared distance. ; Wherein, the centroid vector During the iteration process, it is updated to the average of all analysis unit vectors within the current category: ; in, For the current category The analysis unit set is generated; after the iteration, the category label of the analysis unit reflects its clustering affiliation in the spectral feature, structural feature and thermal response feature space, and the surface state clustering module outputs the category of each analysis unit.

[0012] In some feasible embodiments, the mapping module receives the surface state type label output by the surface state clustering module, as well as the corresponding category feature statistics, wherein the category statistical features include the category center vector of each surface state type in terms of spectral response features, structural features, and thermal response features. In the mapping module, the category center vector corresponding to each land surface state type is first decomposed into spectral feature sub-vectors, structural feature sub-vectors, and thermal response feature sub-vectors. Then, the mapping module performs scalarization processing on the spectral feature sub-vectors, structural feature sub-vectors, and thermal response feature sub-vectors respectively. By statistically aggregating the feature components within each sub-vector after normalization, a comprehensive response value representing the land surface state type in the dimensions of spectral response, land surface structure, and thermal response is obtained, thereby converting the original multidimensional category feature representation into a three-dimensional response vector at a unified scale.

[0013] In some feasible embodiments, the mapping module constructs a comprehensive response index related to stubble for land surface type by performing a linear combination operation on the three-dimensional response vector; then the mapping module sorts the comprehensive response indices corresponding to all land surface type to form a relative order relationship of land surface type in stubble cover feature space, and the sorting result reflects the overall difference structure of different land surface type in stubble cover degree. Based on the sorting results, the mapping module divides the sorted set of land surface condition types into multiple consecutive intervals according to a preset number of levels, with each interval corresponding to a stubble cover condition level.

[0014] In some feasible embodiments, the coverage calculation module receives the stubble coverage status level of each analysis unit output by the mapping module, and obtains the spatial information of the target farmland, wherein each analysis unit corresponds to a unique spatial location identifier and a determined stubble coverage status level label. The coverage calculation module groups each analysis unit according to the stubble coverage status level, merges analysis units with the same stubble coverage status level label into the same coverage status subset, and establishes a corresponding spatial unit index for each coverage status subset.

[0015] (III) Beneficial Effects: Compared with the prior art, this invention has the following beneficial effects: This invention enables real-time observation of crop residue coverage in arid and semi-arid farmland environments by simultaneously acquiring high spatial resolution visible light, near-infrared, and surface temperature images. The visible light and near-infrared images provide information on surface reflectance characteristics to distinguish different types of crop residue from bare soil, while the surface temperature images supplement thermal response characteristics, reflecting surface moisture distribution and the crop residue coverage effect.

[0016] This system integrates spectral, structural, and thermal response features within each analytical spatial unit of farmland, providing a unified multidimensional feature foundation for rapid identification of crop residue cover. Subsequently, unsupervised clustering is used to group the analytical units, classifying units with similar features into the same land condition type, enabling the identification of crop residue cover patterns without prior training.

[0017] The mapping module converts the surface state types obtained from clustering into discrete stubble cover state levels, and combines the set stubble cover ratio ranges corresponding to each level to quickly calculate the stubble cover rate of the target plot, realizing a fast and unified calculation of coverage results from multi-source images. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of an online field stubble coverage detection system based on image recognition, provided in an embodiment of the present invention. Figure 2 A schematic diagram illustrating the connection process of various modules in an online field stubble coverage detection system based on image recognition provided in an embodiment of the present invention. Figure 3 This is a schematic diagram illustrating the process of the surface feature extraction module extracting corresponding features from multi-source images in an online field stubble coverage detection system based on image recognition provided in an embodiment of the present invention. Figure 4 This is a schematic diagram of a field stubble coverage online detection system based on image recognition provided in an embodiment of the present invention, showing a visible light image, a spectrally derived binary image, and a surface temperature image obtained by the same analysis unit. In the picture: 100. Data Acquisition Module; 200. Surface Feature Extraction Module; 300. Surface State Clustering Module; 400. Mapping Module; 500. Coverage Calculation Module. Detailed Implementation

[0019] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0020] It should be noted that, where there is no conflict, the features in the embodiments of the present invention can be combined with each other.

[0021] In the arid or semi-arid farming areas of northern China to which the embodiments of the invention are applicable, the stubble (mulch and straw) and weeds of the previous crop are buried during the tillage before sowing. The stubble mulch in the field has a direct impact on the soil moisture process, which is mainly reflected in reducing soil evaporation, enhancing rainfall infiltration, and improving the soil's water retention capacity.

[0022] Specifically, this can be understood as follows: Stubble cover forms a physical barrier on the soil surface, reducing the intensity of direct solar radiation on the soil surface and slowing down the evaporation of soil moisture into the atmosphere. Stubble cover weakens wind disturbance on the land surface, reducing the combined effects of wind erosion and evaporation. Furthermore, long-term stubble cover helps increase soil organic matter content, improves soil structure, reduces the probability of surface crusting after rainfall, and increases infiltration efficiency, thereby enhancing soil water retention capacity. These factors explain why, in arid, water-saving agricultural areas like northern regions, the level of stubble cover significantly impacts farmland water use efficiency.

[0023] From the perspective of dryland water conservation, the presence or absence of stubble is not sufficient to reflect its water regulation effect. In agricultural management, stubble coverage rate is a key quantitative indicator used to assess the actual intensity of the effect of stubble on the land surface.

[0024] Therefore, as Figures 1 to 4 The above-described image recognition-based online detection system for field stubble coverage is applicable to arid and semi-arid farmland in northern China. It uses the spatial distribution of surface moisture response patterns to invert the equivalent stubble coverage rate, and is used to quickly detect and identify field stubble coverage rate online using remote sensing images.

[0025] Before explaining the online detection system in detail, it is important to clarify that the system no longer uses whether a single pixel in the image is stubble as the judgment target. Instead, it divides the farmland into several analysis units, which can be spatial computing objects such as fixed grids or pixels. Each analysis unit describes a comprehensive surface condition.

[0026] This land surface condition assessment does not aim for detailed classification, but rather focuses on surface reflectance characteristics, response features related to moisture and temperature, and their relative spatial changes. Using single-temporal aerial remote sensing data, the system primarily leverages spatial differences in surface spectrum, structure, and thermal state to rapidly invert farmland stubble coverage, enabling rapid assessment for online monitoring.

[0027] In arid and semi-arid farmlands in the north, soil moisture is a key factor restricting crop emergence and early growth. As an important surface condition affecting surface water retention and evaporation, stubble cover, based on the above analysis, can be understood to mean that the water-retaining effect of stubble cover is not reflected in the specific type of stubble itself, but in the cover state it forms on the surface.

[0028] Under these conditions, relying on detailed identification of individual straw stalks and single stubble types would not only be computationally intensive but also highly sensitive to changes in light intensity and soil background, making it difficult to meet the needs of rapid on-site assessment. Therefore, in this embodiment of the invention, the online detection system focuses more on whether and to what extent stubble has formed effective cover under the current observation conditions, rather than on the morphological classification of the stubble. This is precisely why the acquisition module in this online detection system needs to simultaneously acquire visible light, near-infrared, and surface temperature information and process it within a unified spatial framework.

[0029] More specifically, the acquisition module obtains multi-source images of the target farmland simultaneously under the same observation conditions using various camera units mounted on an aerial photography device similar to a drone, acquiring visible light images, near-infrared remote sensing images, and surface temperature images. It is particularly important to note that subsequent processing modules of this online detection system use the multi-source remote sensing images acquired in a single aerial photograph by the acquisition module as input, without relying on time-series data or historical sample accumulation. This allows the system to quickly analyze farmland stubble coverage under current observation conditions, making it suitable for online or near-real-time operation scenarios.

[0030] Given the rapid changes in surface temperature and moisture in arid and semi-arid regions, temporal discrepancies between different image sources, even if spatially consistent, can lead to a loss of physical consistency among subsequent features, thus affecting the determination of stubble cover status. Therefore, synchronous acquisition is a fundamental prerequisite for the execution of the data acquisition module.

[0031] After acquiring various images, the acquisition module performs geometric correction on the visible light, near-infrared and surface temperature images respectively, and unifies them to the same spatial reference coordinate system. This is mainly to ensure that the same surface location has a one-to-one correspondence in different images.

[0032] During the data acquisition phase, each analysis unit already corresponds to: A spatially identical set of visible light pixels; A spatially identical set of near-infrared pixels; A set of spatially identical pixels representing the Earth's surface temperature.

[0033] The surface feature extraction module takes the set of analysis units output by the acquisition module as its direct input. In this module, each analysis unit is regarded as an independent processing object, and all operations within the module are limited to the pixel range corresponding to that analysis unit.

[0034] Specifically, for a given analysis unit, the surface feature extraction module first reads the red band reflectance of all pixels within that unit in the visible light image and the near-infrared band reflectance in the near-infrared image.

[0035] For each pixel, this module calculates the ratio of its near-infrared reflectance to its red reflectance, obtaining a scalar value as the pixel's spectral response value. After obtaining the spectral response values ​​of all pixels within the analysis unit, the module no longer retains the pixel-level results but instead performs statistical summarization on these response values.

[0036] Specifically, the module summarizes the spectral response values ​​of all pixels to obtain a statistical measure that represents the overall spectral response level of the analysis unit. Finally, the module outputs an overall spectral feature value for each analysis unit. It is important to note that this value no longer distinguishes between pixel locations, but is instead used as the spectral feature output for that analysis unit.

[0037] In the structural feature extraction stage, the surface feature extraction module still processes data within a single analysis unit. First, it reads the grayscale values ​​of all pixels within that unit. These grayscale values ​​originate from the grayscale information used for structural description in visible or near-infrared images. Then, it performs a neighborhood analysis operation on each pixel. Specifically, it determines the surrounding pixels within a preset neighborhood range centered on the current pixel, calculates the difference between the current pixel's grayscale value and the grayscale values ​​of each pixel in the neighborhood, and uses this difference as the local grayscale gradient index for that pixel.

[0038] In this structural feature extraction stage, the output is still pixel-level results, with each pixel corresponding to a local gray-level gradient value. After completing the calculation of the local gray-level gradients of all pixels within the entire analysis unit, the surface feature extraction module performs overall statistics on the local gray-level gradients of all pixels within the analysis unit, and quantifies the spatial distribution characteristics between pixels by combining statistical descriptions related to neighborhood consistency.

[0039] Finally, the module integrates these statistical results into a structural feature value (or a set of fixed-dimensional structural feature quantities), which serves as the structural feature output of the analysis unit.

[0040] In the thermal response feature extraction stage, for the same analysis unit, the surface feature extraction module first reads all temperature pixel values ​​corresponding to that unit in the surface temperature image. It's important to note that since the preceding acquisition module has already performed geometric correction and spatial unification processing, this module can directly read by spatial index without additional matching operations.

[0041] Specifically, the surface feature extraction module performs two types of statistical calculations on the read temperature pixels. One is to calculate the mean of all temperature pixels, which describes the overall thermal level of the analysis unit; the other is to calculate the dispersion of temperature pixels, which quantifies the differences in temperature distribution within the analysis unit. Finally, these two statistical measures are combined within the module into a thermal response feature, which serves as the description result of the analysis unit in the thermal state dimension.

[0042] After completing the calculation of the above three types of features, the surface feature extraction module generates the following for each analysis unit: A spectral feature (a statistical summary of spectral response values); A structural feature (derived from the combined results of gray-level gradient and neighborhood statistics); A thermal response characteristic (consisting of temperature mean and dispersion).

[0043] Each analysis unit generates only one state feature vector, which serves as the final output of the module and can be directly called by the subsequent surface state clustering module.

[0044] Combining the feature extraction stages described above, for example, in an aerial photograph of farmland, an analysis unit covers a 1m × 1m surface area. The module first calculates the near-infrared to red reflectance ratio for all pixels within this area and summarizes these ratios into a spectral feature. Then, it performs neighborhood analysis on the grayscale changes of the pixels within this area, summarizing the spatial differences between pixels into a structural feature. Finally, it statistically analyzes all temperature pixels within the area to generate corresponding thermal response features. Each analysis unit ultimately outputs only a fixed-dimensional state feature vector, which is composed of three types of features: Spectral characteristics (overall values ​​at the analysis unit level); Structural characteristics (analysis unit level global value); Thermal response characteristics (analysis unit level overall value).

[0045] Therefore, after performing the above feature extraction on each analysis unit of the farmland, a set of state feature vectors already exists in the online detection system: ; in, Indicates the first The state feature vector of each analysis unit The total number of analysis units within the target farmland.

[0046] Before performing any clustering operations, the surface state clustering module first performs a data organization operation, which arranges the state feature vectors of all analysis units into a two-dimensional matrix according to the unit number or spatial traversal order: ; In this matrix, each row corresponds to an analysis unit, and each column corresponds to a feature dimension. This matrix serves as the unified input data structure for all subsequent distance calculations and clustering operations.

[0047] When starting clustering, the surface condition clustering module requires pre-setting the number of categories. It is important to note here that... Instead of using a manually set similarity threshold for segmentation, it serves as a structural parameter for unsupervised clustering, constraining the number of groups in a cluster.

[0048] During the specific clustering calculation process, the surface state clustering module first initializes... One centroid vector: ; It is important to note that the dimension of each centroid vector is different from that of the state feature vector. The dimensions are completely consistent. These initial centroids can be obtained by initially selecting samples in the feature matrix or by initializing them with a uniform distribution; no specific limitations are made here.

[0049] In the process of clustering and assigning categories to the analysis units, the following criteria are used: ; in, This distance is derived from a distance metric in Euclidean space, and it reflects the distance of the analysis unit. In the multidimensional space composed of spectral, structural, and thermal response characteristics, and with category The degree of deviation between the central states.

[0050] Wherein, the centroid vector During the iteration process, it is updated to the average of all analysis unit vectors within the current category: ; in, For the current category The set of analytical units; the formula utilizes the fact that when the goal is to minimize the squared distance of all samples to the class center, the optimal center location is the arithmetic mean of the sample vectors.

[0051] Regarding the conditions for stopping the iteration, in some embodiments of the present invention, the surface state clustering module is repeatedly executed, assigning analysis unit categories based on the current centroids and updating the centroid vectors based on the category assignment results, until any of the following conditions are met: The category assignment results will no longer change; Or the change in the centroid vector is less than a preset threshold; Or it may reach the maximum number of iterations.

[0052] The iteration stops when any condition is met. After clustering, each analysis unit is assigned a category label, which indicates the cluster affiliation of the analysis unit in the state space composed of spectral features, structural features, and thermal response features. Regarding the output of this module, it is important to understand that the output of the surface state clustering module is not a physical quantity, but rather a set of discrete labels that classify the analysis units into several surface state types. These labels will serve as the input for subsequent mapping modules.

[0053] The reason why the above-mentioned unsupervised clustering process is performed on the surface state of the target farmland in the embodiments of the present invention is that the clustering operation automatically groups the analysis units under similar observation response conditions into several surface state types, so that the analysis units within the same category have high consistency in the multidimensional feature space, thereby transforming the originally continuous, complex and multi-factor-affected surface feature distribution into several relatively stable internal surface state subsets.

[0054] The surface state type output by clustering does not directly correspond to the stubble cover level, but rather serves as an intermediate expression for different surface observation states, used to isolate interference caused by differences in soil background, hydrothermal conditions, and surface structure.

[0055] After obtaining the surface condition types of each region, the online detection system further uses the category center vector of each type as the overall representative of that type of surface condition, decomposes and scalarizes its spectral response, structural features and thermal response features, compresses the original high-dimensional features into a comprehensive response representation under a unified scale, and constructs a comprehensive response index that is monotonically correlated with the degree of stubble cover.

[0056] Therefore, it can be understood that by globally ranking the comprehensive response indicators of different land surface conditions, the system establishes the relative order of each land surface condition type in the stubble cover feature space.

[0057] The mapping module receives input from the surface state clustering module. Its input data includes a surface state type label for each analysis unit and a category center vector statistically obtained for each surface state type. In some embodiments of the present invention, this category center vector is obtained by averaging the state feature vectors of all analysis units within the same category after clustering, and is used to characterize the overall properties of this type of surface state in three dimensions: spectral response, structural features, and thermal response features.

[0058] When the mapping module is executed, it is important to note that this module primarily processes surface state types, rather than individual analysis units.

[0059] For each land surface condition type, the mapping module reads its corresponding category center vector and, according to the feature composition order of the feature extraction stage, structurally splits the center vector into spectral feature sub-vectors, structural feature sub-vectors, and thermal response feature sub-vectors. This splitting process only involves rearranging and truncating the vector indices and does not introduce new calculations.

[0060] After completing the feature vector splitting, the mapping module performs scalarization processing on the three types of sub-vectors mentioned above.

[0061] Specifically, for each type of subvector, its internal feature components are first normalized to map feature components with different dimensions and numerical ranges to a uniform scale. Subsequently, statistical aggregation calculations are performed on the normalized feature components. For example, in some embodiments of the present invention, the mean or weighted mean is used to compress the original multi-dimensional subvector into a single scalar to characterize the comprehensive response level of the land surface state type in the corresponding feature dimension.

[0062] Through this process, each surface condition type is converted into a three-dimensional response vector consisting of three scalars, corresponding to the spectral response, structural response, and thermal response, respectively.

[0063] After obtaining the three-dimensional response vector, the mapping module performs a linear combination operation on it to calculate a comprehensive response index related to stubble cover. In some embodiments of the present invention, this linear combination process is achieved by assigning weights to the spectral response, structural response, and thermal response components, and then summing the weighted components to integrate the information of the three types of features into a single value. This comprehensive response index serves as the sole basis for subsequent ranking and classification, and its value reflects the relative position of different surface state types in the stubble-related feature space.

[0064] The mapping module then sorts the comprehensive response indices corresponding to all land surface condition types, forming ordered sequences from low to high or high to low. It's important to note that this sorting result is not directly output; rather, it serves as an intermediate structure for classifying land surface conditions, describing the overall differences in stubble-related responses among different land surface condition types.

[0065] After sorting, the mapping module divides the sorted set of land surface status types into several continuous intervals according to the preset number of stubble cover status levels.

[0066] During the classification process, based on the sorting sequence, land surface condition types are sequentially assigned to each level interval, ensuring continuity of land surface condition types within the same level in terms of comprehensive response indicators. Each interval corresponds to a stubble cover status level, and the mapping module uniformly assigns the corresponding cover status level label to all land surface condition types falling into that interval.

[0067] In addition to the above, it is necessary to further explain that the final mapping module outputs the stubble cover status level, which can be understood as the surface status type label of the analysis unit output in the above clustering stage being converted into a stubble cover status level label by the mapping module according to the set coverage ratio range.

[0068] For example, suppose the coverage status levels are divided into several levels: L1, L2, L3...Lm, and the coverage ratio range corresponding to each coverage status level is set to: 0-10%, 11%-20%...91%-100%. That is to say, each analysis unit has a certain degree of residual coverage, but this coverage status level is only used to express the approximate coverage of this unit.

[0069] Based on the above description of cluster analysis, clustering is not for determining whether there are any fragments, but for automatically merging analytical units with similar coverage in a multi-source feature space. That is, the output of clustering is that these analytical units exhibit consistent coverage characteristics in terms of structure, spectrum, and thermal response. The mapping module's role in this process is simply to map this consistency result to a coverage ratio range (level).

[0070] The coverage calculation module in this online detection system uses the output of the aforementioned mapping module as input, and its input data consists of two parts: First, the label for the stubble cover status level (coverage ratio range) corresponding to each analysis unit; Second, analyze the spatial information of the unit in the target farmland.

[0071] Each analysis unit has a unique spatial location identifier and corresponds to a determined stubble cover status level. This spatial information is used to determine the area attribute of the analysis unit and its spatial affiliation within the plot in subsequent calculations.

[0072] The coverage calculation module first iterates through all analysis units, grouping those with the same stubble cover level label into the same coverage level subset. A corresponding spatial index structure is then created for each subset to record the spatial location identifiers of all analysis units within that subset. Continuing with the example of coverage level and coverage ratio ranges, all analysis units classified as coverage level L1 are grouped into one set; all analysis units classified as coverage level L2 are grouped into another set. This process continues, grouping all analysis units within the target farmland area based on their stubble cover ratio.

[0073] After grouping, the coverage calculation module performs area accumulation calculations for each coverage subset. Since the target farmland has been divided into non-overlapping analysis units according to a uniform spatial scale in the acquisition module, the actual surface area corresponding to each analysis unit is a known constant. In addition, it can also be directly calculated by acquiring the spatial resolution of the image.

[0074] Setting the first The surface area of ​​each analysis unit is For the first Each residual cover status level, and the corresponding set of analysis units is denoted as . The cumulative coverage area of ​​this coverage status level in the target farmland is then determined. Calculate as follows: ; The calculation process is implemented in the module by accumulating the area of ​​each analysis unit within the coverage state subset.

[0075] After obtaining the cumulative coverage area for each stubble cover status level, the coverage rate calculation module uses the total analyzed area of ​​the target farmland plot as the normalization benchmark to perform proportional calculations on the area results for each cover status level. For the first... Each stubble cover status level corresponds to a certain coverage rate. The calculation is as follows: ; in, The area of ​​the target farmland, This represents the coverage percentage corresponding to that level. For example, level L2 mentioned above corresponds to 11%-20%. The midpoint of the interval can be taken as 15.5%.

[0076] After calculating the stubble coverage rate, the coverage rate calculation module summarizes and organizes the coverage rate results corresponding to all stubble coverage status levels to form the stubble coverage rate statistics for the target farmland. Simultaneously, the module combines the spatial index information established during the grouping phase to write back the stubble coverage status level labels corresponding to each analysis unit to their spatial location identifiers, thereby synchronously generating spatial distribution statistics for each stubble coverage status level within the target farmland.

[0077] Finally, the coverage calculation module of the online detection system outputs two types of results: The first is the numerical result of stubble coverage rate in units of coverage status level; Second, it analyzes the spatial distribution information of the cover status at the unit level to characterize the spatial distribution pattern of each stubble cover status within the farmland.

[0078] It is important to understand that this online detection system first discretizes the degree of crop residue coverage. This process does not pursue pixel-level precision and does not directly invert the crop residue coverage rate at the pixel scale. Instead, it first performs hierarchical quantification of the coverage degree of farmland analysis units through multi-source feature clustering, and then realizes rapid statistical calculation of farmland crop residue coverage rate based on the area proportion of each coverage level analysis unit.

[0079] In summary, the entire online detection system provides information on the spectral response of surface materials under arid and semi-arid farmland conditions using visible light and near-infrared images, which helps to reflect the differences in reflectance characteristics of stubble, soil, and their mixtures. Surface temperature images supplement the description of surface conditions from the perspective of thermal response, and can reflect the differences in surface heating and heat dissipation behavior under different cover conditions in arid environments.

[0080] After extracting surface features, the system groups the analysis units using unsupervised clustering, automatically merging surface states with similar spectral responses, structural features, and thermal response characteristics. The surface state types output by the clustering are essentially a comprehensive summary of the multi-source response characteristics of farmland surfaces, ultimately yielding spatially consistent and statistically significant stubble cover results. This helps reveal the distribution pattern of surface states caused by differences in stubble cover under current observation conditions.

[0081] This processing procedure ensures that the calculation of stubble coverage retains spatial distribution information and reflects the proportion of different coverage levels in the overall farmland.

[0082] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. The scope of protection of the present invention is defined by the claims. Similarly, any equivalent structural changes made based on the description and drawings of the present invention should also be included within the scope of protection of the present invention.

Claims

1. An online detection system for field stubble coverage based on image recognition, characterized in that, Suitable for arid and semi-arid types of farmland, including: The acquisition module is used to simultaneously acquire high spatial resolution visible light images, near-infrared remote sensing images and surface temperature images of the target farmland. After performing geometric correction and radiometric consistency processing on the acquired images, the target farmland is divided into several analysis units. The surface feature extraction module is used to calculate, within each analysis unit, spectral features reflecting surface moisture based on the visible light and near-infrared images, and structural features characterizing surface roughness and shading properties; calculate and analyze the mean and dispersion of surface temperature based on the surface temperature image to describe the thermal response characteristics of the surface; and form a state feature vector describing the overall surface state of the analysis unit after normalizing various features. The surface state clustering module is used to cluster the analysis units using the state feature vectors corresponding to each analysis unit as input and an unsupervised clustering method. During the clustering process, the analysis units with similar feature distributions are grouped into the same category by calculating the similarity between the analysis units in terms of spectral features, structural features, and thermal state features. Each category obtained by clustering represents a type of surface state caused by differences in stubble cover under the current observation conditions. The mapping module maps the different land surface conditions into multiple stubble cover status levels based on the differences in spectral response, land surface temperature, and structural characteristics of the different land surface condition types. The stubble coverage calculation module is used to calculate the stubble coverage rate of the target farmland by statistically analyzing the area ratio of each stubble coverage status level corresponding to the analysis unit in the target farmland, and combining the set stubble coverage ratio range corresponding to each stubble coverage status level.

2. The online detection system for field stubble coverage based on image recognition according to claim 1, characterized in that, The acquisition module acquires high spatial resolution multi-source images of the target farmland area under the same observation conditions. The multi-source images include visible light images, near-infrared remote sensing images, and land surface temperature images. Each type of image records the corresponding spatial location information and covers the same target farmland spatial range. The acquisition module performs geometric correction processing on the visible light images, near-infrared remote sensing images, and land surface temperature images respectively, unifying the various images to the same spatial reference coordinate system. Based on the processed images, the acquisition module divides the target farmland area into several non-overlapping analysis units according to a preset spatial scale; each analysis unit corresponds to the same spatial range in the visible light image, near-infrared remote sensing image, and surface temperature image.

3. The online field stubble coverage detection system based on image recognition according to claim 1, characterized in that, The surface feature extraction module takes the set of analysis units output by the acquisition module as input and performs feature extraction processing on the multi-source images within each analysis unit; the extraction steps are as follows: For the visible light and near-infrared images within each analysis unit, the surface feature extraction module reads the reflectance ratio of the red light band and near-infrared band of each pixel and uses it as the spectral response value of each pixel; it then summarizes all the spectral response values ​​within the analysis unit to generate spectral features that reflect the overall moisture-related response of the analysis unit. Based on the spatial distribution of pixels within the analysis unit, the surface feature extraction module performs spatial neighborhood analysis on the pixel grayscale values. First, it calculates the difference in grayscale values ​​between each pixel and its surrounding neighboring pixels to quantify the local grayscale gradient. Then, it statistically summarizes the local grayscale gradient and neighborhood consistency index within the entire analysis unit to form the structural features of the analysis unit, which are used to reflect the surface roughness and occlusion characteristics. The surface feature extraction module reads the pixel values ​​of the surface temperature image of the analysis unit, calculates the mean value of the temperature pixels within the analysis unit to reflect the overall heat level, and calculates the dispersion of the temperature pixels to quantify the differences in local temperature distribution. The calculated mean value and dispersion are combined to form the thermal response feature of the analysis unit.

4. The online detection system for field stubble coverage based on image recognition according to claim 3, characterized in that, The surface feature extraction module normalizes the spectral features, structural features, and thermal response features extracted in each analysis unit to unify the numerical scale and dimensions. Then, it combines the three types of features of the analysis unit in a fixed order to form a state feature vector. The surface feature extraction module generates a state feature vector for each analysis unit and maintains the correspondence with the spatial location of the analysis unit.

5. The online detection system for field stubble coverage based on image recognition according to claim 1, characterized in that, The surface state clustering module takes the state feature vector of each analysis unit output by the surface feature extraction module as input. Before performing clustering, it organizes the state feature vectors of all analysis units in spatial order to form a feature matrix. Each row of the feature matrix represents the state feature vector of an analysis unit, and each column of the matrix represents a feature dimension. For each pair of analysis units in the feature matrix, the similarity between the corresponding state feature vectors is calculated using the vector similarity method. The similarity is then used as the input for unsupervised clustering in the surface state clustering module to quantify the closeness of each analysis unit in the three types of features.

6. The online field stubble coverage detection system based on image recognition according to claim 5, characterized in that, During the clustering process, the surface condition clustering module first delineates... There are 10 categories, each category is represented by a corresponding centroid vector. This indicates that during the clustering process, for each analysis unit... The corresponding state feature vector They were assigned to the category that minimized the squared distance. ； Wherein, the centroid vector During the iteration process, it is updated to the average of all analysis unit vectors within the current category: ； in, For the current category The analysis unit set is generated; after the iteration, the category label of the analysis unit reflects its clustering affiliation in the spectral feature, structural feature and thermal response feature space, and the surface state clustering module outputs the category of each analysis unit.

7. The online detection system for field stubble coverage based on image recognition according to claim 1, characterized in that, The mapping module receives the surface state type label output by the surface state clustering module, as well as the corresponding category feature statistics. The category statistical features include the category center vector of each surface state type in terms of spectral response features, structural features, and thermal response features. In the mapping module, the category center vector corresponding to each land surface state type is first decomposed into spectral feature sub-vectors, structural feature sub-vectors, and thermal response feature sub-vectors. Then, the mapping module performs scalarization processing on the spectral feature sub-vectors, structural feature sub-vectors, and thermal response feature sub-vectors respectively. By statistically aggregating the feature components within each sub-vector after normalization, a comprehensive response value representing the land surface state type in the dimensions of spectral response, land surface structure, and thermal response is obtained, thereby converting the original multidimensional category feature representation into a three-dimensional response vector at a unified scale.

8. The online detection system for field stubble coverage based on image recognition according to claim 7, characterized in that, The mapping module constructs a comprehensive response index related to stubble cover for land surface type by performing a linear combination operation on the three-dimensional response vector. Then, the mapping module sorts the comprehensive response indices corresponding to all land surface type to form a relative order relationship of land surface type in stubble cover feature space. The sorting result reflects the overall difference structure of different land surface type in stubble cover degree. Based on the sorting results, the mapping module divides the sorted set of land surface condition types into multiple consecutive intervals according to a preset number of levels, with each interval corresponding to a stubble cover condition level.

9. The online detection system for field stubble coverage based on image recognition according to claim 1, characterized in that, The coverage calculation module receives the stubble coverage status level of each analysis unit output by the mapping module, and obtains the spatial information of the target farmland, wherein each analysis unit corresponds to a unique spatial location identifier and a determined stubble coverage status level label. The coverage calculation module groups each analysis unit according to the stubble coverage status level, merges analysis units with the same stubble coverage status level label into the same coverage status subset, and establishes a corresponding spatial unit index for each coverage status subset.

Images