Medlar planting distribution high-precision remote sensing monitoring method based on multi-temporal fusion and spectral feature optimization

Abstract

The invention belongs to the technical field of remote sensing, and discloses a wolfberry planting distribution high-precision remote sensing monitoring method based on multi-temporal fusion and spectral feature optimization. According to the method, a multi-temporal and multi-spectral satellite image is used as a data source, and preprocessing and multi-temporal image fusion are firstly carried out; the method comprises the following core steps: establishing a time sequence characteristic curve according to a unique phenological period (such as bare soil characteristics in a dormancy period and high vegetation coverage in a rapid growth period) of wolfberry; in a spectral domain, screening out a characteristic spectrum dimension combination with the highest discrimination degree between the wolfberry and other crops through a characteristic wave band optimization algorithm (such as vegetation index difference degree and red edge characteristics); and in combination with an object-oriented classification or deep learning classification model, constructing a space-time coupling classifier, and performing high-precision extraction and distribution mapping on the Chinese wolfberry planting region. The method can effectively solve the problem of confusion classification of Chinese wolfberry and similar ground features (such as other shrubs and orchards), and realizes rapid and accurate monitoring of Chinese wolfberry planting area and spatial distribution.

Description

Technical Field

[0001] This invention belongs to, but is not limited to, the field of remote sensing technology, and particularly relates to a high-precision remote sensing monitoring method for wolfberry planting distribution based on multi-temporal fusion and spectral feature optimization. Background Technology

[0002] As an important characteristic economic crop in Ningxia, the planting scale and spatial distribution of wolfberry (Lycium barbarum L.) are important basic data for wolfberry industry planning, yield assessment and ecological protection. For a long time, the planting area of ​​wolfberry has been mainly formed by manual survey and reporting. This traditional method has exposed many drawbacks in modern wolfberry industry management and sustainable development: (1) poor timeliness: manual surveys have long cycles and slow data updates, and cannot reflect the dynamic changes in planting scale (such as new planting, fallowing and digging) in a timely manner; (2) low spatial accuracy: relying on county-level reporting, the reporting of planned planting by enterprises and cooperatives is relatively accurate, but for small plots planted by farmers and scattered planting, it is difficult to accurately define the boundaries and actual area of ​​the plots, and it is easy to have "fuzzy areas" and "misreporting" and "omission" phenomena; (3) high labor cost: a lot of manpower, material resources and time are required for on-site surveys, measurements and data collection, which is inefficient. In summary, the current method of manually surveying the planting area of ​​goji berries results in inaccurate data. This directly leads to a lack of solid data support when local governments formulate key policies such as goji berry industry development plans, subsidies, and irrigation water allocation, causing resource misallocation. Furthermore, accurate geographical distribution is a prerequisite for effective pest and disease monitoring and early warning. Without knowing the exact location and spread of diseases, it is impossible to promptly identify high-risk areas and implement precise and efficient control measures, easily leading to large-scale outbreaks of pests and diseases and causing significant economic losses. Therefore, traditional manual surveys of goji berry planting areas can no longer meet the requirements of high-quality development of the goji berry industry for data accuracy, timeliness, and objectivity. Therefore, it is urgently necessary to introduce modern information technologies such as remote sensing (e.g., satellite remote sensing, UAV remote sensing) and Geographic Information Systems (GIS) to establish a rapid, accurate, efficient, and dynamic monitoring and evaluation system for the scale and spatial distribution of goji berry planting, providing a reliable data foundation for scientific decision-making and sustainable development of the industry. Summary of the Invention

[0003] The purpose of this invention is to provide a high-precision remote sensing monitoring method for wolfberry planting distribution based on multi-temporal fusion and spectral feature optimization. By introducing time series analysis, feature selection and machine learning classification technology, the method can realize automatic identification and dynamic updating of wolfberry planting areas, effectively solving the problem of confused classification of wolfberry with similar land features (such as other shrubs and orchards).

[0004] This invention is implemented as follows: a high-precision remote sensing monitoring method for the distribution of wolfberry planting based on multi-temporal fusion and spectral feature optimization, the method comprising:

[0005] S1: Obtain data on the number of wolfberry planting plots in the study area, including large-scale planting areas such as enterprises and cooperatives, as well as scattered planting areas by farmers;

[0006] S2: Determine the time range of the growth period of Lycium barbarum in the study area; since Lycium barbarum is a continuously flowering and fruiting plant with indeterminate inflorescences, its growth period is relatively long, and the flowering and fruiting period lasts throughout the entire growing season; among them, the budding / leaf unfolding period is from the end of March to mid-April; the flowering / peak fruiting period is from early May to mid-to-late September; the harvesting period is from early June to early October; the autumn fruit ripening / late growth period is from mid-to-late September to October; and the leaf fall / dormancy period is from November to March of the following year.

[0007] S3: Based on the growth period of wolfberry, obtain remote sensing images covering the key phenological stages of wolfberry in the study area;

[0008] S4: Analysis of remote sensing index characteristics of wolfberry throughout its entire growth period;

[0009] S5: Acquisition and preprocessing of multi-source remote sensing data;

[0010] S6: Multidimensional feature extraction and initial feature space construction;

[0011] S7: Enhanced and optimized spectral and temporal characteristics;

[0012] S8: Construction of a spatiotemporal coupled classification model;

[0013] S9: Results output and accuracy assessment.

[0014] Furthermore, S5 specifically includes:

[0015] (1) Obtain S1 SAR images, S2 and L8 optical images of wolfberry during the key growth period from March to November in the Ningxia study area;

[0016] (2) Perform orbit correction, radiometric calibration, terrain correction, and Speckle filtering on the acquired S1 data, and extract the VV and VH polarization backscattering coefficients;

[0017] (3) The S2 and L8 data obtained are surface reflectance product data that have been atmospherically corrected and radiometrically calibrated; different bands of the S2 image are resampled into image data with uniform resolution.

[0018] Furthermore, S6 specifically includes:

[0019] (1) Extracting optical features based on S2 and L8: Extracting the original spectral bands of all time phases, including red, green, blue, near-infrared, and short-wave infrared, as the original spectral band features F. B (t i ):

[0020]

[0021] Where m1 and m2 are the number of available bands for S2 and L8, respectively;

[0022] Extracting vegetation index features F VI (t i ):

[0023]

[0024] Where V is the number of selected vegetation indices; all indices are calculated based on S2 and L8 data;

[0025] Single-phase optical characteristics:

[0026] F Opt (t i ) = F B (t i )∪F VI (t i )

[0027] Multi-temporal optical characteristics:

[0028]

[0029] (2) Extract radar features based on S1: Extract the VV and VH backscattering coefficients of all time phases and calculate the polarization ratio, such as VV / VH;

[0030] Single-phase radar characteristics:

[0031]

[0032] Where VV and VH are in phase t i The backscattering coefficient; Ratio is the polarization ratio, such as VV / VH;

[0033] General characteristics of multi-temporal radar:

[0034]

[0035] (3) Extraction of temporal phenological features: Based on key vegetation indices, such as NDVI, extract the phenological features F during the growing season of wolfberry (March-November). Phen Temporal features are extracted based on the vegetation index (e.g., NDVI) change curves throughout the growing season (cover T), and temporal phenological features F.Phen :

[0036] F Phen ={NDVI max NDVI min NDVI avg NDVI std ,DOY peak VI Integral}

[0037] Among them, NDVI max NDVI min NDVI ave NDVI std These represent the maximum, minimum, average, and standard values ​​of NDVI during the growing season; DOY peak It is the number of days that the NDVI peak occurs; VI Integral It is the area under the vegetation index curve during the growing season, expressed as an integral value.

[0038] (4) Feature fusion: All features extracted from S1, S2, and L8 at different times and dimensions, including spectral, radar, index, and time series data, are stacked to construct a high-dimensional initial feature space F. initial ;

[0039]

[0040] Furthermore, S7 specifically includes:

[0041] The study focuses on introducing red edge features and vegetation index differences to enhance the separability of wolfberry with similar land features, shrubs, and orchards; for red edge locations, linear interpolation was used, which is suitable for S2 data (using B4, B5, B6, and B7 bands).

[0042]

[0043] Where ρNIR and ρRed are the reflectances in the near-infrared and red light bands, respectively; λ i and λ j It is the center wavelength of the rising red-edge band; ρ j and ρ i It is the reflectivity of the corresponding waveband;

[0044] Vegetation Index Difference (VID) was used to measure the difference in vegetation index during the dormancy period of wolfberry (T). dormant ) and rapid growth period (T peak The NDVI of ) is used to calculate its degree of difference, which is used to capture phenological differences;

[0045] VID NDVI =NDVI(Tpeak -t dormant )

[0046] To determine the most discriminative feature subset F final JM distance is used as the criterion for feature selection;

[0047]

[0048] The JM distance ranges from [0, 2]. The closer the JM value is to 2, the better the separability of the two types of land cover, such as wolfberry and shrubs, on the current feature dimension. Calculate F... final The average JM distance between all features for "goji berries" and "highly confusing features" is used; the top k features with the largest JM distances and their corresponding time phases are selected to form the optimal spatiotemporal feature subset F. final .

[0049] Furthermore, S8 specifically includes:

[0050] The spatiotemporally coupled ST-U-Net model integrates temporal and spatial context information into a semantic segmentation framework. ST-U-Net inherits the encoder-decoder structure of the standard U-Net, but its convolutional layers are designed to handle time-series data. The input layer stacks all Ffinal feature channels of Nt image phases within the optimal temporal combination TOptimal into a spatiotemporal data cube Xinput.

[0051] Xinput∈RH×W×(Nt×Nc_final)

[0052] Where H×W is the spatial size of the image block, and Nc_final is the final number of feature channels.

[0053] The encoder employs 3D convolutional layers to extract features simultaneously in both spatial (X, Y) and temporal (T) dimensions, and automatically learns spatiotemporally coupled features.

[0054] The decoder fuses high-level semantic features captured by the encoder with low-level detail features through upsampling and skip connections, achieving accurate boundary recovery and pixel-level classification.

[0055] The model employs the Adam optimizer with the cross-entropy loss function. CE is trained:

[0056] CE=-1Ni=1Nc=1CYi,clog(Yi,c)

[0057] Where Yi,c is the true label and Yi,c is the model's predicted probability.

[0058] Furthermore, S9 specifically includes:

[0059] Based on the above model, the wolfberry planting areas were extracted, and a wolfberry planting distribution layer was output. The accuracy of the classification results was evaluated using an independent validation sample set. The wolfberry planting area of ​​the entire Ningxia region was extracted, and a wolfberry planting distribution map was created.

[0060] Based on the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solution to be protected by this invention are as follows:

[0061] This invention effectively solves the problem of confusion between goji berries and similar land cover features (such as other shrubs and orchards), enabling rapid and accurate monitoring of goji berry planting area and spatial distribution. By introducing time series analysis, feature selection, and machine learning classification techniques, it achieves automatic identification and dynamic updating of goji berry planting areas, effectively resolving the problem of confusion between goji berries and similar land cover features (such as other shrubs and orchards).

[0062] (1) The expected benefits and commercial value of the technical solution of this invention after transformation are as follows:

[0063] The high-precision remote sensing monitoring method based on multi-temporal fusion and spectral feature optimization provided by this invention has significant economic benefits and commercial value. Its core lies in providing scientific, real-time, and accurate data on the spatial distribution and area of ​​goji berry cultivation, significantly improving the management efficiency and reducing costs of the entire goji berry industry chain. At the macro level, it provides reliable data for macro-control, planting planning, yield forecasting, and disaster assessment by goji berry industry authorities, completely replacing traditional time-consuming and labor-intensive manual surveys and low-precision statistics, greatly saving manpower and time costs. At the micro level, the monitoring results can directly serve planting enterprises and cooperatives, guiding them in precise fertilization, irrigation, and harvesting, achieving optimal resource allocation, thereby effectively improving goji berry yield and quality, and enhancing market competitiveness. Furthermore, the high-precision, dynamically updated data can serve as a scientific basis for agricultural insurance companies to determine rates and assess disaster losses, as well as for financial institutions to conduct credit risk assessments, powerfully promoting the innovation and implementation of agricultural financial instruments. Ultimately, this method, as a core technology, can be developed into specialized remote sensing monitoring system software or data service platforms, and promoted to other major production areas and similar economic crop fields, generating continuous technical service revenue and demonstrating broad commercial application prospects.

[0064] (2) The technical solution of this invention fills a technical gap in the industry both domestically and internationally:

[0065] This invention successfully solves the long-standing technical challenge of accurately distinguishing wolfberry from similar ground features in remote sensing monitoring, filling a gap in high-precision remote sensing mapping for specific economic crops. Existing remote sensing monitoring technologies, when dealing with small- to medium-scale shrubs like wolfberry, easily produce "same species, different spectra" and "different species, same spectra" phenomena due to overlap in spectral characteristics and phenological periods with similar shrubs / orchards such as jujube and sea buckthorn, making it difficult to improve classification accuracy, especially for small plots and scattered wolfberry distributions. The innovation of this invention lies in proposing and implementing an integrated strategy of "multi-temporal fusion + spectral feature optimization." This strategy enables high-precision, rapid, and automated remote sensing monitoring of wolfberry planting distribution against complex ground feature backgrounds. Furthermore, by introducing machine learning, this method efficiently analyzes the multi-dimensional feature space, effectively solving the problems of strong regional dependence and poor cross-regional adaptability of traditional single vegetation index methods or single-temporal methods, making the monitoring results more universal and robust.

[0066] (3) The technical solution of the present invention solves a technical problem that people have long wanted to solve but have never been able to solve successfully:

[0067] This invention successfully solves the long-standing problem of "confusion between goji berries and similar features" that has hindered the modernization of goji berry industry management. This is a major breakthrough that traditional methods have long failed to achieve. Because the shrub morphology and phenological period of goji berries overlap with surrounding arbor orchards (such as jujube trees) and similar shrubs, their spectral curves are difficult to clearly separate in a single temporal phase, making it impossible for traditional methods to achieve high-precision (e.g., classification accuracy exceeding 90%) automated identification. This invention, through the systematic fusion of multi-temporal data, successfully captures the unique spectral-temporal signals of goji berries relative to similar features during key phenological periods (such as flowering and fruiting, and leaf fall), overcoming the limitations of depth and breadth in remote sensing data feature extraction. Simultaneously, combined with feature optimization technology, it automatically selects the subset of features that contribute most to goji berry identification from a vast feature set, greatly improving the discriminative power of the feature space, thus successfully solving this long-standing problem of classification accuracy. Ultimately, this invention integrates complex steps such as time series analysis, feature optimization, and machine learning classification into a standardized technical process, enabling rapid, accurate, and automated extraction of wolfberry planting information. This meets the urgent needs of the wolfberry industry authorities for real-time and efficient monitoring services, and realizes the "automation" and "intelligence" of the monitoring method.

[0068] (4) The technical solution of the present invention overcomes technical bias:

[0069] This invention successfully overcomes long-standing technical biases in the industry through innovative methods. First, it overcomes the reliance on traditional single vegetation index methods (such as NDVI and EVI). While traditional methods are favored for their computational simplicity, their ability to distinguish between goji berries and similar ground features is extremely limited. This invention boldly introduces multi-temporal data (overcoming temporal limitations) and feature optimization (overcoming dimensional limitations), combined with a nonlinear machine learning model, proving that only by constructing a high-dimensional optimal feature space can high-precision identification of goji berries be truly achieved. Second, this invention overcomes the over-reliance on high-resolution data. By utilizing medium-resolution data (such as 10-30m Sentinel-2 / Landsat8), and leveraging the advantages of time series data (trading time for space) and spectral feature optimization, it demonstrates that, while meeting accuracy requirements, high-precision monitoring can also be achieved using lower-cost, wider-coverage medium-resolution data, providing an economical and efficient alternative for large-scale routine monitoring. Finally, this invention breaks the general technical bias that "one set of remote sensing methods is applicable to all crops," and emphasizes that wolfberry, a special economic crop, needs to be customized for multi-temporal and multi-feature integrated analysis based on its unique phenological and spectral characteristics. This customized technical approach "centered on crop characteristics" is the key to achieving high accuracy. Attached Figure Description

[0070] Figure 1 A flowchart of a high-precision remote sensing monitoring method for wolfberry planting distribution based on multi-temporal fusion and spectral feature optimization provided in this invention embodiment;

[0071] Figure 2-1 Map showing the distribution of wolfberry planting areas; Figure 2-2 Initial classification results of typical land parcels; Figure 2-3 Classification results after multi-temporal fusion.

[0072] Figure 3 -A time-series variation curve of NDVI in wolfberry; Figure 3 - B Orchard NDVI Time-Series Variation Curve; Figure 3 -C Shrub NDVI temporal variation curve; Figure: Comparison of NDVI temporal variation curves for typical land features. Detailed Implementation

[0073] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0074] This invention provides a high-precision remote sensing monitoring method for the distribution of wolfberry planting based on multi-temporal fusion and spectral feature optimization. The method includes:

[0075] S1: Obtain data on the number of wolfberry planting plots in the study area, including large-scale planting areas such as enterprises and cooperatives, as well as scattered planting areas by farmers;

[0076] S2: Determine the time range of the growth period of Lycium barbarum in the study area; since Lycium barbarum is a continuously flowering and fruiting plant with indeterminate inflorescences, its growth period is relatively long, and the flowering and fruiting period lasts throughout the entire growing season; among them, the budding / leaf unfolding period is from the end of March to mid-April; the flowering / peak fruiting period is from early May to mid-to-late September; the harvesting period is from early June to early October; the autumn fruit ripening / late growth period is from mid-to-late September to October; and the leaf fall / dormancy period is from November to March of the following year.

[0077] S3: Based on the growth period of wolfberry, obtain remote sensing images covering the key phenological stages of wolfberry in the study area;

[0078] S4: Analysis of remote sensing index characteristics of wolfberry throughout its entire growth period;

[0079] S5: Acquisition and preprocessing of multi-source remote sensing data;

[0080] S6: Multidimensional feature extraction and initial feature space construction;

[0081] S7: Enhanced and optimized spectral and temporal characteristics;

[0082] S8: Construction of a spatiotemporal coupled classification model;

[0083] S9: Results output and accuracy assessment.

[0084] This invention, based on the phenological characteristics of wolfberry—namely, its "infinite inflorescences, simultaneous flowering and fruiting, and long growing season"—constructs a high-precision remote sensing identification framework driven by multi-temporal fusion and spectral feature optimization. This framework enables unified and refined monitoring of both large-scale planting areas and scattered plots of land owned by farmers. The basic principle is to robustly extract planting distribution by leveraging the differences in spectral response, vegetation index, and texture structure at different growth stages of wolfberry through a classification model coupled with multi-temporal remote sensing data.

[0085] First, based on the phenological stages of S1–S2, the time boundaries of each stage of wolfberry growth—from leaf expansion, flowering, peak fruiting, harvesting, to leaf fall and dormancy—were clearly defined. The wolfberry growing season spans from March to the following March, causing its canopy spectral and structural parameters to exhibit regular changes across multiple stages: low reflectance in the early leaf expansion stage, increased leaf area index (LAI) and red light absorption during peak fruiting, and a decline in vegetation indices (such as NDVI and EVI) at the end of the season during autumn fruit ripening. Therefore, S3 selected multi-temporal remote sensing images covering the entire process of "leaf expansion—peak fruiting—harvesting—leaf fall" to achieve continuous observation of wolfberry phenological changes.

[0086] Subsequently, in S4, vegetation indices such as NDVI, EVI, GNDVI, NDRE, and NBR2 were calculated based on multi-phase images to analyze the characteristics of index change curves of wolfberry at typical phenological nodes. Due to the characteristics of "long fruiting period, thick leaves, and high spectral stability", wolfberry exhibits a unique spectral trajectory in the red edge and near-infrared bands during the growing season, which is different from crops such as corn, wheat, and alfalfa. This makes the multi-temporal index sequence a key basis for its identification.

[0087] In S5, multiple sources such as Sentinel-2, Landsat, and GF-6 images are fused together. A consistent temporal data raster set is constructed through radiometric calibration, atmospheric correction, resampling, cloud masking, and region cropping, providing physically consistent input for subsequent classification.

[0088] In S6, multidimensional features such as spectral reflectance, vegetation index, texture (GLCM), and topography (DEM / slope) are extracted to construct an initial feature space. Due to the large number of features and the existence of redundancy and collinearity, the features are optimized in S7. Methods such as spectral sensitivity analysis, stepwise regression, and feature importance (RF-Importance / SHAP) are used to screen out the sensitive bands and key time phases that can best distinguish wolfberry from surrounding crops. For example, the red edge index during the flowering and fruiting period, the combined features of near-infrared and short-wave infrared during the harvesting period, and the multi-temporal NDVI change rate ΔNDVI, etc.

[0089] In S8, a classification model is constructed based on the idea of ​​spatiotemporal coupling. The spectral response sequence of wolfberry at different phenological stages is used as input. Through temporal random forest, spatiotemporal CNN, or Transformer classifier based on attention mechanism, feature fusion and decision reasoning across time phases are realized, thereby significantly improving the ability to identify scattered small plots and complex backgrounds.

[0090] Finally, S9 outputs a map of wolfberry planting distribution and uses indicators such as confusion matrix, overall accuracy (OA), and Kappa coefficient to evaluate accuracy. At the same time, it uses field sampling points and enterprise / cooperative planting data for cross-validation to ensure that the monitoring results have high reliability on both large and small scales.

[0091] S5 specifically includes:

[0092] (1) Obtain S1 SAR images, S2 and L8 optical images of wolfberry during the key growth period from March to November in the Ningxia study area;

[0093] (2) Perform orbit correction, radiometric calibration, terrain correction, and Speckle filtering on the acquired S1 data, and extract the VV and VH polarization backscattering coefficients;

[0094] (3) The S2 and L8 data obtained are surface reflectance product data that have been atmospherically corrected and radiometrically calibrated; different bands of the S2 image are resampled into image data with uniform resolution.

[0095] S6 specifically includes:

[0096] (1) Extracting optical features based on S2 and L8: Extracting the original spectral bands of all time phases, including red, green, blue, near-infrared, and short-wave infrared, as the original spectral band features F. B (t i ):

[0097]

[0098] Where m1 and m2 are the number of available bands for S2 and L8, respectively;

[0099] Extracting vegetation index features F VI (t i ):

[0100]

[0101] Where V is the number of selected vegetation indices; all indices are calculated based on S2 and L8 data;

[0102] Single-phase optical characteristics:

[0103] F Opt (t i ) = F B (t i )∪F VI (t i )

[0104] Multi-temporal optical characteristics:

[0105]

[0106] (4) Extract radar features based on S1: Extract the VV and VH backscattering coefficients of all time phases and calculate the polarization ratio, such as VV / VH;

[0107] Single-phase radar characteristics:

[0108]

[0109] Where VV and VH are in phase t i The backscattering coefficient; Ratio is the polarization ratio, such as VV / VH;

[0110] General characteristics of multi-temporal radar:

[0111]

[0112] (5) Extraction of temporal phenological features: Based on key vegetation indices, such as NDVI, extract the phenological features F during the growing season of wolfberry (March-November). Phen Temporal features are extracted based on the vegetation index (e.g., NDVI) change curves throughout the growing season (cover T), and temporal phenological features F. Phen :

[0113] F Phen ={NDVI max NDVI min NDVI avg NDVI std ,DOY peak VI Integral}

[0114] Among them, NDVI max NDVi min NDVI avg NDVI std These represent the maximum, minimum, average, and standard values ​​of NDVI during the growing season; DOY peak It is the number of days that the NDVI peak occurs; VI Integral It is the area under the vegetation index curve during the growing season, expressed as an integral value.

[0115] (4) Feature fusion: All features extracted from S1, S2, and L8 at different times and dimensions, including spectral, radar, index, and time series data, are stacked to construct a high-dimensional initial feature space F. initial ;

[0116]

[0117] Specifically, S7 includes:

[0118] The study focuses on introducing red edge features and vegetation index differences to enhance the separability of wolfberry with similar land features, shrubs, and orchards; for red edge locations, linear interpolation was used, which is suitable for S2 data (using B4, B5, B6, and B7 bands).

[0119]

[0120] Where ρNIR and ρRed are the reflectances in the near-infrared and red light bands, respectively; λ i and λ j It is the center wavelength of the rising red-edge band; ρ j and ρ i It is the reflectivity of the corresponding waveband;

[0121] Vegetation Index Difference (VID) was used to measure the difference in vegetation index during the dormancy period of wolfberry (T).dormant ) and rapid growth period (T peak The NDVI of ) is used to calculate its degree of difference, which is used to capture phenological differences;

[0122] VID NDVI =NDVI(T peak -T dormant )

[0123] To determine the most discriminative feature subset F final JM distance is used as the criterion for feature selection;

[0124]

[0125] The JM distance ranges from [0, 2]. The closer the JM value is to 2, the better the separability of the two land cover types, such as wolfberry and orchard, on the current feature dimension. Calculate F... final The average JM distance between all features for "goji berries" and "highly confusing features" is used; the top k features with the largest JM distances and their corresponding time phases are selected to form the optimal spatiotemporal feature subset F. final .

[0126] Furthermore, S8 specifically includes:

[0127] The spatiotemporally coupled ST-U-Net model integrates temporal and spatial context information into a semantic segmentation framework. ST-U-Net inherits the encoder-decoder structure of the standard U-Net, but its convolutional layers are designed to handle time-series data. The input layer stacks all Ffinal feature channels of Nt image phases within the optimal temporal combination TOptimal into a spatiotemporal data cube Xinput.

[0128] Xinput∈RH×W×(Nt×Nc_final)

[0129] Where H×W is the spatial size of the image block, and Nc_final is the final number of feature channels.

[0130] The encoder employs 3D convolutional layers to extract features simultaneously in both spatial (X, Y) and temporal (T) dimensions, and automatically learns spatiotemporally coupled features.

[0131] The decoder fuses high-level semantic features captured by the encoder with low-level detail features through upsampling and skip connections, achieving accurate boundary recovery and pixel-level classification.

[0132] The model employs the Adam optimizer with the cross-entropy loss function. CE is trained:

[0133] CE=-1Ni=1Nc=1CYi,clog(Yi,c)

[0134] Where Yi,c is the true label and Yi,c is the model's predicted probability.

[0135] S9 specifically includes:

[0136] Based on the above model, the wolfberry planting areas were extracted, and a wolfberry planting distribution layer was output. The accuracy of the classification results was evaluated using an independent validation sample set. The wolfberry planting area of ​​the entire Ningxia region was extracted, and a wolfberry planting distribution map was created.

[0137] Example 1: Goji Berry Identification Process Based on S1+S2 Multi-Temporal Fusion

[0138] This embodiment is applied to the main wolfberry producing area in Zhongning County, Ningxia, and adopts an integrated approach combining multi-temporal optical (Sentinel-2, S2-MSI) and radar (Sentinel-1, S1-SAR). First, based on the actual phenology of wolfberry in the region, the budding to leaf-fall period is divided into six key phenological periods, and corresponding S1 radar and S2 optical images are acquired for each period. Subsequently, orbital correction, radiometric calibration, topographic correction, and speckle filtering are performed on the S1 data; cloud masking, atmospheric correction, and projection unification are performed on the S2 images. After all images are prepared, an initial feature set including optical bands, radar features, vegetation indices, and temporal phenological characteristics is extracted.

[0139] Subsequently, red-edge feature calculation and vegetation index difference calculation were performed to enhance the distinguishability of wolfberry plants with similar land cover features (such as shrubs and orchards). A feature selection method based on Jeffries-Matusita distance was adopted to select the most separable spatiotemporal feature subset from the candidate feature space. Finally, a classifier based on a random forest model was constructed, and the optimal feature subset was input into the model to classify wolfberry plots. Overall accuracy and user accuracy were calculated using independent validation samples to generate a wolfberry planting distribution map for the study area.

[0140] Example 2: Backscattering Feature Extraction from S1 Radar Data

[0141] In this embodiment, data from May to August, the rapid growth period of wolfberry, is used as the key analysis period, and six Sentinel-1 (S1) radar images are acquired. Radiometric calibration and topographic correction are performed on the S1 radar images, and Lee filtering is used to suppress speckle noise. After obtaining the standardized backscattering coefficients, vertical-vertical (VV) polarization backscattering and vertical-horizontal (VH) polarization backscattering are extracted, and their ratio is calculated as the polarization ratio feature. The radar features from all time phases are stacked to form a radar feature sequence with six time dimensions.

[0142] Typical samples were collected from different wolfberry plots, orchards, and other shrub plots. Analysis of the temporal changes in backscattering from different land cover types revealed that the backscattering coefficient of wolfberry increased significantly more than that of shrubs during their rapid growth period. Based on this, multi-temporal radar features were combined with optical features. In the subsequent feature screening stage, radar features were proven to significantly contribute to distinguishing wolfberry from psammophytic shrubs. This embodiment verifies the effectiveness of radar features as an independent feature system, especially its irreplaceable role in overcoming spectral confusion between wolfberry and psammophytic shrubs.

[0143] Example 3: S2 optical band fusion and vegetation index construction

[0144] In this embodiment, all usable Sentinel-2 (S2) images from March to November 2024 were acquired for the sample area in Zhongning County, Ningxia. First, the 20m resolution red-edge bands (B5, B6, B7) and shortwave infrared bands (B11, B12) were resampled to 10m, forming a dataset with unified spatial resolution along with the visible and near-infrared bands (B2, B3, B4, B8). Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EDI), Soil-Regulated Vegetation Index (SDI), and Red-Edge Vegetation Index (REDI) were calculated using a surface reflectance-based method, resulting in 12 vegetation index feature sets. All vegetation indices were calculated independently for each time phase.

[0145] Subsequently, vegetation indices from all time phases within each plot were multidimensionally stacked, and the temporal sequence of each index was expanded into multidimensional spectral information. Experiments showed that the near-infrared reflectance of wolfberry during its peak fruiting period was significantly higher than that of orchards, while the rate of change in red light reflectance during the autumn leaf-fall period was significantly faster than that of evergreen or late-decaying shrubs. This embodiment provides complete support for the construction and sufficiency of optical vegetation index feature sets.

[0146] Example 4: Extraction of temporal phenological features based on NDVI variation curves

[0147] This embodiment sets up a typical wolfberry planting area in Huinong District, Ningxia, and extracts the annual NDVI sequence. First, all images from March to November are extracted, and cloud-contaminated pixels are removed. To address the temporal gaps caused by cloud cover, Savitzky-Golay filtering or linear interpolation methods are used to reconstruct and smooth the NDVI temporal curve, forming a continuous and smooth growth curve. Based on the reconstructed curve, the maximum value (Max), minimum value (Min), mean (Mean), and standard deviation (Std) are calculated, and the date corresponding to the peak vegetation index is found using first-order differentiation or dynamic thresholding. Furthermore, the area enclosed by the entire season's NDVI curve and the time axis is calculated using the trapezoidal integral method as the growing season integral.

[0148] The results showed that, due to pruning and harvesting management, the peak date of wolfberry often occurred earlier than that of extensively managed orchard vegetation, and its growing season integral differed significantly from that of low-lying desert shrubs. This example fully demonstrates the separability of phenological characteristics along the temporal dimension.

[0149] Example 5: Red-edge linear interpolation feature

[0150] In this embodiment, red band, red edge band, and near-infrared band data from S2 imagery are selected, and the red edge position feature is calculated through linear interpolation. Specifically, linear interpolation is performed using the reflectance of the red edge band to determine the wavelength position corresponding to the maximum value of the first derivative of the red edge. This red edge feature can sensitively reflect subtle changes in vegetation chlorophyll content and biomass.

[0151] Experimental results show that the chlorophyll content of wolfberry changes significantly during its vigorous growth period, with a marked upward shift in the red edge position feature, while the changes in shrubs and orchard vegetation are smaller. By introducing this feature, the separability of wolfberry from similar land features in the classification model is significantly improved. This example provides specific operation and effect demonstration of the spectral detail enhancement feature.

[0152] Example 6: Construction of Vegetation Index Difference (VID)

[0153] In this embodiment, dormant period images (February to March) and rapid growth images (May to June) are used as two key time phases. NDVI is calculated for each set of images, and the difference between the NDVI during the rapid growth phase and the NDVI during the dormant phase is calculated at the pixel scale as the vegetation index difference. This difference reflects the magnitude of crop phenological changes.

[0154] Statistical analysis of wolfberry and shrub samples revealed that the difference in phenological characteristics was approximately 15% higher in wolfberry than in shrubs. This difference reflects the rapid greening and leaf growth of wolfberry. This example fully supports a VID feature constructed based on phenological differences.

[0155] Example 7: Feature Filtering Based on JM Distance

[0156] In this embodiment, distance metrics are calculated for the entire feature set across multiple dimensions, including optical, radar, and phenological data. Goji berries and two main confusing land cover types (orchards and shrubs) are selected as the comparison group, and the Jeffries-Matusita (JM) metric is used to calculate the inter-category distance for each feature. The JM distance ranges from 0 to 2, with higher values ​​indicating stronger separability. Features with the highest average distance values ​​are selected as the optimal feature combination based on their ranking.

[0157] A classification model was constructed using this optimal feature set and compared with a model constructed using all features. The results show that the overall accuracy of the optimal feature model is improved by more than 10%. This embodiment provides a complete feature selection process and performance verification, supporting the technical effectiveness of the feature optimization steps.

[0158] Example 8: Spatiotemporal Coupled Classification Model

[0159] This embodiment constructs a random forest classification model based on optimal features. The data input to the classifier includes multiple spatiotemporal enhancement features such as NDVI during the rapid growth phase, red-edge features, radar polarization ratio features, and phenological integral. The model training uses over six thousand sample points, randomly divided into training and test sets at an 8:2 ratio, covering major land types such as goji berries, orchards, shrubs, and bare land. After model training, the entire region's imagery is input for classification.

[0160] Accuracy was evaluated using independent validation samples, achieving an overall accuracy of 93%, with both user and producer accuracy exceeding 90%. This model validates the effectiveness of multi-temporal and multi-dimensional feature fusion, providing support for the model construction method described in claim 9.

[0161] Example 9: Result Output and Goji Berry Area Statistics

[0162] This embodiment outputs a raster layer depicting the distribution of goji berries based on a classification model, and further uses administrative division vectors to perform regional statistical analysis on the classification results. The goji berry planting area is statistically analyzed for areas such as Zhongning County, Hongsipu District, and Shapotou District. The area calculation uses a pixel-by-pixel cumulative summation method to ensure consistent statistical standards across regions.

[0163] In the accuracy verification stage, verification samples are constructed based on plot boundaries delineated using manual field surveys and high-resolution imagery. By comparing the classification results with the actual plots and calculating the overall accuracy and Kappa coefficient, the final accuracy meets the requirements for regional agricultural monitoring. This embodiment fully supports the technical requirements of claim 10 regarding "accuracy evaluation and result output".

[0164] Example 10: Stabilization Analysis Based on Multi-Source Data Fusion

[0165] This embodiment analyzes the contribution of different data sources to different growth stages of wolfberry by jointly fusing three types of images: S1, S2, and L8. Four feature combinations were selected for the experiment: single-source optical imagery, single-source radar imagery, optical + radar fusion, and optical + radar + phenological fusion. The classification accuracy under different combinations was compared, and the importance of each feature was analyzed.

[0166] The results show that single-source optical methods have the lowest accuracy, radar features can effectively improve the identification ability during the rapid growth period, while three-source fusion performs the best. When phenological features are added, the separation degree between wolfberry and shrubland features is the highest. This embodiment provides enhancement, supplementation, and verification of the overall technical effect of the system method.

[0167] Evidence related to the technical effects obtained by the embodiments of the present invention.

[0168] (1) Quantitative evaluation of classification accuracy

[0169] To verify the final recognition accuracy of the method of this invention, 1200 independent verification sample points were collected using high-resolution imagery combined with ground-based point marking. Based on the constructed "multi-temporal S1+S2+feature optimization+random forest" model, the classification confusion matrix is ​​shown in the table below:

[0170] Table 1. Confusion matrix extracted from wolfberry planting distribution.

[0171]

[0172] Calculations show that the overall accuracy of this method reaches 92.9%. The Kappa coefficient reaches 0.9, indicating that the classification results have extremely high consistency with the actual ground conditions. Key breakthrough: For the most difficult-to-distinguish category, "goji berries," the producer accuracy (low omission rate) reaches 93.4%, and the user accuracy (low misclassification rate) reaches 91.7%, proving that this invention effectively solves the problem of confusion between goji berries and similar ground features.

[0173] (2) Gain analysis of multi-source feature combinations (ablation experiment)

[0174] To demonstrate the necessity of "multi-temporal fusion" and "radar feature introduction," four sets of comparative experimental schemes were designed. The experimental conditions remained consistent, with only the input feature source being changed.

[0175] Table 2: Comparison of classification accuracy for different feature combination schemes

[0176]

[0177] The ablation experiment results show that the accuracy increases stepwise from scheme A to scheme D using different feature combinations, which intuitively reflects the cumulative effect of each technical improvement point of the present invention.

[0178] (3) Analysis of temporal phenological characteristic curves

[0179] The accompanying figures in this specification include three categories of content: the distribution of wolfberry plantings, classification results, and NDVI temporal characteristics of typical land cover types. These figures support the completeness and understandability of the multi-temporal fusion and spectral feature optimization method of this invention. Among them, Figure 2 It consists of three sub-graphs: Figure 2-1 This is a map showing the distribution of wolfberry planting areas, used to illustrate the overall spatial pattern of large-scale planting plots and scattered farmer plots within the study area; Figure 2-2 This is an initial classification result map of a typical plot, reflecting the classification model's ability to distinguish between wolfberry and surrounding land features before the integration of key phenological periods; Figure 2-3 The image shows the classification results after multi-temporal fusion. By introducing the spectral temporal characteristics of the wolfberry growth period, the boundary extraction accuracy and background suppression effect of the wolfberry region were significantly improved. Figure 3 This demonstrates the temporal variation patterns of NDVI for different vegetation types, among which... Figure 3 -A is the time-series variation curve of NDVI of wolfberry, reflecting its continuous growth period characteristics from leaf expansion, flowering, fruiting to harvest; Figure 3 -B is the orchard NDVI time-series variation curve, used to compare time-series differences with those of goji berries; Figure 3 -C is a time-series variation curve of NDVI for shrubs, showing the seasonal fluctuations of non-goji berry vegetation. All figures together illustrate the technical principle and implementation effect of this invention in achieving high-precision identification of goji berry planting areas using multi-temporal remote sensing features.

[0180] Comparative analysis reveals that goji berry growing areas exhibit a typical "bimodal" or "broad-peak" structure. Rapid greening occurs in mid-to-late March, peak fruiting occurs in May (peak 1), pruning in June causes a slight decline, and high values ​​are maintained again in August (peak 2), followed by rapid leaf drop in November. In orchards (such as jujube trees), peak values ​​occur from July to October, with a relatively smooth curve and no significant fluctuations. For shrubs, NDVI values ​​remain relatively stable at low levels between February and June, with an almost flat curve, indicating that their greening and leaf growth periods are significantly delayed compared to goji berries. The peak occurs in August and lasts for a shorter period. This provides a solid biological and statistical basis for the automated classification using temporal characteristics in this invention.

[0181] It should be noted that embodiments of the present invention can be implemented in hardware, software, or a combination of both. The hardware portion can be implemented using dedicated logic; the software portion can be stored in memory and executed by a suitable instruction execution system, such as a microprocessor or dedicated-design hardware. Those skilled in the art will understand that the above-described devices and methods can be implemented using computer-executable instructions and / or included in processor control code, for example, such code provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuitry such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field-programmable gate arrays, programmable logic devices, etc., or by software executed by various types of processors, or by a combination of the above-described hardware circuitry and software, such as firmware.

[0182] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention, and within the spirit and principles of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for monitoring the distribution of Chinese wolfberry planting based on multi-temporal fusion and spectral feature optimization, comprising the following steps: Step 1: obtaining the number information of Chinese wolfberry planting plots in the study area; Step 2: determining the time range of the growth period of Chinese wolfberry and the key phenological period according to the growth period; Step 3: obtaining multi-temporal remote sensing images covering the key phenological period of Chinese wolfberry; Step 4: extracting optical features, radar features and time-series phenological features based on the multi-temporal images, and constructing an initial feature set containing multiple dimensions; Step 5: introducing red edge features and vegetation index difference to enhance the distinguishability of Chinese wolfberry and similar objects; Step 6: selecting the highest separability features from the initial feature set using distance-based method to form an optimal spatio-temporal feature subset; Step 7: constructing a classification model based on the optimal spatio-temporal feature subset, outputting the Chinese wolfberry planting distribution results and performing accuracy evaluation. 2.The method of claim 1, wherein the step 3 comprises: obtaining multi-temporal radar remote sensing data and optical remote sensing data covering the period from March to November during the growth period of Chinese wolfberry; performing orbit correction, radiation calibration, terrain correction and speckle noise filtering on the radar data, and extracting the vertical polarization backscattering coefficient and the horizontal polarization backscattering coefficient; performing atmospheric correction and radiation calibration on the optical data, and performing uniform spatial resolution processing on different optical bands. 3.The method of claim 1, wherein the optical features comprise: extracting the reflectivity of red light, green light, blue light, near-infrared light and short-wave infrared light at each time phase; calculating a plurality of vegetation indices based on the reflectivity, and fusing the vegetation index features of each time phase to form a multi-temporal optical feature set. 4.The method of claim 1, wherein the radar features comprise: extracting the vertical polarization backscattering coefficient, the horizontal polarization backscattering coefficient and the ratio of the two at each time phase; stacking the radar features of all time phases to form a multi-temporal radar feature set. 5.The method of claim 1, wherein the time-series phenological features comprise: extracting the maximum value, minimum value, average value and standard deviation of the vegetation index curve during the entire growing season; extracting the date when the vegetation index reaches the peak value; calculating the area of the region enclosed by the vegetation index curve and the time axis during the growing season as the integral quantity. 6.The method of claim 1, wherein the red edge position feature is determined by linear interpolation, and the linear interpolation is based on the red light band reflectivity, the near-infrared band reflectivity and the center wavelength of the red edge rising segment to calculate the red edge position value, which is used as a feature to enhance the separability of Chinese wolfberry and shrubs and orchards. 7.The method of claim 1, wherein the vegetation index difference is obtained by calculating the difference between the vegetation index in the dormancy period of Chinese wolfberry and the vegetation index in the rapid growth period, which is used to reflect the seasonal phenological difference of Chinese wolfberry. 8.The method of claim 1, wherein the distance metric is Jeffries-Matusita distance, and the value range is 0 to 2; calculating the average distance between Chinese wolfberry and confusing objects for each candidate feature, and selecting a number of features with the largest distance value as the optimal feature set.

9. The method of claim 1, wherein the classification model is constructed based on an optimal spatiotemporal feature subset, the model simultaneously utilizes multi-temporal information and multi-dimensional spectral features, classifies the wolfberry pixels through machine learning, and outputs a wolfberry planting distribution layer.

10. The method of claim 1, wherein the precision evaluation uses independent validation samples, calculates overall precision, user precision, and producer precision based on a confusion matrix, and determines the wolfberry planting area and spatial distribution according to the evaluation results.

Images