A method and system for extracting cultivation attributes of farmland

Abstract

This invention belongs to the field of remote sensing image processing and farmland planting attribute extraction technology. It provides a method and system for extracting farmland planting attributes. The method uses acquired radar and optical remote sensing images to construct a multi-source crop distribution dataset to generate preliminary samples and standard phenological curves; it uses the DTW algorithm to match and screen candidate samples, and combines spatial constraints and manual cleaning to obtain high-confidence samples; it uses a monthly time discretization strategy to reconstruct the spatiotemporal feature matrix of multi-source feature fusion; it constructs a Fuzzy-TWDTW model to solve for the optimal matching path, and simultaneously outputs a farmland planting attribute classification map and a classification uncertainty evaluation map; this invention realizes fully automatic sample construction, enhances model robustness, and improves classification accuracy and stability.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing image processing and farmland planting attribute extraction technology, specifically relating to a method and system for extracting farmland planting attributes. Background Technology

[0002] Farmland planting attributes are core data for agricultural monitoring and food security. Traditional manual surveys are costly and have significant time delays; existing remote sensing methods mostly rely on single optical time series, sample acquisition depends on manual labor, and standard DTW is sensitive to phenological variations and mixed pixels. In areas with frequent cloud and rain, fragmented plots, and complex crop rotation, such as the middle and lower reaches of the Yangtze River, the accuracy is insufficient, making it impossible to achieve automated, high-precision, and robust extraction. The technical solution in patent document CN115512233A still has obvious defects: First, it only uses Sentinel-2 single optical data to construct NDVI time series, which is greatly affected by cloud and rain weather, is prone to data loss, and has limited adaptability; Second, the sample generation relies on previous land use survey data and cluster analysis, which cannot achieve automatic sample amplification, consumes manpower, and has low sample update efficiency; Third, it uses the standard DTW algorithm for category matching, which has poor robustness to crop phenological shifts, mixed pixels, and fragmented plots, making it difficult to further improve classification accuracy; Fourth, it only outputs basic planting attributes, does not involve the identification of fine cover types such as crop rotation patterns, and does not output classification uncertainty assessment, which cannot quantify the reliability of the identification results and is difficult to meet the needs of fine monitoring of cultivated land in complex areas. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and system for extracting arable land planting attributes.

[0004] To achieve the objectives of this invention, the following technical solutions are adopted.

[0005] A method for extracting arable land planting attributes includes the following steps: S1: Preprocess the radar and optical remote sensing data of the area to be measured; S2: Construct a multi-source public crop distribution dataset with unified spatial coordinates and classification system. Through spatial consistency test, remove heterogeneous pixels at the edge of plots or with low confidence, generate preliminary crop samples, and construct initial standard crop phenological curves. S3: Construct the time series feature curve of the pixel to be tested using the preprocessed radar and optical remote sensing data. Use the DTW algorithm to match and calculate the minimum distance between the time series feature curve of the pixel to be tested and the initial standard crop phenology curve. Select the pixels to be tested with the largest minimum distance as candidate samples. Combine with the high-resolution cultivated land base map to implement spatial constraints and eliminate non-cultivated land pseudo samples. After manual visual cleaning, obtain high confidence samples. S4: Based on high-confidence samples, standardized crop phenological reference curves are extracted through spatial aggregation functions and spatial low-pass filtering. Logistic nonlinear time weights are introduced to complete pixel-level dynamic phenological alignment and reconstruct the spatiotemporal feature matrix of multi-source feature fusion. S5. Construct a Fuzzy-TWDTW model. Using the spatiotemporal feature matrix as input, construct the fuzzy feature space through Gaussian fuzzy membership, apply Logistic nonlinear time weighting constraints, and carry out differential weighted fusion of optical and radar multi-source features. Solve the optimal matching path for accumulated fuzzy cost through dynamic programming. Use the normalized centroid method to complete the defuzzification process and simultaneously output the plot-level cultivated land planting attribute classification map and classification uncertainty assessment map.

[0006] Furthermore, the preprocessing includes: The radar remote sensing images were sequentially subjected to thermal noise removal, radiometric calibration and terrain correction, and the VV and VH dual-polarization backscattering coefficients were extracted. Clouds and shadows were removed from optical remote sensing images using the QA60 band, and interpolation and filtering algorithms were used to fill in time gaps and smooth the images.

[0007] Furthermore, the multi-source crop distribution dataset includes CCD-Maize, China rapeseed maps, ChinaRice, China Wheat, ChinaCP, and ChinaCP-Wheat10m.

[0008] Furthermore, the high-confidence sample covers 11 land cover types: fallow, 4 single-season planting types, 5 crop rotation patterns, and other multiple cropping types.

[0009] Furthermore, a time discretization strategy is adopted to divide the time steps of high-confidence samples by month. The median synthesis algorithm is used to suppress noise in each monthly window. When optical observation fails, the radar priority complementarity mechanism is triggered to fill the time gap.

[0010] Furthermore, the multi-source characteristics include: biomass characteristics: NDVI, EVI, MSAVI; water characteristics: LSWI, NDWI; pigment and structural characteristics: NDRE, VV, VH.

[0011] Furthermore, the Gaussian fuzzy membership degree is used to convert temporal hard distance matching into soft membership degree mapping, reducing the interference of mixed pixels and crop phenological spectral variations on classification accuracy.

[0012] Furthermore, the Logistic nonlinear time-weighted constraint employs the Logistic function, which assigns a low penalty within a set fluctuation threshold, and the penalty increases exponentially beyond the fluctuation threshold, thereby achieving asynchronous and flexible differentiation of phenology.

[0013] Furthermore, the defuzzification process converts fuzzy membership degrees into hard classification results and outputs a classification uncertainty assessment graph using complementary values ​​of fuzzy membership degrees.

[0014] A system for extracting arable land planting attributes, comprising: The image preprocessing module is used to preprocess the acquired radar remote sensing data and optical remote sensing data; The module for constructing preliminary samples and initial standard curves is used to build a multi-source crop distribution dataset with unified spatial coordinates and classification system. After spatial consistency testing, low-confidence marginal heterogeneous pixels are removed to generate preliminary crop samples and construct initial standard crop phenological curves. The DTW sample automatic amplification module constructs the time-series feature curve of the pixel to be measured using preprocessed radar remote sensing data and optical remote sensing data. The DTW algorithm is used to match the time-series feature curve of the pixel to be measured with the initial standard crop phenology curve to calculate the minimum distance. Pixels close to the minimum distance are selected as candidate samples. Spatial constraints are implemented by combining high-resolution cultivated land base map to eliminate non-cultivated land pseudo samples. After manual visual cleaning, high-confidence samples are obtained. The spatiotemporal feature matrix reconstruction module, based on high-confidence samples, extracts standardized crop phenological reference curves through spatial aggregation statistics and spatial low-pass filtering, introduces Logistic nonlinear time weights to complete pixel-level dynamic phenological alignment, and reconstructs the spatiotemporal feature matrix of multi-source feature fusion. The Fuzzy-TWDTW classification and uncertainty output module constructs a Fuzzy-TWDTW model. Taking the spatiotemporal feature matrix as input, it sequentially constructs a fuzzy feature space through Gaussian fuzzy membership, applies Logistic nonlinear time weighting constraints, and performs differential weighted fusion of optical and radar multi-source features. The optimal matching path for cumulative fuzzy cost is solved through dynamic programming. The normalized centroid method is used to complete the defuzzification process, and the module simultaneously outputs a plot-level farmland planting attribute classification map and a classification uncertainty assessment map.

[0015] The advantages of this invention are: 1. Fully automated sample construction: No manual pixel-by-pixel interpretation is required; large-scale single-crop and rotation samples are automatically generated. 2. The model is extremely robust: fuzzy logic tolerates spectral variations, nonlinear time weights adapt to phenological shifts, multi-source fusion enhances recognition, and it is suitable for fragmented plots. 3. High accuracy and stability: OA reaches 86%~94%, F1 for wheat / rapeseed is improved by more than 5%, and the uncertainty of synchronous output is reduced; 4. Integrated process, easy to promote: adaptable to cloudy and rainy areas, multiple sensors, and multiple regions, with strong parameter self-adaptation. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the method for extracting arable land planting attributes provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the crop standard time-series curve template construction process provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the Fuzzy-TWDTW model construction and calculation process provided in the embodiments of the present invention; Figure 4 This is a schematic diagram of the fuzzy distance calculation process based on feature values ​​and center values ​​provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the nonlinear time weight generation process provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of the multi-feature fusion distance calculation process based on Fuzzy-TWDTW provided in an embodiment of the present invention; Figure 7 This is a schematic diagram of the system architecture for extracting arable land planting attributes provided in an embodiment of the present invention. Detailed Implementation

[0017] The present invention will be further described in conjunction with the embodiments and accompanying drawings.

[0018] As an embodiment 1 of the present invention, such as Figure 1 As shown, a method for extracting arable land planting attributes aims to address the problems of low sample acquisition efficiency, sensitivity of classification models to phenological variations, and inability to quantify the reliability of identification results in existing technologies. This method achieves automated and high-precision extraction of arable land planting attributes by constructing a closed-loop logical process of "multi-source data acquisition - sample amplification - feature reconstruction - Fuzzy model classification." The method specifically includes the following steps: Step S1: Acquire Sentinel-1 radar and Sentinel-2 optical remote sensing images of the area to be measured using the GEE platform and perform preprocessing.

[0019] Specifically, this step aims to address the adaptability of a single data source under complex meteorological conditions. The middle and lower reaches of the Yangtze River are characterized by frequent cloud cover and rain, making reliance solely on optical imagery prone to data gaps in time series. This embodiment leverages the Google Earth Engine (GEE) cloud platform, which features open-source cloud removal, DTW algorithms, and the TWDTW core algorithm, to collaboratively acquire Sentinel-1 radar imagery and Sentinel-2 optical imagery. Radar imagery possesses all-weather imaging capabilities, penetrating cloud layers to capture crop structure information; optical imagery is rich in spectral information, enabling detailed characterization of crop biomass and pigment features. By preprocessing both data sources, sensor noise, atmospheric interference, and terrain effects are eliminated, providing a high-quality data foundation for subsequent feature extraction.

[0020] Step S2: Construct a multi-source crop distribution dataset with unified spatial coordinates and classification system. Remove low-confidence edge heterogeneous pixels through spatial consistency test, generate preliminary crop samples, and construct initial standard crop phenological curves.

[0021] Specifically, this step aims to address the issues of scattered and inconsistent preliminary sample sources and standards. This embodiment does not rely on a single survey data point, but rather integrates existing multi-source publicly available crop distribution data to construct a unified spatial reference framework. Since different datasets differ in classification accuracy and spatial resolution, direct fusion can easily introduce noise. Therefore, this step introduces a spatial consistency check mechanism. By analyzing the consistency of pixel neighborhoods, heterogeneous pixels located at plot edges or with low classification confidence are removed, retaining only contiguous and homogeneous pixels as preliminary crop samples. The mean of their temporal characteristics is extracted to construct initial standard crop phenological curves, providing a benchmark for subsequent sample expansion.

[0022] Step S3: Construct the temporal feature curve of the pixel to be measured using the preprocessed radar and optical remote sensing images. Use the DTW algorithm to match and calculate the distance between the temporal feature curve of the pixel to be measured and the initial standard crop phenology curve DTWmin. Select the top 20% of pixels in terms of DTWmin distance as candidate samples. Combine the high-resolution cultivated land base map to implement spatial constraints and eliminate non-cultivated land pseudo samples. After manual visual cleaning, obtain high-confidence samples.

[0023] Specifically, this step is crucial in addressing the issue of insufficient sample size. Preliminary crop samples are often limited in number and insufficient to cover complex planting patterns. This embodiment utilizes the nonlinear warping capability of the Dynamic Time Warping (DTW) algorithm to calculate the minimum distance (DTWmin) between the pixel to be tested and the initial standard curve; a smaller distance indicates higher similarity. To balance sample quantity and quality, this embodiment selects the top 20% of pixels with the highest similarity as candidate samples. Considering the phenomenon of "heterogeneous objects sharing the same spectrum" (e.g., grassland and crops) with similar spectra but different land types, this step further introduces a high-resolution cultivated land base map for spatial constraints, forcibly eliminating pseudo-samples falling in non-cultivated areas. Finally, a small amount of manual visual cleaning ensures high sample confidence, achieving automated amplification from "small samples" to "large-scale, high-quality samples."

[0024] Step S4: Based on high-confidence samples, the year is divided into 12 standard time steps using a monthly time discretization strategy on the GEE platform; each monthly window uses a median synthesis algorithm to suppress noise, and a radar priority complementarity mechanism is triggered to fill the time gap when optical observation fails; standardized crop phenological reference curves are extracted through spatial aggregation statistics and spatial low-pass filtering, and Logistic nonlinear time weights are introduced to complete pixel-level dynamic phenological alignment, reconstructing an 8×12-dimensional spatiotemporal feature matrix of multi-source feature fusion.

[0025] Specifically, this step aims to address the issues of temporal feature standardization and data gaps. The original remote sensing images have inconsistent temporal resolutions and are severely affected by cloud shadows. This embodiment employs a "monthly temporal discretization" strategy, dividing the entire year into 12 standard time steps and using median synthesis to effectively suppress outliers. In particular, for cases where optical imagery is completely ineffective in a specific month, a radar priority complementarity mechanism is triggered, using radar data to fill in the gaps and ensure the integrity of the temporal sequence. Building upon this, a Logistic nonlinear temporal weight is introduced to flexibly constrain the temporal shift of crop phenology, allowing the standard curve to adapt to differences in sowing periods across different regions. Ultimately, the reconstructed 8×12-dimensional spatiotemporal feature matrix includes multi-dimensional features such as biomass and water content, while also possessing temporal standardization and robustness.

[0026] Step S5: Construct a Fuzzy-TWDTW model. Using an 8×12-dimensional spatiotemporal feature matrix as input, the model sequentially constructs a fuzzy feature space through Gaussian fuzzy membership, applies Logistic nonlinear time weighting constraints, and performs differential weighted fusion of optical and radar multi-source features. The optimal matching path for accumulated fuzzy cost is solved through dynamic programming. Defuzzification is completed using the normalized centroid method, and a plot-level farmland planting attribute classification map and a classification uncertainty assessment map are output simultaneously.

[0027] Specifically, this step is crucial for improving classification accuracy and reliability. Traditional hard classification methods are extremely sensitive to mixed pixels and spectral variations. This embodiment constructs a Fuzzy-TWDTW model. First, it uses Gaussian fuzzy membership to transform "hard distance" into "soft membership," allowing for certain spectral differences among similar crops. Second, it uses Logistic nonlinear time weighting to reasonably reward or penalize phenological time shifts. Finally, it integrates the advantages of optical and radar through multi-source feature differential weighted fusion. The model solves for the optimal matching path through dynamic programming and outputs the final classification result using the normalized centroid method. Unlike traditional methods, this step simultaneously outputs a classification uncertainty assessment map, quantifying the confidence level of each pixel's classification result. This provides an intuitive reliability basis for subsequent manual verification and decision-making, overcoming the deficiency of traditional methods that "only provide results without confidence levels."

[0028] As Embodiment 2 of the present invention, based on Embodiment 1, this embodiment focuses on a detailed description of the specific implementation process of data preprocessing and sample construction. For the cloudy and rainy regions such as the middle and lower reaches of the Yangtze River, data quality and sample accuracy are the cornerstones of subsequent model training and classification accuracy.

[0029] First, regarding the preprocessing in step S1, this embodiment employs the following specific methods: thermal noise removal, radiometric calibration, and terrain correction are performed sequentially on the Sentinel-1 radar remote sensing image, and the VV and VH dual-polarization backscattering coefficients are extracted; clouds and shadows are removed from the Sentinel-2 optical remote sensing image using the QA60 band, and interpolation and filtering algorithms are used to complete temporal gap filling and smoothing.

[0030] Specifically, Sentinel-1, as a side-looking radar sensor, produces radar images containing significant thermal noise, and its imagery is affected by terrain undulations, resulting in overlay and shadowing. This embodiment utilizes a preprocessing algorithm on the GEE platform. First, thermal noise is removed to purify the signal. Second, radiometric calibration is performed, converting digital quantization values ​​into physically meaningful backscattering coefficients. Finally, terrain correction is performed using the SRTM digital elevation model to eliminate the distortion effect of terrain on backscattering intensity, thereby accurately extracting VV (vertical transmit, vertical receive) and VH (vertical transmit, horizontal receive) dual-polarization data reflecting crop structure and moisture content. For Sentinel-2 optical imagery, although its high-resolution multispectral information is rich, it is highly susceptible to cloud and shadow obstruction. This embodiment utilizes cloud masking information from the QA60 band to accurately identify and remove contaminated pixels. For the temporal gaps resulting from the removal, linear interpolation combined with Savitzky-Golay filtering or Whittaker filtering is used for filling and smoothing, ensuring the continuity and smoothness of the crop growth curve and avoiding deviations in phenological feature extraction due to missing data.

[0031] Secondly, regarding the construction of the multi-source crop distribution dataset in step S2, this embodiment selects specific data sources: the multi-source crop distribution dataset includes CCD-Maize, China rapeseed maps, China Rice, ChinaWheat, ChinaCP, and ChinaCP-Wheat10m.

[0032] Because the aforementioned datasets originate from different research institutions or projects, their original projected coordinate systems, spatial resolutions, and attribute classification standards vary significantly. Direct overlaying would result in severe spatial misalignment and attribute conflicts. Therefore, this embodiment first performs rigorous coordinate transformation and resampling operations when constructing the dataset, unifying all datasets to the same geographic coordinate system (such as WGS84) and projection zone, and uniformly resampling them to the same resolution benchmark. Based on this, attribute fields are mapped and unified to construct a multi-source crop distribution dataset with a consistent classification system. This operation effectively eliminates noise caused by data source heterogeneity, providing a reliable data foundation for the subsequent generation of high-quality preliminary samples.

[0033] Finally, regarding the coverage of the high-confidence samples, this embodiment clarifies that the high-confidence samples cover 11 types of land cover: fallow, 4 types of single-season planting, 5 types of crop rotation, and other multiple cropping types.

[0034] Specifically, the planting structure in the middle and lower reaches of the Yangtze River is complex, including not only single-season crops such as corn, rice, wheat, and rapeseed, but also a wide variety of crop rotation patterns such as rice-corn, wheat-corn, double-cropping rice, wheat-rice, and rapeseed-rice. Traditional sample sets often focus only on a single crop type, ignoring crop rotation information over time, resulting in classification results that fail to reflect the true planting system. This embodiment introduces DTW sample amplification and manual visual cleaning to construct a refined sample set covering the above 11 types. Figure 2 As shown, after removing heterogeneous pixels at the edges through spatial consistency testing, contiguous and homogeneous crop areas were retained, ensuring the purity of the sample. This sample design, which covers fallow, single-season, and complex crop rotation patterns, can comprehensively characterize the planting attributes of the study area, providing solid training data support for the subsequent Fuzzy-TWDTW model to distinguish easily confused crop rotation patterns.

[0035] As Embodiment 3 of the present invention, based on the above embodiments, this embodiment focuses on a detailed explanation of the multi-source feature selection strategy and the spatiotemporal feature matrix reconstruction process. The selection of features directly determines the classification model's ability to identify crop types, especially in areas with fragmented plots and complex planting systems, where a single feature is often ineffective.

[0036] Specifically, the multi-source features selected in this embodiment include: biomass features: NDVI, EVI, MSAVI; water features: LSWI, NDWI; pigment and structural features: NDRE, VV, VH.

[0037] The selection of the above features is not an arbitrary combination, but rather a deep consideration of crop physiological characteristics and remote sensing physical mechanisms. Specifically, NDVI (Normalized Difference Vegetation Index) is the most commonly used indicator reflecting vegetation biomass, but it is prone to saturation when vegetation cover is high; EVI (Enhanced Vegetation Index) effectively solves the saturation problem in high biomass areas by correcting for soil background and atmospheric effects; MSAVI (Modified Soil-Regulated Vegetation Index) further reduces soil background noise and is suitable for early crop growth. Among the moisture features, LSWI (Land Surface Water Index) is sensitive to vegetation water content and can capture crop water stress; NDWI (Normalized Difference Water Index) helps distinguish between water bodies and irrigated farmland. Among the pigment and structural features, NDRE (Normalized Red Edge Index) is highly sensitive to chlorophyll content and leaf structure changes and is often used for precise identification of crop growth stages. In particular, this embodiment introduces the VV and VH dual-polarized radar backscattering coefficients, which is key to the adaptability of this invention to cloudy and rainy environments. Radar signals have the ability to penetrate clouds and fog. VV polarization is sensitive to crop canopy structure and stem thickness, while VH polarization is significantly responsive to crop biomass and geometry. When optical images are obscured by clouds and become ineffective, radar features can provide indispensable structural information, achieving a complementary relationship between optical and radar features in terms of physical mechanisms.

[0038] Based on the above characteristics, this embodiment constructs an 8×12-dimensional spatiotemporal feature matrix fused from multiple sources. The construction logic of this matrix is ​​as follows: First, a monthly time discretization strategy is adopted to divide the whole year into 12 standard time steps. The revisit cycle of the original remote sensing images is irregular and has a large number of gaps due to the influence of clouds and rain. Directly using daily series data will lead to the curse of dimensionality and a surge in noise. Discretizing the time axis into 12 monthly windows not only preserves the key phenological nodes of the entire crop life cycle but also significantly reduces data redundancy. Second, within each monthly window, a median synthesis algorithm is used to suppress noise. Compared with traditional mean synthesis, median synthesis can effectively remove abnormally high or low values ​​caused by cloud shadows and sensor failures, and retain the typical state of pixels within that month. Finally, in the case of optical observation failure, a radar priority complementarity mechanism is triggered to fill the time gaps. Specifically, when the number of valid optical observations within a monthly window falls below a preset threshold (e.g., less than 2) or the cloud cover percentage exceeds a preset proportion (e.g., exceeding 80%), the system determines that the optical observations are invalid. In this case, Sentinel-1 radar data from that period is prioritized for filling the gap, ensuring the integrity and continuity of the spatiotemporal feature matrix. Through the above processing, a standardized spatiotemporal feature matrix containing 8 feature dimensions and 12 time dimensions is finally formed, providing high-quality input data for the subsequent Fuzzy-TWDTW model.

[0039] As Embodiment 4 of the present invention, based on the above embodiments, this embodiment focuses on a detailed description of the core algorithm of the Fuzzy-TWDTW model. This model is key to achieving high-precision classification and uncertainty assessment in the present invention. Its core logic lies in transforming the traditional "hard matching" into a "soft metric" and introducing nonlinear time constraints, thereby effectively addressing the problems of spectral variability and asynchronous phenology in complex farmland environments.

[0040] First, regarding the process of constructing the fuzzy feature space in step S5, this embodiment uses a Gaussian fuzzy membership function. The Gaussian fuzzy membership function converts temporal hard distance matching into a soft membership mapping, reducing the interference of mixed pixels and crop phenological spectral variations on classification accuracy.

[0041] Specifically, traditional Dynamic Time Warping (DTW) algorithms typically employ hard distance metrics such as Euclidean distance, calculating the absolute difference between observed and standard values. However, in fragmented areas, a single pixel often contains a mixed spectrum of crops and background (e.g., soil, water), causing observed values ​​to deviate from the standard curve. Hard distance metrics can impose excessive penalties for this, leading to misclassification. This embodiment introduces a Gaussian membership function to transform hard distance into soft membership in the [0, 1] interval. Figure 4 As shown, for the observed value xt,k of feature k at time point t and the standard reference curve value rt,k, their similarity is calculated using the Gaussian membership function: μt,k=exp(-(xt,k-rt,k)² / 2σ²). Here, σ is the bandwidth parameter, controlling the decay rate of the membership function. The fuzzy distance dspec in the feature space is defined as: dspec(xt,k,rl,k)=1-μt,k. In this way, even if there is a certain deviation between the observed value and the standard value, as long as it is within the allowable range of the bandwidth σ, the fuzzy distance remains at a low level, effectively tolerating spectral variations caused by differences in planting density, growth vigor, or mixed pixel effects in similar crops, thus improving the robustness of the model.

[0042] Secondly, regarding the time dimension matching problem in step S5, this embodiment applies a Logistic nonlinear time-weighted constraint. The Logistic nonlinear time-weighted constraint uses the Logistic function, assigning a low penalty within a set fluctuation threshold, and an exponentially increasing penalty beyond the fluctuation threshold, thus achieving asynchronous and flexible differentiation of phenological phenomena.

[0043] Specifically, crop sowing and harvesting periods are often affected by climate, variety, and human management, resulting in interannual shifts. While the standard DTW algorithm allows for flexible scaling of the time axis, it lacks constraints on the scaling range, easily leading to mismatches of crop curves from different seasons (e.g., matching spring crops to autumn crops). This embodiment constructs a Logistic-type time weighting function, such as... Figure 5 As shown, the calculation formula is as follows: ωtime(ti,tj)=1 / (1+exp(-α·(|ti-tj|-β))). Where |ti-tj| is the time difference between the observation time point and the standard time point; β is the fluctuation threshold (tolerance band parameter), for example, set to 50 days, representing the maximum allowable phenological shift; α is the penalty slope, controlling the steepness of the penalty increase. When the time difference is less than β, the function value is close to 0, assigning a very low penalty weight, that is, allowing crops to experience phenological shifts within a reasonable time window; when the time difference exceeds β, the function value increases exponentially, rapidly applying a high penalty, forcibly distinguishing planting patterns in different seasons. This "tolerance band + penalty mechanism" design achieves flexible distinction of asynchronous phenological phenomena, avoids errors caused by rigid matching, and significantly improves the model's ability to identify crop rotation patterns.

[0044] Finally, for the output and evaluation of the classification results, this embodiment uses the normalized centroid method. The defuzzification process converts fuzzy membership degrees into hard classification results and outputs a classification uncertainty evaluation map using the complementary values ​​of fuzzy membership degrees.

[0045] Specifically, after constructing the fuzzy feature space and applying nonlinear time weighting constraints, the model uses a dynamic programming algorithm to find the optimal matching path that minimizes the accumulated fuzzy cost, obtaining the fuzzy membership degree set of each category to which the pixel belongs. To output the final classification result, this embodiment uses the normalized centroid method for defuzzification. The formula for normalizing the membership degree of each candidate category is: μc = μc / Σμc. The system selects the category with the highest membership degree as the final classification attribute (hard classification result) for that pixel. Unlike traditional methods, this invention simultaneously outputs a classification uncertainty assessment map. Uncertainty is quantified using the complementary value of fuzzy membership degrees, for example, by calculating the difference between the highest and second-highest membership degree, or by directly subtracting the highest membership degree value from 1. If a pixel has a low highest membership degree (e.g., 0.5), it indicates that it is located at the intersection of multiple categories in the feature space, resulting in low classification confidence and high uncertainty. This feature allows users to intuitively identify areas where classification results are unreliable (such as plot edges and mixed pixel areas), providing clear guidance for subsequent manual verification and refined management, and solving the problem of traditional methods that "only provide results without confidence levels".

[0046] As an embodiment 5 of the present invention, such as Figure 7 As shown, this embodiment provides a system for extracting arable land planting attributes. This system is used to execute the methods described in any one of embodiments 1 to 4 above, and through a modular architecture design, achieves fully automated processing from data acquisition to classification output. The system specifically includes:

[0047] The image acquisition and preprocessing module is used to acquire and preprocess Sentinel-1 radar and Sentinel-2 optical remote sensing images of the area to be measured based on the GEE platform.

[0048] Specifically, this module serves as the data entry point for the entire system. It integrates the API interface of the GEE (Google Earth Engine) cloud platform, enabling it to automatically retrieve and download Sentinel-1 and Sentinel-2 imagery data covering the area to be measured. After acquiring the raw imagery, the module automatically triggers a preprocessing workflow: for Sentinel-1 radar imagery, it performs thermal noise removal, radiometric calibration, and terrain correction, outputting VV and VH dual-polarization backscattering coefficients; for Sentinel-2 optical imagery, it uses the QA60 band for cloud and shadow identification and applies interpolation and filtering algorithms to fill temporal gaps. The module's output is a clean, time-continuous image dataset that has undergone radiometric and geometric correction, providing a high-quality data source for subsequent feature extraction.

[0049] The preliminary sample and initial standard curve construction module is used to construct a multi-source crop distribution dataset with unified spatial coordinates and classification system. After spatial consistency test, low-confidence edge heterogeneous pixels are removed to generate preliminary crop samples and construct initial standard crop phenological curves.

[0050] Specifically, this module addresses the "cold start" problem for samples. It has pre-built interfaces for reading multiple publicly available datasets such as CCD-Maize and Chinarapeseed maps, enabling automatic coordinate transformation and attribute unification. The module incorporates a spatial consistency check algorithm that analyzes the homogeneity of pixel neighborhoods through a sliding window, automatically removing heterogeneous pixels at the edges. Finally, the module outputs a cleaned preliminary sample set and its corresponding initial standard crop phenological curves, providing a benchmark for subsequent sample expansion.

[0051] The DTW sample automatic amplification module is used to construct the temporal feature curve of the pixel to be measured using preprocessed radar and optical remote sensing images, and to calculate the DTWmin distance by matching the temporal feature curve of the pixel to be measured with the initial standard crop phenological curve using the DTW algorithm. The top 20% of pixels in terms of DTWmin distance are selected as candidate samples, and spatial constraints are implemented to eliminate non-cultivated land pseudo samples in combination with the high-resolution cultivated land base map. After manual visual cleaning, high-confidence samples are obtained.

[0052] Specifically, this module is the core of the system's fully automated sample construction. It receives image data from the preprocessing module and a baseline curve from the preliminary sample module, and calculates the DTWmin distance between the pixel to be tested and the baseline curve. The module has built-in sorting and filtering logic, automatically extracting the top 20% of pixels with the highest similarity, and uses a high-resolution farmland base map for spatial masking operations to eliminate non-farmland pseudo-samples. The module's output is a large-scale, high-confidence sample set, significantly reducing the cost of manual annotation.

[0053] The spatiotemporal feature matrix reconstruction module is used to divide the whole year into 12 standard time steps based on high-confidence samples using a monthly time discretization strategy on the GEE platform. Each monthly window uses a median synthesis algorithm to suppress noise, and triggers a radar priority complementarity mechanism to fill the time gap when optical observation fails. Standardized crop phenological reference curves are extracted through spatial aggregation statistics and spatial low-pass filtering, and Logistic nonlinear time weights are introduced to complete pixel-level dynamic phenological alignment, reconstructing an 8×12-dimensional spatiotemporal feature matrix of multi-source feature fusion.

[0054] Specifically, this module is responsible for transforming raw image data into standardized model input. It integrates temporal discretization and feature synthesis algorithms, enabling the normalization of irregular time-series data into 12 standard time steps. The module incorporates a radar-priority complementarity mechanism, automatically switching to radar data to fill in missing optical data. Ultimately, the module outputs an 8×12-dimensional spatiotemporal feature matrix containing multi-dimensional features such as biomass, moisture, and structure, providing standardized input data for the classification model.

[0055] The Fuzzy-TWDTW classification and uncertainty output module is used to construct the Fuzzy-TWDTW model. Taking an 8×12-dimensional spatiotemporal feature matrix as input, it sequentially constructs a fuzzy feature space through Gaussian fuzzy membership, applies Logistic nonlinear time weighting constraints, and performs differential weighted fusion of optical and radar multi-source features. The optimal matching path for accumulated fuzzy cost is solved through dynamic programming. The normalized centroid method is used to complete the defuzzification process, and the plot-level cultivated land planting attribute classification map and classification uncertainty assessment map are output simultaneously.

[0056] Specifically, this module serves as the system's decision-making hub. Internally, it encapsulates a Gaussian fuzzy membership function, a Logistic time weighting function, and a dynamic programming solver. The module receives the spatiotemporal feature matrix, performs elastic matching calculations in the fuzzy feature space, and outputs a set of fuzzy membership degrees. Using the normalized centroid method, the module transforms the fuzzy membership degrees into hard classification results, while simultaneously calculating the difference between the maximum and second-largest membership degrees to generate a classification uncertainty assessment map. The module's output is the final classification map and uncertainty map of cultivated land planting attributes, providing users with intuitive decision support.

[0057] Through the collaborative work of the five modules described above, this embodiment constructs a complete system for extracting arable land planting attributes. This system not only automates the entire process from data acquisition to classification output, but also reduces system coupling through modular design, facilitating subsequent maintenance and upgrades. The close connection between the data flow relationships of each module ensures the stability of system operation and the reliability of classification results.

[0058] As Embodiment 6 of the present invention, in order to verify the practical application effect of the method and system for extracting arable land planting attributes provided by the present invention, this embodiment selects the middle and lower reaches of the Yangtze River as a specific application scenario for detailed description. This region has typical cloudy and rainy climate characteristics, high degree of land fragmentation, and planting systems including various complex patterns such as single-season rice, double-season rice, rice-wheat rotation, and rice-oilseed rotation, making it an ideal test field for verifying the robustness and accuracy of the arable land planting attribute extraction method.

[0059] Regarding the optimization of model parameters, this embodiment uses the controlled variable method to optimize the key parameters in the Fuzzy-TWDTW model across the entire domain. Specifically, for the parameters in the Logistic nonlinear time-weighted constraint, after repeated experiments, the time penalty slope α is preferably set to -0.1, and the time window parameter β is preferably 50 days. The physical meaning of this parameter combination is that it allows the crop phenological period to shift within a tolerance range of 50 days without imposing a high penalty, which perfectly covers the maximum growth cycle deviation of crops caused by differences in sowing time in the study area; once the shift exceeds 50 days, the penalty weight increases exponentially, thereby effectively distinguishing planting patterns in different seasons. For the spectral ambiguity dimension, the ambiguity index is set to 1.5, and the feature bandwidth is set to 0.8. This setting achieves the optimal balance between solving the problem of spectral mixing of mixed pixels and preserving the spectral separability of similar crops. Furthermore, the weights of the multi-source features are allocated based on the importance metric calculated by the random forest algorithm: NDVI and EVI, as core biomass indicators, have the highest weighting of approximately 42%; LSWI and NDWI, as water features, account for approximately 24%; VH and VV radar features account for approximately 18%; and MSAVI and NDRE account for approximately 16%. The above parameter optimization process demonstrates the adjustability and adaptability of the technical solution of this invention in practical applications.

[0060] Table 1 presents the accuracy assessment of the mapping results based on ground reference points for crop patterns. This table uses independent ground-measured reference points as the true values ​​to conduct standardized accuracy verification of the generated crop pattern mapping results. It is used to clarify the degree of agreement between the mapping results and the actual crop distribution on the ground, quantitatively assess the overall accuracy and reliability of the mapping results in this paper, and prove that the research results are supported by solid measured data.

[0061] Table 1. Accuracy assessment of mapping results based on crop model ground reference points.

[0062] To quantitatively evaluate the actual technical effect of this invention, this embodiment compares the Fuzzy-TWDTW model with traditional random forest and standard time-weighted dynamic time warping algorithms based on the same validation dataset. Multiple sample area experiments show that in the study areas such as the Jianghan Plain, Poyang Lake Plain, and Taihu Plain, the overall classification accuracy of the method of this invention is significantly better than the two traditional algorithms. Table 2 summarizes the overall accuracy (OA) of the three methods in different years, which can quantitatively reflect the impact of interannual phenological differences, climate fluctuations, and changes in planting structure on the model's classification performance, and is used to compare the long-term crop extraction capabilities and temporal robustness of different algorithms. The results show that the method of this invention maintains higher overall accuracy in all years, with smaller interannual accuracy fluctuations, and exhibits superior stability and generalization performance in long-term crop identification tasks.

[0063] Table 2. Statistical results of the overall accuracy (OA) of the three methods in different years.

[0064] As shown in Table 2, taking 2023 data as an example, in the Jianghan Plain sample area, the overall accuracy of Random Forest was 85.90%, Standard TWDTW was 90.50%, while the Fuzzy-TWDTW of this invention reached 94.50%. Especially in challenging areas with fragmented terrain and severe pixel mixing, the accuracy of traditional machine learning methods significantly declined, while the model of this invention maintained extremely strong resistance to fragmentation interference. The comparison results under different terrain features are shown in the following table: Table 3 compares the single-category extraction accuracy of each method for different crops, verifying the fine classification advantage of the method of this invention; Table 4 examines the overall performance and generalization ability of each method across regions. Both sets of experimental results show that the extraction performance of the method proposed in this invention is significantly better than the other two comparative methods.

[0065] Table 3. F1-Scores of the three methods in rice, corn, wheat, rapeseed, and other land types.

[0066] Table 4. Overall accuracy (OA) and Kappa coefficient of the three methods in different regions.

[0067] Specifically, regarding crop type identification accuracy, this invention excels in distinguishing between winter wheat and rapeseed, which have overlapping sowing periods and highly similar spectra. Experimental data shows that the F1-Score extracted using the method of this invention reaches 0.924 for winter wheat and 0.918 for rapeseed, representing improvements of 5.3% and 5.4% respectively compared to the Standard TWDTW model. This significant improvement is mainly attributed to the Gaussian fuzzy membership degree and Logistic nonlinear time weighting mechanism introduced in this invention. The Gaussian fuzzy membership degree transforms hard distance matching into soft membership degree mapping, effectively tolerating interference from mixed pixels and spectral variations in crop phenology; while the Logistic nonlinear time weighting achieves flexible differentiation of asynchronous phenological phenomena, avoiding misclassification caused by rigid matching. In addition, this embodiment also simultaneously outputs a classification uncertainty assessment map, quantifying the reliability of the identification results through complementary fuzzy membership degree values, providing intuitive guidance for subsequent manual verification. The above application examples fully demonstrate the robustness and high accuracy of this invention in complex agricultural mapping scenarios, effectively meeting the needs of refined farmland monitoring in areas with frequent cloud cover and fragmented plots.

[0068] As Embodiment 7 of the present invention, in order to verify the adaptability of the present invention under different climatic and topographical conditions, this embodiment selects the high-altitude agricultural area of ​​the Qinghai-Tibet Plateau as an alternative application scenario for detailed description. This region has high altitude, low temperature, is shrouded in clouds and fog all year round, and has a long snow cover period, making it extremely difficult to acquire optical remote sensing images. It is an ideal test field to verify the robustness of the present invention under extreme data shortage conditions.

[0069] In response to the severe lack of optical data in high-altitude and cold regions, this embodiment has made adaptive adjustments to the data source strategy. Specifically, in step S1, this embodiment primarily relies on Sentinel-1 radar imagery as the core data source, with Sentinel-2 optical imagery serving only as an auxiliary data source. This is because high-altitude and cold regions frequently experience cloud and fog, resulting in extremely short effective observation windows for optical imagery and data gaps in many months. Radar imagery possesses all-weather, all-time imaging capabilities, capable of penetrating clouds and fog to capture crop structure information, making it the only reliable data source for monitoring cultivated land in high-altitude and cold regions. In the preprocessing stage, this embodiment has particularly strengthened terrain correction processing for radar imagery, employing higher-precision DEM data (such as ALOS PALSAR DEM) to eliminate the distortion effects of complex terrain in high-altitude areas on backscattered signals.

[0070] In the sample selection step S2, this embodiment adaptively adjusts the sample selection threshold to suit the relatively simple terrain and large plot size of high-altitude and cold regions. Specifically, high-altitude agricultural areas are mostly located in river valleys or gentle slopes of plateaus, with high contiguous farmland and relatively weak heterogeneity at plot edges. Therefore, this embodiment relaxes the DTWmin distance ranking threshold from 20% in the middle and lower reaches of the Yangtze River to 30%, selecting the top 30% of pixels with the highest similarity as candidate samples. This adjustment is based on the following considerations: the planting system in high-altitude and cold regions is relatively simple, mainly consisting of single-season crops such as highland barley, rapeseed, and spring wheat, with simple crop rotation patterns and naturally high sample purity; at the same time, the large plot size and weak edge effect mean that relaxing the selection threshold can effectively increase the number of samples and compensate for the information loss caused by the lack of optical data. Experimental verification shows that in the Qinghai-Tibet Plateau sample area, the purity of the sample set selected using the 30% threshold remains above 92%, meeting the requirements for model training.

[0071] Regarding the feature fusion strategy in step S4, this embodiment performs weight reconstruction for scenarios dominated by radar data. Specifically, since optical imagery becomes invalid in most months, this embodiment constructs a 4×12-dimensional spatiotemporal feature matrix with radar features as the core, including VV polarization backscattering coefficient, VH polarization backscattering coefficient, VV / VH polarization ratio, and radar texture features (such as the contrast and homogeneity of the gray-level co-occurrence matrix). In terms of feature weight allocation, the radar feature weight is increased to approximately 70%, while the optical feature weight (NDVI, EVI, etc.) is reduced to approximately 30%, and it only participates in the calculation during the effective observation months. In particular, this embodiment introduces radar polarization decomposition features, decomposing VV and VH polarization data into surface scattering, dihedral scattering, and volume scattering components to more finely characterize the structural features of alpine crops. Experimental results show that in the Qinghai-Tibet Plateau sample area, the overall accuracy of the radar-dominated feature fusion strategy reaches 88.5%, which is about 12 percentage points higher than the scheme using only optical features, fully demonstrating the adaptability of this invention under extreme data shortage conditions.

[0072] Furthermore, considering the short phenological periods and concentrated growing seasons of crops in high-altitude and cold regions, this embodiment adjusts the nonlinear time-weighted parameters of the Logistic algorithm. Specifically, the time window parameter β is shortened from 50 days in the middle and lower reaches of the Yangtze River to 30 days to adapt to the short growth cycle and small phenological shifts of high-altitude and cold-region crops. Simultaneously, the time penalty slope α is adjusted to -0.15 to enhance sensitivity to phenological shifts and avoid mismatching crops at different growth stages. These parameter adjustments enable the model to more accurately capture the rapid growth process of high-altitude and cold-region crops, improving classification accuracy.

[0073] As Embodiment 8 of the present invention, in order to further verify the performance limit of the present invention under ideal observation conditions, this embodiment selects the contiguous cultivated area of ​​the Northeast Plain as another alternative application scenario for detailed description. This area has flat terrain, regular plots, and a high degree of contiguousness. The planting system is mainly single-season crops such as corn, soybeans, and rice. The conditions for acquiring optical images are good, making it an ideal test field to verify the classification accuracy and efficiency of the present invention under conditions of sufficient data and simple terrain.

[0074] Taking advantage of the superior conditions for acquiring optical data in contiguous farmland areas on the plains, this embodiment optimizes and adjusts the data source strategy. Specifically, in step S1, this embodiment uses Sentinel-2 optical imagery as the core data source and Sentinel-1 radar imagery as a supplementary data source. This is because there is relatively little cloud and rain in the plains, resulting in a long effective observation window for optical imagery, which can provide rich spectral information for crop identification. Radar imagery is only used as a supplement in months when optical imagery is occasionally missing, to ensure the integrity of the timeline. In the preprocessing stage, for optical imagery, this embodiment strengthens atmospheric correction processing and employs a more refined aerosol inversion algorithm to improve the accuracy of the spectral signal.

[0075] In the sample screening step S2, this embodiment adaptively adjusts the sample screening threshold to suit the characteristics of high homogeneity and simple cropping systems in the plains region. Specifically, the Northeast Plain has extremely high contiguous farmland, with plots generally exceeding tens of hectares in size and exhibiting very weak edge heterogeneity. Therefore, this embodiment adjusts the DTWmin distance ranking threshold from 20% in the Yangtze River mid-lower reaches region to 25%, selecting the top 25% of pixels with the highest similarity as candidate samples. This adjustment is based on the following considerations: the cropping system in the plains region is mainly single-season crops with simple crop rotation patterns, resulting in naturally high sample purity; simultaneously, the large plot size and strong homogeneity mean that appropriately relaxing the screening threshold can increase sample diversity while ensuring sample quality, thereby improving the model's generalization ability. Experimental verification shows that in the Northeast Plain sample area, the purity of the sample set screened using the 25% threshold reaches over 95%, significantly higher than that of the fragmented plot areas in the Yangtze River mid-lower reaches region.

[0076] Regarding the feature fusion strategy in step S4, this embodiment performs weight reconstruction for scenarios dominated by optical data. Specifically, since optical images are valid for most months, this embodiment constructs an 8×12-dimensional spatiotemporal feature matrix with optical features as the core, including spectral features such as NDVI, EVI, MSAVI, LSWI, NDWI, and NDRE, as well as VV and VH radar features as supplements. In terms of feature weight allocation, the weight of optical features is increased to approximately 65%, while the weight of radar features is reduced to approximately 35%. In particular, this embodiment introduces red-edge features (such as NDRE and CIred-edge) as high-weight features. This is because crops in plain areas grow uniformly, and the sensitivity of red-edge features to chlorophyll content and leaf structure can more finely distinguish different crop types. Experimental results show that in the Northeast Plain sample area, the overall accuracy of the optical-dominated feature fusion strategy reaches 96.2%, which is about 3 percentage points higher than the radar-dominated strategy, fully demonstrating the high-accuracy potential of this invention under conditions of sufficient data.

[0077] Furthermore, considering the relatively concentrated sowing period and small phenological shifts in plain areas, this embodiment adjusts the nonlinear time-weighted parameters of the Logistic algorithm. Specifically, the time window parameter β is shortened from 50 days in the middle and lower reaches of the Yangtze River to 35 days to adapt to the smaller differences in crop sowing periods in plain areas; simultaneously, the time penalty slope α is adjusted to -0.12 to moderately reduce sensitivity to phenological shifts and avoid imposing excessive penalties on normal sowing period differences. These parameter adjustments enable the model to adapt more flexibly to the planting characteristics of plain areas, improving classification accuracy.

[0078] A comparative analysis of Examples 7 and 8 clearly demonstrates the strong adaptability of the technical solution of this invention under different geographical environments and data conditions. In high-altitude and cold regions, effective classification is achieved under extreme data shortage conditions through strategies such as radar data dominance, relaxed sample selection thresholds, and feature weight reconstruction. In contiguous farmland areas on plains, high-precision classification is achieved under ideal observation conditions through strategies such as optical data dominance, appropriate adjustment of sample selection thresholds, and feature weight optimization. These two alternative embodiments fully demonstrate that the technical solution of this invention has broad applicability and tunability, and can meet the needs of farmland planting attribute extraction under different regions, terrains, and data conditions.

[0079] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A method for extracting planting attributes of arable land, characterized in that, Includes the following steps: S1: Preprocess the radar and optical remote sensing data of the area to be measured; S2: Construct a multi-source public crop distribution dataset with unified spatial coordinates and classification system. Through spatial consistency test, remove heterogeneous pixels at the edge of plots or with low confidence, generate preliminary crop samples, and construct initial standard crop phenological curves. S3: Construct the time series feature curve of the pixel to be tested using the preprocessed radar and optical remote sensing data. Use the DTW algorithm to match and calculate the minimum distance between the time series feature curve of the pixel to be tested and the initial standard crop phenology curve. Select the pixels to be tested with the largest minimum distance as candidate samples. Combine with the high-resolution cultivated land base map to implement spatial constraints and eliminate non-cultivated land pseudo samples. After manual visual cleaning, obtain high confidence samples. S4: Based on high-confidence samples, a time discretization strategy is adopted to divide the time step of high-confidence samples by month. The median synthesis algorithm is used to suppress noise in each monthly window. When optical observation fails, the radar priority complementarity mechanism is triggered to fill the time gap. The standardized crop phenological reference curve is extracted by spatial aggregation function and spatial low-pass filtering. Logistic nonlinear time weight is introduced to complete the pixel-level dynamic phenological alignment and reconstruct the spatiotemporal feature matrix of multi-source feature fusion. S5. Construct a Fuzzy-TWDTW model. Using the spatiotemporal feature matrix as input, construct the fuzzy feature space through Gaussian fuzzy membership, apply Logistic nonlinear time weighting constraints, and carry out differential weighted fusion of optical and radar multi-source features. Solve the optimal matching path for accumulated fuzzy cost through dynamic programming. Use the normalized centroid method to complete the defuzzification process and simultaneously output the plot-level cultivated land planting attribute classification map and classification uncertainty assessment map.

2. The method for extracting arable land planting attributes according to claim 1, characterized in that, The preprocessing includes: The radar remote sensing images were sequentially subjected to thermal noise removal, radiometric calibration and terrain correction, and the VV and VH dual-polarization backscattering coefficients were extracted. Clouds and shadows were removed from optical remote sensing images using the QA60 band, and interpolation and filtering algorithms were used to fill in time gaps and smooth the images.

3. The method for extracting arable land planting attributes according to claim 1, characterized in that, The multi-source crop distribution datasets include CCD-Maize, China rapeseed maps, China Rice, China Wheat, ChinaCP, and ChinaCP-Wheat10m.

4. The method for extracting arable land planting attributes according to claim 1, characterized in that, The high-confidence sample covers 11 land cover types: fallow, 4 single-season planting types, 5 crop rotation patterns, and other multiple cropping types.

5. The method for extracting arable land planting attributes according to claim 1, characterized in that, The multi-source characteristics include: biomass characteristics: NDVI, EVI, MSAVI; water characteristics: LSWI, NDWI; pigment and structural characteristics: NDRE, VV, VH.

6. The method for extracting arable land planting attributes according to claim 1, characterized in that, The Gaussian fuzzy membership degree is used to convert temporal hard distance matching into soft membership degree mapping, reducing the interference of mixed pixels and crop phenological spectral variations on classification accuracy.

7. The method for extracting arable land planting attributes according to claim 1, characterized in that, The Logistic nonlinear time-weighted constraint uses the Logistic function, which assigns a low penalty within a set fluctuation threshold, and the penalty increases exponentially when the fluctuation threshold is exceeded, thereby achieving asynchronous and flexible differentiation of phenology.

8. The method for extracting arable land planting attributes according to claim 1, characterized in that, The defuzzification process converts fuzzy membership degrees into hard classification results and outputs a classification uncertainty assessment graph using complementary values ​​of fuzzy membership degrees.

9. A system for extracting arable land planting attributes, characterized in that, include: The image preprocessing module is used to preprocess the acquired radar remote sensing data and optical remote sensing data; The module for constructing preliminary samples and initial standard curves is used to build a multi-source crop distribution dataset with unified spatial coordinates and classification system. After spatial consistency testing, low-confidence marginal heterogeneous pixels are removed to generate preliminary crop samples and construct initial standard crop phenological curves. The DTW sample automatic amplification module constructs the time-series feature curve of the pixel to be measured using preprocessed radar remote sensing data and optical remote sensing data. The DTW algorithm is used to match the time-series feature curve of the pixel to be measured with the initial standard crop phenology curve to calculate the minimum distance. Pixels close to the minimum distance are selected as candidate samples. Spatial constraints are implemented by combining high-resolution cultivated land base map to eliminate non-cultivated land pseudo samples. After manual visual cleaning, high-confidence samples are obtained. The spatiotemporal feature matrix reconstruction module, based on high-confidence samples, adopts a time discretization strategy to divide the time steps of high-confidence samples by month. Each monthly window uses a median synthesis algorithm to suppress noise. When optical observation fails, a radar priority complementarity mechanism is triggered to fill the time gap. Standardized crop phenological reference curves are extracted through spatial aggregation statistics and spatial low-pass filtering. Logistic nonlinear time weights are introduced to complete pixel-level dynamic phenological alignment and reconstruct the spatiotemporal feature matrix of multi-source feature fusion. The Fuzzy-TWDTW classification and uncertainty output module constructs a Fuzzy-TWDTW model. Taking the spatiotemporal feature matrix as input, it sequentially constructs a fuzzy feature space through Gaussian fuzzy membership, applies Logistic nonlinear time weighting constraints, and performs differential weighted fusion of optical and radar multi-source features. The optimal matching path for cumulative fuzzy cost is solved through dynamic programming. The normalized centroid method is used to complete the defuzzification process, and the module simultaneously outputs a plot-level farmland planting attribute classification map and a classification uncertainty assessment map.

Images