A method and apparatus for corn remote sensing mapping

Abstract

This invention discloses a remote sensing mapping method and device for maize. It constructs a multi-dimensional temporal spectral feature set by acquiring remote sensing image data of the crop growing season in the target area, and uses this set to identify key time nodes during the tasseling stage of maize. Based on the temporal spectral features of these key time nodes, a temporal irrigation index is constructed. The irrigation index is then fused with the multi-dimensional temporal spectral feature set to construct a high-dimensional enhanced feature set. Finally, the high-dimensional enhanced feature set is used to train and classify a machine learning classifier, outputting a spatial distribution map of maize in the target area. By introducing a temporal irrigation index that characterizes the temporal changes before and after the tasseling stage of maize, the method can better represent the differences between maize and other crops at key physiological stages, enhance the discriminative power of classification features, reduce classification confusion between easily confused crops, and thus improve the extraction effect of the spatial distribution map of maize in the target area.

Description

Technical Field

[0001] This invention belongs to the field of agricultural remote sensing and information technology, and more specifically, relates to a method and equipment for remote sensing mapping of maize. Background Technology

[0002] Maize is one of the world's most important food crops, with a wide planting area and high yield. Accurately obtaining spatial distribution information of maize planting areas is of great significance for food security assessment, agricultural resource management, yield forecasting, and agricultural policy formulation. Especially in typical agricultural planting areas such as the US Corn Belt, maize and easily confused crops (such as soybeans) are often widely planted in rotation. Both crops show high vegetation cover and strong biomass accumulation in the middle and late stages of the growing season. Their remote sensing images have high similarity in spectral responses in the visible and near-infrared bands, making it difficult to stably distinguish between the two crops when relying solely on conventional spectral features or general temporal features. This has become an important factor restricting the improvement of maize remote sensing classification accuracy.

[0003] In existing technologies, classification methods based on time-series remote sensing features throughout the entire growing season are commonly used to improve the accuracy of maize identification. Although this method can improve identification capabilities to some extent by utilizing phenological differences within the crop growth cycle, it usually relies on continuous observation data throughout the entire growing season. However, in practical applications, remote sensing time-series data often suffers from missing data, discontinuity, or high noise levels due to factors such as cloud and rain weather, shading, and observation cycle limitations, which affects the stability of subsequent classification results.

[0004] The prior art patent CN120219975A proposes a phenologically adaptive automatic mapping method for maize, including the following steps: selecting and acquiring Sentinel-2 time-series satellite image data of the target area and preprocessing it to synthesize a semi-monthly clear-sky time-series image dataset; filtering the clear-sky time-series image dataset to obtain a denoised and smoothed time-series satellite image dataset; adaptively acquiring key phenological periods for maize identification in the target area based on the NDVI index peak and valley values ​​of the filtered images; analyzing the typical spectral differences between maize and other major background features, proposing a corrected red-edge normalized vegetation index, and comparing it with the NDVI index... The VI index and LSWI index constitute a typical index dataset for identifying key phenological stages of maize. By utilizing the division rules of typical index data for key phenological stages of maize and other major background land cover, rapid mapping of maize in the target area can be automatically achieved. This invention mainly focuses on feature extraction and classification of key phenological stages. The identification of key phenological stages still relies on relatively complete time-series observation data, and is therefore more suitable for maize mapping under mid-to-late growing season or full growing season conditions. In early monitoring scenarios, since the key phenological nodes of crops have not yet fully emerged, the temporal and spectral differences between maize and crops such as soybeans are not prominent enough, resulting in limited early identification capabilities. Summary of the Invention

[0005] This invention aims to overcome the problems of insufficient utilization of the dynamic characteristics of the maize growing season and weak ability to distinguish between maize and easily confused crops such as soybeans in existing maize remote sensing mapping methods, and provides a maize remote sensing mapping method and equipment.

[0006] The primary objective of this invention is to solve the aforementioned technical problems. The technical solution of this invention is as follows: The first aspect of this invention provides a method for remote sensing mapping of maize, comprising the following steps: Acquire remote sensing image data of the target area during the crop growing season, calculate the normalized vegetation index and land surface water index, and construct a multidimensional temporal spectral feature set. The key time nodes of maize tasseling stage are identified using the multidimensional time-series spectral feature set, and a time-series irrigation index is constructed using the spectral features of the time window corresponding to the key time node and the previous equal-length time window. The time-series irrigation index is fused with a multi-dimensional time-series spectral feature set to construct a high-dimensional enhanced feature set. The high-dimensional enhanced feature set is used to train and classify a machine learning classifier, and the spatial distribution map of corn in the target area is output.

[0007] Furthermore, the method for constructing a multidimensional time-series spectral feature set includes the following steps: Acquire multi-band remote sensing image data of the target area during the crop growing season; Scene classification maps are used to preprocess multi-band remote sensing image data, images with cloud cover exceeding a preset threshold are removed, and cloud, shadow and thin cloud areas are masked to obtain effective observation image data. Define the growing season time range of crops in the target area, and divide the time range into several consecutive equal-length windows. Calculate the multi-temporal spectral features corresponding to the effective observation image data in each time window. The multi-temporal spectral features include normalized vegetation index, land surface water index, and near-infrared reflectance. The median composite method was used to perform time-series composite processing on the effective observation image data within each time window to obtain cloudless continuous time-series images. The cloudless continuous time series images are stacked in chronological order, and the multi-band remote sensing image data and multi-temporal spectral features corresponding to each time window are combined to construct a multi-dimensional time series spectral feature set.

[0008] Furthermore, the multi-band remote sensing image data is multi-temporal Sentinel-2 surface reflectance data, including red band data, near-infrared band data, and short-wave infrared band data.

[0009] Furthermore, the method for constructing the time-series irrigation index includes the following steps: Obtain the normalized vegetation index time series data of the multidimensional time-series spectral feature set, and perform pixel-by-pixel traversal of the normalized vegetation index time series data to construct the normalized vegetation index growth curve of each pixel during the crop growing season. The effective peak time of the normalized vegetation index growth curve was determined by the dynamic peak detection algorithm, and the date corresponding to the effective peak time was determined as the key time node t2 of the maize tasseling stage. Centered on t2, a dynamic observation window T2 is constructed using a preset time span. A reference node t1 is determined by tracing back a preset time interval from t2. A dynamic reference window T1 of the same length as t2 is constructed with t1 as the center. The preset time interval is greater than or equal to the preset time span. Obtain the time series sequence of land surface moisture index for each pixel in the multidimensional time-series spectral feature set within T1 and T2, denoted as follows: and And calculate the median land surface moisture index within T1 respectively. and the median land surface moisture index within T2 ; Using the and The time-series irrigation index (TSII) for each pixel is obtained by performing interpolation calculation, as shown in the following expression:

[0010] Furthermore, the effective peak time of the normalized vegetation index growth curve is determined using a dynamic peak detection algorithm, including the following steps: The normalized vegetation index growth curve was smoothed using the n-point sliding window averaging method to obtain the smoothed normalized vegetation index growth curve. The smoothed normalized vegetation index growth curve is filtered for invalid nodes using a preset vegetation determination threshold, and the effective normalized vegetation index data is output. The global maximum value of the effective normalized vegetation index data is then extracted. Extract nodes that are greater than or equal to a preset proportion of the global maximum value as a candidate peak set; The candidate peak set is sorted in chronological order, and the time node corresponding to the earliest appearing candidate peak is selected as the effective peak time of the corresponding pixel.

[0011] Furthermore, the time-series irrigation index is fused with a multi-dimensional time-series spectral feature set, including the following steps: Obtain the multidimensional temporal spectral feature set corresponding to each pixel in the target area, and represent the multidimensional temporal spectral feature set as a multidimensional feature layer arranged in time order; Obtain the time-series irrigation index corresponding to each pixel, and use the time-series irrigation index as a new feature layer; The newly added feature layer and the multi-dimensional feature layer are band-stitched together to obtain the high-dimensional enhanced feature vector corresponding to each pixel. The high-dimensional enhanced feature vectors corresponding to each pixel in the target area are reorganized according to their spatial location to construct a high-dimensional enhanced feature set.

[0012] Furthermore, the machine learning classifier is trained and classified using a high-dimensional enhanced feature set, including the following steps: Based on the existing arable land dataset, we used stratified random sampling to extract sample points of corn and easily confused crops, and obtained the real crop labels corresponding to each sample point. Extract the high-dimensional enhanced feature vectors corresponding to each sample point in the high-dimensional enhanced feature set, and construct a training sample set containing the real crop label and the high-dimensional enhanced feature vectors; Build a machine learning classifier, input the training sample set into the machine learning classifier for training, and obtain the trained crop classification model; Input the high-dimensional enhanced feature set corresponding to each pixel in the target area into the trained crop classification model, and output a preliminary spatial distribution map of crops in the target area; The preliminary spatial distribution map is filtered to extract the spatial distribution results corresponding to corn, and the final corn spatial distribution map is output.

[0013] Furthermore, the machine learning classifier is a random forest classification model or a support vector machine model.

[0014] Furthermore, the filtering process is a mode filtering process, which includes the following steps: By using a preset sliding window to traverse the spatial distribution map pixel by pixel, the frequency of occurrence of each crop category in each sliding window is counted, and the crop category with the highest frequency is determined as the mode category of the corresponding sliding window. Replace the crop category of the center pixel of each sliding window with the corresponding mode category to obtain the filtered classification result image; The filtered classification result map is subjected to mode filtering again until the preset number of iterations is reached, and the final crop spatial distribution map is output.

[0015] A second aspect of the present invention provides a maize remote sensing classification electronic device, including a memory and a processor. The memory includes a maize remote sensing classification method program, which, when executed by the processor, implements the steps of a maize remote sensing mapping method.

[0016] Compared with the prior art, the beneficial effects of the technical solution of the present invention are: This invention constructs a multi-dimensional temporal spectral feature set by acquiring remote sensing image data of crops during the growing season in the target area. This feature set is then used to identify key time nodes during the tasseling stage of maize. Based on the temporal spectral features of these key time nodes, a temporal irrigation index is constructed. This irrigation index is then fused with the multi-dimensional temporal spectral feature set to construct a high-dimensional enhanced feature set. Finally, this high-dimensional enhanced feature set is used to train and classify a machine learning classifier, outputting a spatial distribution map of crops in the target area. Compared to classification methods that rely solely on conventional temporal spectral features, this invention, by introducing a temporal irrigation index that characterizes the temporal changes before and after the tasseling stage of maize, can better represent the differences between maize and other crops at key physiological stages, enhance the discriminative power of classification features, reduce classification confusion between easily confused crops, and thus improve the extraction effect of the spatial distribution map of crops in the target area. Attached Figure Description

[0017] To make the objectives and technical solutions of this invention clearer, the following drawings are provided and described: Figure 1 A flowchart of a corn remote sensing mapping method provided in an embodiment of the present invention; Figure 2 A technical roadmap provided for embodiments of the present invention; Figure 3 This is a schematic diagram illustrating the calculation principle and timing window of the irrigation index during the tasseling period provided in an embodiment of the present invention; Figure 4 A comparison of corn-soybean classification results provided in embodiments of the present invention. Detailed Implementation

[0018] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0019] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0020] Example 1: This invention provides a remote sensing mapping method for maize, focusing on the key phenological stage unique to maize—the tasseling stage—based on crop physiological mechanisms. During this stage, maize is extremely sensitive to water stress. To ensure yield, irrigation is often implemented during this period, leading to a significant short-term increase in canopy moisture content. In contrast, soybeans exhibit a relatively mild water response during the same period. This difference in water signals, driven by management measures and manifested at a key phenological node, provides a stable and physically meaningful source of characteristics for remote sensing to distinguish between the two crops. Dynamically capturing and quantifying this characteristic through remote sensing is an effective new approach to solving the spectral confusion problem between maize and soybeans and achieving accurate early mapping.

[0021] like Figure 1 The diagram shown is a flowchart of a remote sensing mapping method for maize. Figure 2 As a technical roadmap, this invention primarily relies on the Google Earth Engine (GEE) cloud computing platform for data processing and computation. This embodiment selects the northern region of Iowa, the core area of ​​the US Corn Belt, as the specific verification area. The central coordinates of this region range from approximately 42.91°N to 43.08°N and 93.74°W to 94.08°W, covering an area of ​​approximately 15.2 km², and it is characterized by flat terrain and typical corn-soybean rotation.

[0022] The specific steps are as follows: S1: Acquire remote sensing image data of the target area during the crop growing season, calculate the normalized vegetation index and land surface moisture index, and construct a multidimensional time-series spectral feature set.

[0023] In order to construct a high-quality temporal feature set that can accurately depict the crop growth process, this study acquired multi-temporal Sentinel-2 remote sensing images of the target area during the crop growing season and completed the entire process from data cleaning to temporal synthesis based on the GEE platform.

[0024] The specific process is as follows: S1.1: Acquire multi-band Sentinel-2 remote sensing image data of the target area during the early growing season of crops, selecting images that include red band, near-infrared band and short-wave infrared band; S1.2: Use Scene Classification Map (SCL) to preprocess multi-band remote sensing image data, remove images with cloud cover exceeding a preset threshold (in this embodiment, the threshold is 20%), and mask cloud, shadow and thin cloud areas to obtain quality-controlled effective observation image data.

[0025] During the data acquisition and preprocessing stage, the Sentinel-2 MSI Level-2A surface reflectance (SR) dataset was used. To address the issue of optical imagery being susceptible to cloud and rain interference, the image's built-in Scene Classification Layer (SCL) was utilized for rigorous quality control. Specifically, a cloud removal masking algorithm was developed to automatically identify and remove pixels in the SCL bands marked as clouds (high / medium probability), cloud shadows, and thin clouds, retaining only clear and valid surface observations to ensure the accuracy of subsequent feature extraction.

[0026] S1.3: Define the growing season time range of crops in the target area (in this embodiment, it is from May 1 to August 4), and divide the growing season time range into several consecutive equal-length windows (in this embodiment, the time window interval is 8 days); for each valid observation image data in each time window, calculate the corresponding multi-temporal spectral features, which include the normalized difference vegetation index (NDVI), the land surface water index (LSWI), and the near-infrared reflectance (B8).

[0027] S1.4: In order to overcome the residual noise in single-scene images and generate temporally continuous data, the median composition method is used to perform time-series synthesis processing on the effective observation image data with calculated multi-temporal spectral characteristics within each time window. Using the aggregation function of the GEE platform, the median of all effective observations after masking within each time window is calculated (i.e., invalid pixels removed by masking are directly ignored during synthesis calculation), to obtain cloudless continuous time series images for each time window.

[0028] S1.5: Stack the cloudless continuous time series images in chronological order, and combine the multi-band remote sensing image data and multi-temporal spectral features corresponding to each time window to construct a multi-dimensional time series spectral feature set. This dataset integrates the time series information of NDVI, LSWI and spectral bands, laying a solid data foundation for subsequent capture of key phenological nodes.

[0029] S2: The key time nodes of the corn tasseling stage are identified using the multidimensional time-series spectral feature set, and the time-series irrigation index is constructed using the spectral features of the time window corresponding to the key time node and the spectral features of the preceding time window. It does not depend on fixed calendar dates, but rather on the biological characteristics of the crop itself for adaptive discrimination.

[0030] The specific process is as follows: S2.1: Obtain the normalized vegetation index (NDVI) time series data corresponding to the target area of ​​the multidimensional time series spectral feature set, and perform pixel-by-pixel traversal of the NDVI time series data to construct the NDVI growth curve of each pixel during the crop growing season, so as to reflect the phenological rhythm of the crop.

[0031] S2.2: The effective peak time of the NDVI growth curve is determined by the dynamic peak detection algorithm, and the date corresponding to the effective peak time is determined as the key time node t2 of the maize tasseling period.

[0032] Specifically, the tasseling stage of maize is a crucial turning point in its transition from vegetative to reproductive growth. At this time, the plant's biomass accumulation typically reaches its peak, reflected in remote sensing features where the NDVI value reaches its maximum throughout the growing season. Based on this physiological mechanism, this example uses a peak detection algorithm to process each pixel within the study area. The system automatically retrieves the NDVI growth curve and locates the moment when the value reaches its maximum. The date of the time window corresponding to this maximum value is dynamically determined as the critical time node t2 of the tasseling stage for that pixel.

[0033] Through this pixel-by-pixel dynamic recognition strategy, the system generates an independent tasseling time label for each location in the entire map. Compared with traditional phenological determination methods based on fixed dates, this method can effectively overcome the fluctuations in phenological periods between plots caused by different sowing times, soil fertility differences, or microclimate changes. It achieves accurate anchoring of the most water-sensitive physiological stage of maize, providing an accurate time benchmark for the subsequent construction of the Time Series Irrigation Index (TSII).

[0034] The specific process is as follows: S2.2.1: The NDVI growth curve is smoothed using the n-point sliding window averaging method (in this embodiment, n is 3) to obtain the smoothed NDVI growth curve. S2.2.2: The smoothed NDVI growth curve is filtered for invalid nodes using a preset vegetation determination threshold (in this embodiment, the threshold is NDVI>0.2), and valid NDVI data is output. The global maximum value of the valid NDVI data is then extracted. S2.2.3: Extract nodes that are greater than or equal to a preset percentage of the global maximum value (in this embodiment, the preset percentage is 95%) as a candidate peak set; S2.2.4: Sort the candidate peak set in chronological order to obtain the sorted candidate peak sequence, and select the time node corresponding to the earliest appearing candidate peak from the sorted candidate peak sequence as the effective peak time of the corresponding pixel (if the effective peak cannot be extracted due to missing data, a preset date such as August 1st is used as the default replacement date). The moment corresponding to this effective peak value represents the turning point when maize biomass accumulation reaches its peak and the growth shifts from vegetative to reproductive growth.

[0035] S2.3: Centered on t2, a dynamic observation window T2 is constructed using a preset time span (10 days before and after in this embodiment). A reference node t1 is determined by retrospectively determining a preset time interval (15 days in this embodiment) from t2. A dynamic reference window T1 of the same length as t2 is then constructed centered on t1 to ensure consistency in the comparison. The preset time interval is greater than or equal to the preset time span. Since the biological occurrence time of the tasseling period t2 (usually July to August) lags far behind the start time (e.g., May), this retrospective t1 window will necessarily exist effectively in the time-series data. This design ensures that the comparison time base is immediately before the tasseling event, thus enabling the sensitive capture of sudden changes in water levels caused by irrigation events.

[0036] S2.4: Obtain the LSWI time series of the land surface moisture index for each pixel in the multidimensional time-series spectral feature set within T1 and T2, denoted as follows: and And calculate the median land surface moisture index within T1 respectively. and the median land surface moisture index within T2 This represents the stable moisture state during that period.

[0037] S2.5: Utilizing the aforementioned and The time-series irrigation index (TSII) for each pixel is obtained by performing interpolation calculation, as shown in the following expression:

[0038] The time-series irrigation index is used to quantify the change in canopy or topsoil moisture before and after the tasseling critical period of maize. By the numerical difference of this time-series irrigation index, it is possible to quantitatively distinguish between maize, which shows a significant increase in moisture during the tasseling period (i.e., TSII shows a significant positive value), and soybean, which shows a relatively slow change in moisture (i.e., TSII shows a lower value), thus serving as a key discriminant feature to distinguish between the two types of crops.

[0039] In the Iowa study area, maize is extremely sensitive to water stress during the tasseling stage, and agricultural management typically involves significant irrigation or soil water conservation measures, resulting in a significantly higher LSWit2 than LSWit1, thus leading to a higher positive TSII value. In contrast, soybean water requirements change relatively smoothly during the same period, and its TSII value usually remains at a low level. This characteristic effectively amplifies the physiological differences between the two crops during their vigorous growth phase, which is subject to severe spectral confusion, and constitutes a key discriminant factor in the subsequent classification model. Figure 3 As shown in the figure, the time-series evolution curves of maize and soybean indices extracted from Sentinel-2 remote sensing data under actual growing season conditions (taking 2023 as an example) are displayed. The horizontal axis of the figure represents consecutive dates within the growing season, and the vertical axes represent the Normalized Difference Vegetation Index (NDVI) and the Land Surface Water Index (LSWI), respectively. The observation results in the figure above show that retrieving the maximum value of the NDVI curve (i.e., ndvi_max) can accurately pinpoint the key node t2 of maize's tasseling stage. Furthermore, comparing this with the magnified local features in the figure below, it can be seen that within the observation window from t1 to tasseling stage t2, because maize is extremely sensitive to water stress during the tasseling stage, agricultural management typically involves significant irrigation or soil water conservation measures, resulting in a significant positive increase in its LSWI (i.e.,...). Figure 3 The dLSWI (corn) varies greatly; in contrast, the water requirement of soybean crops changes relatively slowly during the same period, and its LSWI curve fluctuates less (i.e., Figure 3 The differences in dLSWI (soybeans) were relatively small. This result intuitively and powerfully demonstrates that the TSII constructed in this invention can effectively capture the unique water response signals of maize during key phenological stages, thereby achieving precise differentiation at the physiological mechanism level during the peak growing season when the spectra of the two crops highly overlap.

[0040] S3: The temporal irrigation index (1-dimensional feature) is fused with a multi-dimensional temporal spectral feature set (including NDVI, LSWI, and B8 band sequences, totaling 36 temporal features) at the pixel scale to form a complete 37-dimensional high-dimensional enhanced feature set. This fusion process fully preserves the original spectral and phenological physical signals, eliminating the need for additional data dimensionality reduction or weighted calculations, and can be directly used as high-dimensional feature input for subsequent machine learning algorithms.

[0041] More specifically, fusing the time-series irrigation index with a multidimensional time-series spectral feature set includes the following steps: S3.1: Obtain the multidimensional temporal spectral feature set corresponding to each pixel in the target area, and represent the multidimensional temporal spectral feature set as a multidimensional feature layer arranged in time order.

[0042] S3.2: Obtain the time-series irrigation index corresponding to each pixel, and use the time-series irrigation index as a new feature layer to characterize the water change characteristics before and after the tasseling period of maize.

[0043] S3.3: Perform band stacking on the newly added feature layer and the multidimensional feature layer to obtain the high-dimensional enhanced feature vector corresponding to each pixel.

[0044] S3.4: Reorganize the high-dimensional enhanced feature vectors corresponding to each pixel in the target area according to their spatial location to construct a high-dimensional enhanced feature set.

[0045] S4: Use the high-dimensional enhanced feature set to train and classify the machine learning classifier, and output a spatial distribution map of corn in the target area.

[0046] The specific process is as follows: S4.1: Based on existing farmland datasets (such as CDL data), use stratified random sampling to extract sample points of corn and easily confused crops, and obtain the real crop labels corresponding to each sample point.

[0047] S4.2: Extract the high-dimensional enhanced feature vector corresponding to each sample point in the high-dimensional enhanced feature set, construct a sample set containing real crop labels (in this embodiment, corn or soybean) and high-dimensional enhanced feature vectors (including 36-dimensional basic time series features and 1-dimensional TSII features), and divide it into training set and test set according to a preset ratio (in this embodiment, the preset ratio is 7:3).

[0048] S4.3: Construct a machine learning classifier and set the classification parameters of the classifier. Input the training sample set into the machine learning classifier for training. Use the training set to supervise the learning of the enhanced feature set so that the model can fully learn the weight distribution of TSII features in distinguishing the "corn-soybean" category. Establish the classification mapping relationship between crop category and high-dimensional enhanced feature vector to obtain the trained crop classification model.

[0049] More specifically, the machine learning classifier is a support vector machine model or a random forest classification model. In this embodiment, a random forest classification model is used. The hyperparameters of the random forest algorithm are set as follows: the number of decision trees is set to 100; the number of features considered when splitting each tree is set to the square root of the total number of features; and the minimum number of leaf node samples is set to 3. The random forest classification model constructs multiple decision trees based on the sample data. During the node splitting process, by evaluating the contribution of each feature dimension to reducing classification impurity, it automatically learns and establishes the optimal splitting threshold and decision weights of the Time Series Irrigation Index (TSII) when distinguishing crops, thus completing the model training.

[0050] S4.4: Input the high-dimensional enhanced feature set corresponding to each pixel in the target area into the trained crop classification model. Multiple decision trees vote to predict the crop category pixel by pixel and output a preliminary spatial distribution map of crops in the target area.

[0051] S4.5: Filter the preliminary spatial distribution map to remove small patches, extract the spatial distribution results corresponding to corn, and output the final corn spatial distribution map.

[0052] More specifically, considering that the preliminary thematic map generated based on pixel classification is discrete category label data, it may contain a small number of isolated misclassified pixels (visually appearing as fragmented patches similar to "salt and pepper noise") due to factors such as mixed pixels. To avoid destroying the discrete properties of the classification labels, this invention abandons median filtering, which is suitable for continuous variables, and instead uses a majority filter to smooth the classification results, including the following steps: S4.5.1: Use a preset sliding window (a 3×3 square window is used in this embodiment) to traverse the spatial distribution map pixel by pixel, count the frequency of occurrence of each crop category in each sliding window, and determine the crop category with the highest frequency of occurrence as the mode category of the corresponding sliding window; S4.5.2: Replace the crop category of the center pixel of each sliding window with the corresponding mode category to obtain the filtered classification result image; S4.5.3: To completely eliminate stubborn isolated pixels, the filtered classification result map is repeatedly subjected to mode filtering until the preset number of iterations is reached (the number of iterations is 2 in this embodiment), the spatial distribution result corresponding to corn is extracted, the final corn spatial distribution map is output, and the corn planting area is counted.

[0053] This post-processing step effectively removes small, isolated patches, thereby significantly improving the spatial continuity and visual quality of the final map.

[0054] To verify the effectiveness of this invention, a typical corn-soybean rotation area (Site 1) in Iowa, USA, during the early season (May-July) in 2023 was used as the study area. Based on Sentinel-2 remote sensing imagery, the mapping effects of two classification algorithms, Support Vector Machine (SVM) and Random Forest (RF), under different feature combinations were compared and analyzed. Figure 4 As shown in the image, yellow represents corn and green represents soybeans. Figure 4 (a) Represents USDA Farmland Data Layer (CDL) reference data (standard value); Figure 4 (b) represents the classification result of only multi-feature + SVM; Figure 4 (c) represents the classification result with only multiple features and RF; Figure 4 (d) represents the Sentinel-2 true-color composite reference image; Figure 4 (e) represents the classification result of multi-feature + TSII + SVM; Figure 4(f) represents the classification results of multi-feature + TSII + RF. A confusion matrix was constructed using the reserved 30% test set samples to quantitatively evaluate the classification results, and the overall classification accuracy (OA), Kappa coefficient, producer accuracy (Recall), user accuracy (Precision), and F1 score (F1-score) for corn and soybean categories were calculated. The results are shown in Table 1.

[0055] Table 1

[0056] As shown in Table 1, under the conditions of the early growing season (May to July), after introducing the Time Series Irrigation Index (TSII) feature, the overall classification accuracy (OA) of the Random Forest (RF) model increased from 95.34% to 96.64%, and the Kappa coefficient increased from 0.9261 to 0.9327. This indicates that the TSII feature can effectively enhance the discrimination ability between easily confused crops such as corn and soybean, and reduce classification confusion caused by spectral similarity. Furthermore, after introducing the TSII feature, the overall classification accuracy of the classification model reached over 96%, and the Kappa coefficient exceeded 0.93. In particular, the F1 score of the corn category was improved, indicating that the present invention has good effects in terms of corn identification accuracy and category discrimination ability.

[0057] Furthermore, compared to the full growing season model (May to September), the accuracy of early identification (May to July) in this scheme decreased by only about 1.04 to 1.55 percentage points, but the monitoring period was significantly shortened by about 2 months. This strongly validates its core timeliness value in early crop monitoring and yield forecasting. The final output maize distribution map is highly consistent with USDA statistics in terms of spatial pattern, further demonstrating the reliability and excellent operational application potential of this technical solution.

[0058] This invention constructs a multi-dimensional temporal spectral feature set by acquiring remote sensing image data of crops during the growing season in the target area. This feature set is then used to identify key time nodes during the tasseling stage of maize. Based on the temporal spectral features of these key time nodes, a temporal irrigation index is constructed. This irrigation index is then fused with the multi-dimensional temporal spectral feature set to construct a high-dimensional enhanced feature set. Finally, this high-dimensional enhanced feature set is used to train and classify a machine learning classifier, outputting a spatial distribution map of crops in the target area. Compared to classification methods that rely solely on conventional temporal spectral features, this invention, by introducing a temporal irrigation index that characterizes the temporal changes before and after the tasseling stage of maize, can better represent the differences between maize and other crops at key physiological stages, enhance the discriminative power of classification features, reduce classification confusion between easily confused crops, and thus improve the extraction effect of the spatial distribution map of crops in the target area.

[0059] Example 2: This embodiment provides a corn remote sensing classification electronic device, including a memory and a processor. The memory includes a corn remote sensing classification method program. When the corn remote sensing classification method program is executed by the processor, it implements the steps of a corn remote sensing mapping method as described in Embodiment 1.

[0060] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art can make other variations or modifications based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A method for remote sensing mapping of maize, characterized in that, Includes the following steps: Acquire remote sensing image data of the target area during the crop growing season, calculate the normalized vegetation index and land surface water index, and construct a multidimensional temporal spectral feature set. The key time nodes of maize tasseling stage are identified using the multidimensional time-series spectral feature set, and a time-series irrigation index is constructed using the spectral features of the time window corresponding to the key time node and the previous equal-length time window. The time-series irrigation index is fused with a multi-dimensional time-series spectral feature set to construct a high-dimensional enhanced feature set. The high-dimensional enhanced feature set is used to train and classify a machine learning classifier, and the spatial distribution map of corn in the target area is output.

2. The method for remote sensing mapping of maize according to claim 1, characterized in that, The method for constructing a multidimensional time-series spectral feature set includes the following steps: Acquire multi-band remote sensing image data of the target area during the crop growing season; Scene classification maps are used to preprocess multi-band remote sensing image data, images with cloud cover exceeding a preset threshold are removed, and cloud, shadow and thin cloud areas are masked to obtain effective observation image data. Define the growing season time range of crops in the target area, and divide the time range into several consecutive equal-length windows. Calculate the multi-temporal spectral features corresponding to the effective observation image data in each time window. The multi-temporal spectral features include normalized vegetation index, land surface water index, and near-infrared reflectance. The median composite method was used to perform time-series composite processing on the effective observation image data within each time window to obtain cloudless continuous time-series images. The cloudless continuous time series images are stacked in chronological order, and the multi-band remote sensing image data and multi-temporal spectral features corresponding to each time window are combined to construct a multi-dimensional time series spectral feature set.

3. The method for remote sensing mapping of maize according to claim 2, characterized in that, The multi-band remote sensing image data is multi-temporal Sentinel-2 surface reflectance data, including red band data, near-infrared band data, and short-wave infrared band data.

4. The method for remote sensing mapping of maize according to claim 1, characterized in that, The method for constructing time-series irrigation indices includes the following steps: Obtain the normalized vegetation index time series data of the multidimensional time-series spectral feature set, and perform pixel-by-pixel traversal of the normalized vegetation index time series data to construct the normalized vegetation index growth curve of each pixel during the crop growing season. The effective peak time of the normalized vegetation index growth curve was determined by the dynamic peak detection algorithm, and the date corresponding to the effective peak time was determined as the key time node t2 of the maize tasseling stage. Centered on t2, a dynamic observation window T2 is constructed using a preset time span. A reference node t1 is determined by tracing back a preset time interval from t2. A dynamic reference window T1 of the same length as t2 is constructed with t1 as the center. The preset time interval is greater than or equal to the preset time span. Obtain the time series sequence of land surface moisture index for each pixel in the multidimensional time-series spectral feature set within T1 and T2, denoted as follows: and And calculate the median land surface moisture index within T1 respectively. and the median land surface moisture index within T2 ; Using the and The time-series irrigation index (TSII) for each pixel is obtained by performing interpolation calculation, as shown in the following expression: 。 5. The method for remote sensing mapping of maize according to claim 4, characterized in that, Determining the effective peak time of the normalized vegetation index growth curve using a dynamic peak detection algorithm includes the following steps: The normalized vegetation index growth curve was smoothed using the n-point sliding window averaging method to obtain the smoothed normalized vegetation index growth curve. The smoothed normalized vegetation index growth curve is filtered for invalid nodes using a preset vegetation determination threshold, and the effective normalized vegetation index data is output. The global maximum value of the effective normalized vegetation index data is then extracted. Extract nodes that are greater than or equal to a preset proportion of the global maximum value as a candidate peak set; The candidate peak set is sorted in chronological order, and the time node corresponding to the earliest appearing candidate peak is selected as the effective peak time of the corresponding pixel.

6. The method for remote sensing mapping of maize according to claim 1, characterized in that, The fusion of time-series irrigation indices with multidimensional time-series spectral feature sets includes the following steps: Obtain the multidimensional temporal spectral feature set corresponding to each pixel in the target area, and represent the multidimensional temporal spectral feature set as a multidimensional feature layer arranged in time order; Obtain the time-series irrigation index corresponding to each pixel, and use the time-series irrigation index as a new feature layer; The newly added feature layer and the multi-dimensional feature layer are band-stitched together to obtain the high-dimensional enhanced feature vector corresponding to each pixel. The high-dimensional enhanced feature vectors corresponding to each pixel in the target area are reorganized according to their spatial location to construct a high-dimensional enhanced feature set.

7. A method for remote sensing mapping of maize according to claim 6, characterized in that, Training and classifying a machine learning classifier using a high-dimensional enhanced feature set includes the following steps: Based on the existing arable land dataset, we used stratified random sampling to extract sample points of corn and easily confused crops, and obtained the real crop labels corresponding to each sample point. Extract the high-dimensional enhanced feature vectors corresponding to each sample point in the high-dimensional enhanced feature set, and construct a training sample set containing the real crop label and the high-dimensional enhanced feature vectors; Build a machine learning classifier, input the training sample set into the machine learning classifier for training, and obtain the trained crop classification model; Input the high-dimensional enhanced feature set corresponding to each pixel in the target area into the trained crop classification model, and output a preliminary spatial distribution map of crops in the target area; The preliminary spatial distribution map is filtered to extract the spatial distribution results corresponding to corn, and the final corn spatial distribution map is output.

8. A method for remote sensing mapping of maize according to claim 7, characterized in that, The machine learning classifier is either a random forest classification model or a support vector machine model.

9. A method for remote sensing mapping of maize according to claim 7, characterized in that, The filtering process is a mode filtering process, which includes the following steps: By using a preset sliding window to traverse the spatial distribution map pixel by pixel, the frequency of occurrence of each crop category in each sliding window is counted, and the crop category with the highest frequency is determined as the mode category of the corresponding sliding window. Replace the crop category of the center pixel of each sliding window with the corresponding mode category to obtain the filtered classification result image; The filtered classification result map is subjected to mode filtering again until the preset number of iterations is reached, and the final crop spatial distribution map is output.

10. An electronic device for remote sensing classification of corn, characterized in that, The electronic device includes a memory and a processor. The memory includes a corn remote sensing classification method program. When the corn remote sensing classification method program is executed by the processor, it implements the steps of a corn remote sensing mapping method as described in any one of claims 1 to 9.

Images