Method and device for generating and automatically mapping remote sensing of corn samples in monoculture and intercropping

Abstract

The application provides a kind of single cropping and intercropping corn sample generation and remote sensing automatic mapping method and device, comprising: obtaining Sentinel-2 ground reflectivity image, generating composite time series image using the median of reflectivity and index data, obtaining spectral characteristics;Using the applicable spectral characteristics, a two-step hierarchical sample generation strategy is used to generate single cropping and intercropping corn samples, including identifying potential corn areas based on vegetation pigment time variation VVP index within the optimal phenology window and generating single cropping and intercropping corn samples based on water greenness ratio WGR index;Single cropping corn samples, intercropping corn samples and non-corn field reference samples are used as training data set, random forest classifier is trained, and the mapping of single cropping and intercropping corn is completed.The application originally proposes a classification framework suitable for single cropping and intercropping corn mapping, which combines automatic sample generation with random forest classification, reduces the dependence on manual marking while maintaining the classification effect.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image processing and classification technology, and in particular to a method and apparatus for generating monoculture and intercropping maize samples and for automatic remote sensing mapping. Background Technology

[0002] Monoculture and intercropping of maize typically coexist in areas dominated by smallholder farmers, playing distinct roles in agricultural production. Monoculture maize refers to planting only maize in the same field during the growing season, characterized by high intensification and specialization, meeting the demands of large-scale production and food security. Intercropping maize refers to planting two or more crops, including maize, in the same field during the growing season, maximizing water and land resource utilization for smallholder farmers with limited resources, and mitigating risks such as pest outbreaks, soil degradation, and climate change. Although intercropping maize is prevalent in smallholder areas, its spatial distribution data is largely lacking in current agricultural surveys and mapping datasets, primarily due to the scarcity of automatic classification and remote sensing mapping methods for monoculture and intercropping maize. Currently, most maize distribution remote sensing classification and mapping technologies are designed and mapped primarily for major agricultural regions dominated by large-scale monoculture, such as the US Corn Belt, North China, and the Northeast Plain of China. For mapping complex maize planting patterns, existing technologies mainly utilize the high spatiotemporal resolution of Sentinel-2 as the data foundation, with implementation schemes focusing on feature analysis and manual sample interpretation.

[0003] (1) Local phenological analysis

[0004] This method utilizes multi-temporal Sentinel-2 data for phenological analysis, distinguishing between monoculture and intercropping maize by identifying useful time windows. For example, Ibrahim et al. mapped phenological stages in regions such as Nigeria using spectral differences.

[0005] (2) Spectral difference analysis

[0006] A few methods have proposed schemes for analyzing the differences in spectral characteristics between monoculture and intercropped maize at different phenological stages within small regions, and have provided suggestions for effectively distinguishing between monoculture and intercropped maize. For example, Mahlayeye et al., in Busia County, Kenya, used a small sample set to analyze the spectral differences between monoculture and intercropped maize at different phenological stages, and proposed a phenological stage and sensitive feature selection scheme applicable to effectively distinguish between monoculture and intercropped maize.

[0007] (3) Manual interpretation and sample label generation

[0008] To obtain the sample data needed to train machine learning models, existing methods often employ manual interpretation of high-resolution images. For example, Qi et al. combined their field survey experience with manual interpretation of high-resolution images to generate training data for smallholder farmers in the Hetao Irrigation District of China to classify crops (including intercropping maize).

[0009] Although existing remote sensing mapping techniques have made significant progress in identifying monoculture maize, and some studies have analyzed the possibility of distinguishing between monoculture and intercropping maize, the following limitations exist when using remote sensing to map monoculture and intercropping maize in complex agricultural landscapes where intercropping is prevalent in small-scale farming areas and monoculture maize is interspersed:

[0010] 1. Challenges in Feature and Phenological Window Selection for Complex Planting Pattern Recognition

[0011] In smallholder farming areas, crops (such as intercropping maize and legumes) are mixed on small, scattered plots of land, resulting in satellite pixels exhibiting mixed spectral characteristics. Existing remote sensing classification methods for maize distribution are mostly based on the assumption that "one pixel represents a single crop," which masks spectral differences at key growth stages when identifying intercropped maize. This makes it difficult to identify the sensitivity indices and optimal time windows for distinguishing between monoculture and intercropped maize. This severe spectral signal confusion first leads to unclear phenological characteristics. Although existing technologies utilize basic indices such as NDVI, they struggle to capture subtle differences in canopy greenness and moisture dynamics in maize intercropping patterns. Second, in rainfed agriculture areas, crop planting dates vary greatly due to climate and geographical conditions. Defining absolute phenological windows based on calendar dates further complicates identifying the optimal windows for discerning the subtle differences between monoculture and intercropped maize.

[0012] 2. High cost of sample acquisition and lack of data

[0013] While machine learning methods demonstrate stability and reliability in multi-crop classification, they heavily rely on large-scale, high-quality ground-labeled samples. In small-scale farming areas, due to inconvenient transportation, low economic levels, and the absence of intercropping information in statistical or map data, obtaining sufficient intercropping training samples covering the entire country is extremely costly and inefficient. Existing remote sensing classification explorations of intercropped maize are mostly concentrated in local pilot areas, lacking a scheme to automatically identify high-purity monoculture and intercropped maize samples on a large scale. This leads to a data-scarce bottleneck for supervised learning methods in intercropping mapping applications. Regarding automatic sample generation, existing schemes, when applied to the complex identification of monoculture and intercropped maize, face challenges due to the high spectral similarity between monoculture and intercropped maize, making it difficult to eliminate inter-class confusion and lacking hierarchical selection strategies. Furthermore, existing methods often treat the distinction between monoculture and intercropped maize as a simple binary distinction problem, failing to consider the interference of other non-maize crops in the identification process. In practical applications, spectral noise from non-maize crops also severely affects further subdivision of monoculture and intercropped maize. This makes phenological identification methods that rely solely on a single feature or simple empirical threshold prone to misjudgment or omission.

[0014] 3. There is a lack of large-scale classification methods that take into account both phenological experience and the generalization ability of machine learning.

[0015] While hybrid approaches combining phenological experience with machine learning offer an alternative for complex multi-class classifications where field sample data is scarce, they have not yet been applied to address specific challenges in classifying monoculture and intercropping maize. Due to the lack of an effective automated framework for converting expert knowledge into machine-learnable generalized information in monoculture and intercropping maize mapping, existing local phenological rules are difficult to generalize to dynamic monitoring at the national scale, resulting in poor spatial consistency and temporal continuity in the generated monoculture and intercropping maize maps. Summary of the Invention

[0016] To address the above technical problems, this invention provides a method and apparatus for generating monoculture and intercropping maize samples and for automatic remote sensing mapping. The specific technical solution is as follows:

[0017] A method for generating samples of monoculture and intercropping maize and for automatic remote sensing mapping includes the following steps:

[0018] Step 1: Obtain Sentinel-2 surface reflectance images, and use the median of reflectance and exponential data to generate synthetic time series images to obtain a time series dataset of spectral features;

[0019] Step 2: Using applicable spectral features, a two-step stratified sample generation strategy is adopted to generate monoculture and intercropping maize samples, including identifying potential maize areas based on vegetation pigment time variation (VVP) index within the optimal phenological window and generating monoculture and intercropping maize samples based on water-to-greenness ratio (WGR) index.

[0020] Step 3: Using monoculture maize samples, intercropping maize samples, and non-maize field reference samples as training datasets, train a random forest classifier and complete the mapping of monoculture and intercropping maize.

[0021] An automatic remote sensing mapping device for generating monoculture and intercropping maize samples includes:

[0022] The spectral feature acquisition module acquires Sentinel-2 surface reflectance images, uses the median of reflectance and exponential data to generate synthetic time-series images, and obtains a time-series dataset of spectral features.

[0023] The sample generation module uses a two-step hierarchical sample generation strategy to generate monoculture and intercropping maize samples using applicable spectral features. This includes identifying potential maize areas based on the vegetation pigment time change VVP index within the optimal phenological window and generating monoculture and intercropping maize samples based on the water-to-greenness ratio WGR index.

[0024] The mapping module uses monoculture maize samples, intercropping maize samples, and non-maize field reference samples as training datasets. It is trained using a random forest classifier and completes mapping of monoculture and intercropping maize.

[0025] An electronic device, characterized in that it comprises: one or more processors; and a memory for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the method.

[0026] A computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, cause the processor to implement the method described thereon.

[0027] The present invention has the following beneficial effects:

[0028] 1. The applicable index features and phenological time windows for the automatic generation of monoculture and intercropped maize samples were determined. For example, the applicability of VVP and WGR were analyzed and used as the main features for generating monoculture and intercropped maize samples. Through separation and histogram analysis of the above two indices, the phenological time windows for VVP and WGR calculation were determined—that is, VA and VVC in VVP are 40-60 days and 10-30 days before heading, respectively, and NDVI and SWIR1 in WGR are 40-60 days before heading and 0-20 days after heading, respectively.

[0029] 2. A two-step stratified method for automatically generating samples of monoculture and intercropped maize is proposed. The invention originally employs a two-step stratified sample generation strategy in the automatic generation of samples for monoculture and intercropped maize. Figure 2It solves the problems of spectral confusion and insensitivity to key features. That is, relying on the defined decision threshold, the VVP index first identifies fields dominated by maize, and the WGR index further distinguishes between monoculture maize and intercropped maize based on canopy greenness-moisture dynamics.

[0030] 3. A hybrid classification method combining sample generation and machine learning, applicable to large-scale mapping of monoculture and intercropping maize, is proposed. Figure 1 This invention proposes an original classification framework applicable to mapping of monoculture and intercropping maize, combining automatic sample generation with random forest classification, which reduces reliance on manual labeling while maintaining classification effectiveness. Attached Figure Description

[0031] Figure 1 This is a flowchart of the method of the present invention;

[0032] Figure 2 Generate a flowchart for stratified samples;

[0033] Figure 3 The NDVI and SWIR1 time series curves for monoculture and intercropping maize;

[0034] Figure 4 For the location and overview of the test area;

[0035] Figure 5 The index representing the separation of VVC and VA values ​​for maize and non-maize at different growing season dates;

[0036] Figure 6 The dissociation index of SWIR1 and NDVI values ​​for monoculture and intercropped maize on growing season dates;

[0037] Figure 7 A schematic diagram showing the VVP frequency histograms for corn and non-corn, as well as the integral curves and percentile markings of the VVP histograms for corn and non-corn.

[0038] Figure 8 The WGR frequency histograms for monoculture and intercropping maize, the integral curves of the WGR histograms for monoculture and intercropping maize, and the schematic diagram of percentile marker thresholds are shown.

[0039] Figure 9 Map showing the distribution of monoculture and intercropping maize during the long rainy season of 2023;

[0040] Figure 10 This is a schematic diagram of the confusion matrix and accuracy evaluation of the hybrid classification method of the present invention;

[0041] Figure 11 A diagram showing the comparison of area estimates from multiple sources. Detailed Implementation

[0042] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other. To achieve the above objectives, this invention adopts the following technical solution.

[0043] The flowchart of the method for generating monoculture and intercropping maize samples and automatically mapping remote sensing data proposed in this invention is as follows: Figure 1 As shown. The main data inputs required for this invention are Sentinel-2 satellite time-series data (reflectivity bands and indices) and reference auxiliary data (cultivated land masking and field non-maize survey data). The main process of the invention is a phenology-driven two-step hierarchical sample generation and random forest classifier mapping. First, through a two-step hierarchical identification process based on phenological knowledge, maize areas are identified from different crop pixels using the temporal variation of the vegetation-pigment (VVP) index, and monoculture and intercropping maize samples are automatically generated from the maize areas using the water-to-greenness ratio (WGR) index. Next, the generated monoculture maize samples, intercropping maize samples, and non-maize field reference samples are used to train a random forest model, which is then deployed in the study area to complete the mapping of monoculture and intercropping maize using features from the Sentinel-2 satellite time-series data. The detailed steps are as follows:

[0044] Step 1: Data Input and Preprocessing

[0045] The invention primarily utilizes Sentinel-2 multispectral instrument surface reflectance images. First, cloud-masking products such as S2_CLOUD_PROBABILITY are used to label and remove pixels affected by cloud cover from the Sentinel images. Then, a 10-day synthetic time-series image is generated using the median of Sentinel-2 reflectance and index data over 10 days, and time-linear interpolation is used to fill in missing data values ​​in the time series. Finally, the invention primarily uses six spectral bands, including green (G), red (R), Red Edge 1 (RE1), near-infrared (NIR), and shortwave infrared 1 (SWIR 1), as well as four spectral indices commonly used for maize phenological analysis and extraction (see Table 1).

[0046] Table 1. Required Indices for the Scheme

[0047]

[0048] The primary reference data used in this invention includes farmland masking and field non-maize survey data. Farmland masking is mainly used to exclude non-farmland pixels in the target area, such as forests, built-up areas, grasslands, wetlands, water bodies, and shrublands. Field non-maize survey data comes from the center coordinates of other non-maize plots recorded by GPS devices, ensuring that sampling points are not near field boundaries to avoid pixel mixing. Non-maize samples include various crop types, such as wheat, rice, perennial crops (e.g., coffee, tea, sugarcane), peas, and other crops (e.g., beans, mung beans, potatoes, sorghum). In addition, the collection of non-maize samples is supplemented by manually interpreting high-resolution Google Earth images and field photographs.

[0049] Step 2: Generation of monoculture and intercropping maize samples based on a hierarchical strategy

[0050] Collecting sufficient training samples in the field to distinguish different maize planting patterns is constrained by the high cost of field surveys in smallholder areas and the fragmented nature of smallholder farmland. This invention employs a stratified sample generation strategy to overcome the challenge of identifying the complexity of monoculture and intercropping maize in dispersed and complex landscapes. Although maize exhibits distinct spectral-temporal characteristics compared to other crops, distinguishing between monoculture and intercropping maize remains challenging due to the similarity of their phenological signals. Therefore, see... Figure 2 The sample generation process employed a stratified strategy using two phenological time series indices, applied progressively to reduce spectral obfuscation. First, the VVP index was applied to distinguish maize from non-maize crops by leveraging the significant canopy vegetation-pigmentation variations. After identifying maize-dominated fields, the WGR index was then used to further identify subtle spectral differences between monoculture and intercropping, generating corresponding samples in the absence of interference from other crop types.

[0051] (1) First step of stratification: Identifying potential maize regions based on VVP index

[0052] The first step is to distinguish maize from other crops within numerous fields in the target area. This invention further illustrates how this feature can be used to identify maize in complex smallholder farming systems. Whether monoculture or intercropping is used, maize fields exhibit spectral characteristics distinct from other crops: maize fields show higher OSAVI values ​​during the growing season, typically not lower than the TCARI value, which is sensitive to chlorophyll. This specific vegetation pigment relationship differs from patterns observed in other crops, forming the basis for distinguishing maize from non-maize crops based on crop phenology. Furthermore, maize exhibits lower ARI anthocyanin values ​​during the tasseling stage, unlike similar upland crops (such as sorghum), which facilitates spectral differentiation between them. The VVP index is calculated as follows:

[0053] (1)

[0054] Among them, VVC (Variation coupling of Vegetation-Chlorophyll) refers to the vegetation-chlorophyll variation coupling index, which quantifies the high photosynthetic efficiency of maize by calculating the difference between normalized OSAVI and TCARI values; VA (Variation of Anthocyanin dynamic) refers to the anthocyanin dynamic change index, which quantifies the dynamic pattern of anthocyanin content by calculating the cumulative first-order difference of ARI values ​​during the maize tasseling period. The VVC calculation formula is as follows:

[0055] (2)

[0056] in, and , n1, and n2 represent the normalized OSAVI and TCARI values ​​for the i-th date within the time window, respectively, where n1 is the length of the phenological time window used to calculate VVC. VVC identifies maize by depicting a stable relative relationship of "higher OSAVI values ​​but relatively lower TCARI values ​​during the growing season." The formula for calculating VA is as follows:

[0057] (3)

[0058] in, Let be the ARI value for the i-th date within the time window, and n2 be the length of the phenological time window for calculating VA. VA enhances the distinction between maize and other dryland crops (such as sorghum) by characterizing the rapid temporal dynamic changes in anthocyanin content. In summary, the product of VVC and VA, forming VVP, can serve as a reliable feature for identifying maize distribution areas. To utilize VVP to distinguish maize from other crops within a target cultivated area, it is also necessary to determine the optimal phenological window for calculating VVP and the decision rule threshold for VVP. Based on field sample analysis in Kenya, a smallholder farming area, the optimal time windows for calculating VVC and VA are 30 to 10 days and 60 to 40 days before the heading date, respectively, serving as the calculation time windows for Formulas 2 and 3. The 95th percentile of the VVP value for non-maize crops—0.05—is determined as the decision threshold for identifying high-purity maize samples and non-maize crops; that is, a VVP value greater than 0.05 indicates maize.

[0059] (2) Second step of stratification: Generate monoculture and intercropping maize samples based on WGR index

[0060] Since the changes in maize canopy vegetation and pigment spectrum show a consistent trend regardless of whether it is monoculture or intercropping, the VVP index based on OSAVI, TCARI, and ARI indices has limitations when further dividing maize fields. Therefore, a second step is needed: utilizing the specific spectral bias introduced by short-term legume crops intercropped with maize, a relevant index is constructed to complete the identification of monoculture and intercropped maize. Two synergistic indicators commonly used to identify legume crops—NDVI and SWIR1—are used to jointly characterize the crop canopy greenness-water dynamics. Figure 3 (The solid line represents the mean curve, and the corresponding area is a range of one plus or minus standard deviation.) This shows a significant difference between the two indicators: monoculture maize, due to its denser planting density and higher vegetation cover, exhibits a higher and more stable NDVI value in the early to mid-growth stages. Conversely, intercropped maize, due to the rapid water loss caused by the high senescence and early harvest rates of the intercropped legumes, exhibits a higher SWIR1 value in the later stages of growth. Therefore, based on this observed spectral phenological difference, the WGR index was used to quantify the difference between the high greenness in the early stages of monoculture maize fields and the late-stage SWIR1 signal enhancement in intercropped maize fields. The formula for calculating the WGR index is as follows:

[0061] (4)

[0062] in, It is the average of the SWIR1 values ​​for all dates within the corresponding phenological time window, where n3 is the length of the time window. The SWIR1 value for the i-th date; It is the average NDVI value of all dates within the corresponding phenological time window, where n4 is the length of the time window. Let represent the NDVI value for the i-th date. The WGR index can integrate the opposite temporal trends of NDVI and SWIR1 for monoculture and intercropping maize, thereby enhancing the spectral separability between the two planting patterns throughout the vegetative growth period. To further distinguish monoculture and intercropping within potential maize regions using WGR, it is also necessary to determine the optimal phenological window for calculating WGR and the decision rule threshold for WGR. Based on field sample analysis in Kenya's smallholder farming areas, the optimal time window for NDVI was determined to be 40 to 60 days before tasseling (i.e., the early vegetative growth stage); the optimal time window for SWIR1 was determined to be 0 to 20 days after tasseling (i.e., the late vegetative growth stage). The determined time windows were applied to the calculation in Equation 4. Furthermore, the thresholds for determining high-purity monoculture / intercropping maize samples were determined as follows: intercropping maize: WGR > 2.0; monoculture maize: WGR < 1.0.

[0063] Step 3: Random Forest Classification Mapping Based on Generated Samples

[0064] Based on the samples generated in step two, Random Forest (RF) is selected as the classifier. It can adaptively learn useful features from different planting patterns and provide accurate classification results for multi-class discrimination (such as monoculture maize, intercropping maize, and non-maize crops). RF, as an ensemble learning method that aggregates multiple decision trees, is widely used in remote sensing mapping applications due to its resistance to overfitting and its ability to handle high-dimensional datasets. The input features of the Random Forest classifier used in this invention include raw reflectance bands, such as green band, red band, first red edge band, near-infrared band, and shortwave infrared band, as well as corresponding feature indices, including NDVI and TCARI. The training dataset for the classifier includes monoculture and intercropping maize samples automatically generated in step two, as well as collected non-maize reference data. Considering both computational complexity and efficiency, based on commonly used grid search strategies, two important parameters of the Random Forest classifier are set in this invention: the number of subtrees used to form the ensemble (ntree) is 200, and the number of variables used to grow the trees (nfeature) is the square root of the total number of features.

[0065] Specific Implementation Examples: Introduction to Technical Effect Testing Areas and Field Reference Sample Sets:

[0066] The selection of test areas needs to consider both areas where maize is the primary crop and smallholder farming areas where both monoculture and intercropping exist. Kenya is one of the largest maize producers in sub-Saharan Africa, with an official estimated production of 4.28 million tons in 2023. Kenya's terrain extends from the coastal plains to the eastern edge of the East African Plateau and the East African Rift Valley, with an average annual rainfall of 508 mm (e.g., ...). Figure 4As shown in (a), (b), and (c), most areas experience two rainy seasons (long rainy season: March to May; short rainy season: October to December). Maize is a major crop in Kenya, primarily grown as monoculture for commercial purposes, or intercropped with other short-term staple or cash crops (such as legumes) for smallholder livelihoods. During the long rainy season of 2023 and the short rainy season of 2023-2024, a field survey meticulously labeled monoculture and intercropped maize, collecting authentic reference samples. A total of 1515 maize samples (891 monoculture and 299 intercropped maize samples during the long rainy season; 222 monoculture and 103 intercropped maize samples during the short rainy season) and 5400 non-maize crop samples were used for technical testing effectiveness verification, such as characteristic analysis and classification mapping results validation. In addition, various agricultural statistics corresponding to 2023 were collected, such as the Kenya Agriculture and Food Authority's (AFA) "Food Crop Statistics Yearbook", statistics from the Food and Agriculture Organization of the United Nations (FAO), and data from the Copernicus4GEOGLAM service of the United States Department of Agriculture (USDA) and the European Commission's Joint Research Centre (including Kenyan crop mapping and area estimation statistics), for technical testing and comparison of the mixed classification methods.

[0067] Example 1: Explanation of the test results of index characteristics and phenological windows

[0068] Considering that maize planting times may vary across large arable land areas, the effectiveness of the determined optimal phenological window for the VVP index can be demonstrated based on the separability analysis of a field sample set. The separation of VVC and VA values ​​for maize and non-maize samples at different growth stages using the SI (Separability Index) was calculated, and the times with the highest separation were used as the optimal phenological window for VVP calculation. The formula for calculating SI is as follows:

[0069] (5)

[0070] in, and Let h and k represent the average spectral characteristics m of two comparison crops (e.g., maize and non-maize, or subsequently, monoculture and intercropping maize) over a time period n. and The values ​​represent the standard deviations of the spectral characteristics m of the two comparative crops within time periods h and k, respectively. Indicates inter-class spectral heterogeneity. Indicates intraclass spectral heterogeneity. A higher SI value indicates that the two crop types have a stronger ability to distinguish specific characteristics at a given time. Figure 5 The separability of VVC and VA time series for maize and non-maize crops (black bars) shows that the peak SI value of VVC occurs within 30 days before the tasseling date, with all values ​​above 0.4, indicating high separability of VVC for both crops during the tasseling stage. VA shows peak separability between 40 and 60 days before the tasseling date, indicating a significant change in maize anthocyanin content from sowing to tasseling. To address the differences in sowing time across large-scale smallholder areas, a relative time window definition is used instead of the common absolute calendar date format (day of the year), i.e., the day before the tasseling date. The tasseling date within the maize growing season is determined based on the maximum NDVI of the NDVI curve during the actual planting cycle, matched to variations in actual NDVI conditions in different regions. Furthermore, Figure 5 The separability of VVC and VA for monoculture and intercropping maize (gray bars) is shown. Figure 5 In the figures, (a) represents the VVC of maize and non-maize, and (b) represents the VA value. The separation index (SI) values ​​for these crops were consistently low on different growing season dates, indicating that monoculture and intercropping maize share similar VVP characteristics. This also suggests that, regardless of whether it's monoculture or intercropping, maize-dominated fields exhibit significant differences in chlorophyll and anthocyanin levels, which can be distinguished from other non-maize types using the VVP index. Separability analysis of the VVP index on different dates shows that the corresponding phenological time window determined in this invention is reasonable and effective, and can best distinguish maize from non-maize at a reasonable time.

[0071] Figure 6 The separability of NDVI and SWIR1 time series for monoculture and intercropped maize demonstrated shows that the maximum SI index occurs in SWIR1 one month after tasseling and in NDVI 40 to 60 days before tasseling, respectively. Therefore, the determined optimal time window for NDVI calculation (40 to 60 days before tasseling, i.e., the early vegetative growth stage) and the optimal time window for SWIR1 calculation (0 to 20 days after tasseling, i.e., the late vegetative growth stage) are consistent with reality and can best distinguish monoculture and intercropped maize at reasonable times. Figure 6 In the table, (a) represents the separation index of SWIR1 values ​​for monoculture and intercropping maize on growing season dates, and (b) represents the separation index of NDVI values ​​on growing season dates.

[0072] Example 2: Explanation of the effect of the two-step stratification method in the automatic generation of samples for monoculture and intercropping maize.

[0073] In practical applications, other non-maize crops constitute significant noise in distinguishing between monoculture and intercropping maize. Stratified identification is an effective strategy to mitigate uneven inter-class differences and isolate easily confused classes. This strategy can identify pure samples of monoculture and intercropping maize from complex crop systems.

[0074] The first step, considering the unique temporal variations in vegetation, chlorophyll, and anthocyanin indices in maize-dominated fields, uses the VVP index to preferentially separate maize from other non-maize crops, thereby improving the subsequent identification of the two maize planting patterns. As a crucial parameter determining the effectiveness of this first step, the effectiveness of the VVP decision rule threshold in separating maize from other non-maize crops can be obtained through histogram analysis. A frequency histogram is a statistical chart that uses the area of ​​rectangles to display the distribution characteristics of continuous data. The horizontal axis represents the grouping intervals of the data, and the vertical axis represents the "frequency / group interval," with each bar representing the frequency of data occurrence within that group. By selecting appropriate quantile values ​​on the frequency histogram, a strict decision threshold is defined. This approach prioritizes high-purity sample pixels, ensuring that samples exceeding the threshold represent the target crop while maximizing representativeness and minimizing contamination from other crop types. Figure 7 The histogram distribution of VVP frequencies for maize and non-maize crops shows that maize VVP values ​​are concentrated in the positive range, while non-maize crops are more widely dispersed in the negative range. Figure 7 In the diagram, (a) shows the VVP frequency histogram for maize and non-maize crops; (b) shows the integral curve (solid line) of the VVP histogram for maize and non-maize crops and the percentile markers. When the VVP threshold is >0.05, the selection rate of maize samples is 24%, while only 5% of non-maize crop samples are selected. Based on this, the threshold determined in this invention is located at the 95th percentile of the VVP value, which can effectively identify high-purity maize samples and non-maize crops.

[0075] The second step utilized the temporal variations of vegetation and moisture indices in monoculture and intercropping maize. Influenced by planting spacing and the mixing signal from legume crops, the slightly lower greenness and canopy moisture during the growth process reflected the unique characteristics of maize. The WGR index can capture the dynamic characteristics of canopy greenness-moisture in intercropped maize, and its effectiveness lies primarily in the effectiveness of the corresponding decision threshold. The effectiveness of the WGR decision rule threshold is also explained based on histogram analysis. For example... Figure 8 The histograms of WGR frequencies for monoculture and intercropped maize partially overlap, but significant differences emerge at the tails of the histograms. The WGR values ​​for monoculture maize are mostly below 1.25, while those for intercropped maize are concentrated above 2.0. Figure 8 In the image, (a) shows the WGR frequency histograms for monoculture and intercropping maize; (b) shows the integral curves (solid lines) of the WGR histograms for monoculture and intercropping maize and the percentile marker thresholds. Figure 8Quantitative analysis showed that samples with WGR > 2.0 accounted for 32% of intercropped maize samples and only 5% of monocropped maize samples. Conversely, samples with WGR < 1.0 accounted for 28% of monocropped maize samples and only 5% of intercropped maize samples. Based on this analysis, the determined threshold can distinguish high-purity monocropped and intercropped maize samples.

[0076] Example 3: Explanation of the effect of using a mixed classification method for large-scale mapping of monoculture and intercropping maize.

[0077] like Figure 9 As shown in (a), (b), (c), and (d), these images illustrate the distribution maps of monoculture and intercropping maize in Kenya during the long rainy season of 2023, generated using the hybrid classification method proposed in this invention. To illustrate the advantages and effectiveness of the new mapping product, the latest 10-meter maize distribution map from Copernicus4GEOGLAM (Copernicus Group on Earth Observations for Global Agriculture Monitoring) in 2023 is used for comparison. Field reference plots were collected as accurate geographical references. Figure 9 In the study, for site A, where monoculture maize is the dominant crop, both the hybrid classification method of this invention and the Copernicus4GEOGLAM maize distribution map accurately identified the geographical locations of the actual reference maize sample fields. For site B, where intercropping maize is the dominant crop, the hybrid classification method successfully extracted the distribution of intercropped maize plots, while the Copernicus4GEOGLAM crop type map failed to include the locations of most of the smaller actual reference intercropped maize sample fields. For site C, where both monoculture and intercropping maize are distributed, the Copernicus4GEOGLAM maize distribution map also missed crucial information on the actual intercropping distribution, while the hybrid classification method of this invention demonstrated advantages in identifying and extracting the actual reference sample fields for both monoculture and intercropping maize. Furthermore, the mapping results produced by the hybrid classification method of this invention also have clearer field boundaries and more complete field shapes.

[0078] Based on real reference samples collected in the field during the long rainy season of 2023 and the short rainy season of 2023-2024, the confusion matrix and F1 score between the classification results and the reference data were calculated to evaluate the accuracy of the hybrid classification method. Figure 10 ). Figure 10In the long rainy season, the overall accuracy (OA) reached 88.74%, and in the short rainy season, it reached 85.79%, indicating robust accuracy in mapping monoculture and intercropping maize. During the long rainy season, producer accuracy was 89%, 87.9%, and 88.2% for monoculture maize, intercropping maize, and non-maize crops, respectively, with corresponding user accuracy of 86.9%, 70.3%, and 95.6%. Furthermore, the F1 scores for both maize types exceeded 0.75: 0.8796 for monoculture maize and 0.7815 for intercropping maize, demonstrating the effectiveness of the hybrid classification method in identifying monoculture and intercropping maize. In the short rainy season, although classification accuracy decreased slightly, the hybrid classification method maintained an F1 score above 0.72 and a kappa coefficient exceeding 0.75, showing improved performance and advantages compared to existing complex planting pattern mapping techniques. Additionally, Figure 11 The study compared the maize acreage statistics based on a hybrid classification method with existing official statistics. The figure shows that the estimated total maize planting area in Kenya is 3,345,250 hectares, which is 31.03% higher than the surveyed area reported in the AFA Yearbook, 37.66% higher than the FAO statistics, 52.05% higher than the USDA statistics, and 49.91% higher than the Copernicus4GEOGLAM statistics. The invention of the hybrid classification method, with its advantage of comprehensively identifying both monoculture and intercropping maize, captures a more comprehensive total planted area and provides richer and more complete statistical information.

[0079] Alternative Solution 1: When applying the hybrid classification method of the invention, the specific classifier model can use other machine learning model methods, such as support vector machines, convolutional neural networks, and other machine learning / deep learning classifiers; the specific input features can be introduced according to the actual situation of the target area, including other reflectance bands or indices that are beneficial to corn identification; if there are field-collected samples in the target area, the specific optimal phenological window or threshold can be adjusted to a more favorable value according to the actual situation of the target area in accordance with the implementation scheme of the invention.

[0080] Another aspect of the present invention provides an apparatus for generating monoculture and intercropping maize samples and for automatic remote sensing mapping, comprising:

[0081] The spectral feature acquisition module acquires Sentinel-2 surface reflectance images, uses the median of reflectance and exponential data to generate synthetic time-series images, and obtains a time-series dataset of spectral features.

[0082] The sample generation module uses a two-step hierarchical sample generation strategy to generate monoculture and intercropping maize samples using applicable spectral features. This includes identifying potential maize areas based on the vegetation pigment time change VVP index within the optimal phenological window and generating monoculture and intercropping maize samples based on the water-to-greenness ratio WGR index.

[0083] The mapping module uses monoculture maize samples, intercropping maize samples, and non-maize field reference samples as training datasets. It is trained using a random forest classifier and completes mapping of monoculture and intercropping maize.

[0084] Another aspect of the present invention provides an electronic device, characterized in that it includes: one or more processors; and a memory for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the method.

[0085] Another aspect of the present invention provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, cause the processor to implement the method described thereon.

[0086] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of the present invention can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0087] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0088] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for generating samples of monoculture and intercropping maize and for automatic remote sensing mapping, characterized in that, Includes the following steps: Step 1: Obtain Sentinel-2 surface reflectance images, and use the median of reflectance and exponential data to generate synthetic time series images to obtain a time series dataset of spectral features; Step 2: Using applicable spectral features, a two-step stratified sample generation strategy is adopted to generate monoculture and intercropping maize samples, including identifying potential maize areas based on vegetation pigment time variation (VVP) index within the optimal phenological window and generating monoculture and intercropping maize samples based on water-to-greenness ratio (WGR) index. Step 3: Using monoculture maize samples, intercropping maize samples, and non-maize field reference samples as training datasets, train a random forest classifier and complete the mapping of monoculture and intercropping maize.

2. The method for generating monoculture and intercropping maize samples and automatically mapping remote sensing data according to claim 1, characterized in that, The spectral bands used in step one include green light band, red light band, red edge band, near-infrared band, and short-wave infrared band. The four spectral indices used include the Optimized Soil Corrected Vegetation Index (OSAVI), the Chlorophyll Absorption Reflectance Variation Index (TCARI), the Anthocyanin Reflectance Index (ARI), and the Normalized Difference Vegetation Index (NDVI).

3. The method for generating monoculture and intercropping maize samples and automatically mapping remotely according to claim 2, characterized in that, Identifying potential maize regions based on the vegetation pigment time variation (VVP) index includes: calculating the VVP index, which is obtained by multiplying the vegetation-chlorophyll variation coupling index (VVC) and the anthocyanin dynamic change index (VA); wherein, VVC quantifies the high photosynthetic efficiency of maize by calculating the sum of the normalized differences between OSAVI and TCARI values ​​within a time window of 30 to 10 days before tasseling, and VA quantifies the dynamic pattern of anthocyanin content by calculating the cumulative first-order difference component of ARI values ​​within a time window of 60 to 40 days before tasseling; pixels with a VVP value greater than 0.05 are identified as maize regions.

4. The method for generating monoculture and intercropping maize samples and automatically mapping remote sensing data according to claim 2, characterized in that, Generating monoculture and intercropping maize samples based on the water-to-greenness ratio (WGR) index includes: calculating the WGR index, which is the ratio of the average SWIR1 value within a time window of 0 to 20 days after tasseling to the average NDVI value within a time window of 40 to 60 days before tasseling; identifying pixels with WGR > 2.0 as intercropping maize samples and pixels with WGR < 1.0 as monoculture maize samples.

5. The method for generating monoculture and intercropping maize samples and automatically mapping remotely according to claim 2, characterized in that, The formula for calculating the VVP index is: ; in, The formula for calculating VVC is: ; The formula for calculating VA is: ; In the formula, and These are the normalized OSAVI and TCARI values ​​for the i-th date within the time window, respectively. The length of the time window is 30 to 10 days before heading. The ARI value for the i-th date within the time window. The time window length is 60 to 40 days before heading.

6. The method for generating monoculture and intercropping maize samples and automatically mapping remote sensing data according to claim 5, characterized in that, The formula for calculating the WGR index is: ； in, It is the average SWIR1 value for all dates within the corresponding phenological time window, and n3 is the length of the time window from 0 to 20 days after heading. The SWIR1 value for the i-th date; It is the average NDVI value for all dates within the corresponding phenological time window, where n4 is the length of the time window from 40 to 60 days before heading. Let be the NDVI value for the i-th date.

7. The method for generating monoculture and intercropping maize samples and automatically mapping remotely according to claim 1, characterized in that, The input features of the random forest classifier in step three include green light band G, red light band R, red edge band RE1, near-infrared band NIR, short-wave infrared band SWIR 1, normalized difference vegetation index NDVI, and chlorophyll absorption reflectance variation index TCARI.

8. A device for generating monoculture and intercropping maize samples and for automatic remote sensing mapping, characterized in that, include: The spectral feature acquisition module acquires Sentinel-2 surface reflectance images, uses the median of reflectance and exponential data to generate synthetic time-series images, and obtains a time-series dataset of spectral features. The sample generation module utilizes applicable spectral features and employs a two-step stratified sample generation strategy to generate monoculture and intercropping maize samples. This includes identifying potential maize regions based on the vegetation pigment temporal variation (VVP) index within the optimal phenological window and generating monoculture and intercropping maize samples based on the water-to-greenness ratio (WGR) index. The mapping module uses monoculture maize samples, intercropping maize samples, and non-maize field reference samples as training datasets, trains a random forest classifier, and completes mapping of monoculture and intercropping maize.

9. An electronic device, characterized in that, include: One or more processors; A memory for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the method of any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, It stores executable instructions that, when executed by a processor, cause the processor to perform the method described in any one of claims 1 to 7.

Images