A method for evaluating the natural quality of cultivated land based on all-remote sensing data-driven

Abstract

The present invention relates to the technical field of land use and quality assessment, and particularly relates to a method for evaluating the natural quality of cultivated land driven by all-remote sensing data. The present invention combines remote sensing monitoring technology for cultivated land quality evaluation, which can frequently and persistently provide surface information, and can monitor the dynamic changes of cultivated land quality on a long time scale. Considering that unsupervised machine learning has strong objectivity and does not require human intervention, it automatically finds the internal laws of cultivated land quality within the evaluation unit and the complex non-linear relationships with various evaluation indicators, and realizes a method for evaluating cultivated land quality driven by remote sensing indicators based on a competitive neural network. The present invention overcomes the limitations of traditional cultivated land quality evaluation relying on expert experience and the acquisition of evaluation indicators relying on ground surveys, greatly reduces the huge consumption of human and material resources in traditional cultivated land quality evaluation, and with its long-time multi-scale monitoring advantage, realizes the timely and efficient supervision of cultivated land resources.

Description

Technical Field

[0001] The present invention relates to the technical field of land use and quality assessment, and particularly relates to a method for evaluating the natural quality of cultivated land driven by all remote sensing data. Background Art

[0002] As a specific type of land, cultivated land is scarce and irreplaceable, and is the basic resource for human survival. Compared with the tangible reduction of cultivated land area, the decline in cultivated land quality is implicit and not easily noticed by people. However, its impact is no less than the reduction in cultivated land quantity. The change in cultivated land quality is posing a serious threat to food security, ecological environment and social and economic development.

[0003] The monitoring and evaluation of cultivated land quality are important prerequisites for realizing the scientific management and protection of cultivated land. The data sources of traditional cultivated land quality monitoring and evaluation rely on the methods of zoning layout and on-site sampling, but there are problems such as low monitoring accuracy at large scales, long cycle and slow update, and time-consuming and laborious, which are difficult to meet the needs of real-time update and rapid monitoring of cultivated land quality results.

[0004] The main methods used in traditional cultivated land quality evaluation include some comprehensive evaluation methods such as empirical judgment index method, analytic hierarchy process, fuzzy comprehensive evaluation method, matter element analysis method, regression analysis method, grey relational analysis method, etc. In the evaluation process of the above methods, the determination of index weights exists, and most of the weight determination relies on expert knowledge and experience, making the evaluation results not objective.

[0005] Remote sensing technology is known for its wide range of ground object information acquisition, short revisit cycle, strong currency, accuracy and reliability, and low cost. Using remote sensing classification technology can effectively extract the quantity and spatial distribution information of cultivated land. Combining multi-temporal image data and relying on remote sensing change detection technology, the extraction of crop planting structure can be realized. Most of the existing studies on evaluating cultivated land quality using remote sensing technology are to establish an inversion model of specific evaluation indicators by combining ground data on the basis of remote sensing data. Its specific indicators often can only reflect the quality of cultivated land in a certain connotation, making the evaluation results incomplete and not objective. The evaluation model follows some comprehensive evaluation methods in traditional cultivated land quality evaluation and does not get rid of the limitation of artificially determining weights. In recent years, there have also been some supervised machine learning methods used to determine index weights for cultivated land quality evaluation, but the model construction depends on prior knowledge and the robustness of the model is poor in actual application. Summary of the Invention

[0006] In view of the above existing problems or deficiencies, in order to overcome the limitations of traditional cultivated land quality evaluation relying on expert experience, the acquisition of evaluation indicators relying on ground surveys, and the construction of models relying on prior knowledge, as well as the technical problem of poor robustness of the model in practical applications; the present invention provides a method for evaluating the natural quality of cultivated land driven by all-remote sensing data. In this method, a competitive neural network is used to provide a general method for identifying patterns in multi-dimensional complex systems, which is an "unsupervised learning" model with strong self-adaptability; at the same time, remote sensing technology can monitor the information of the cultivated land surface and its coverings in a long-term and multi-scale manner. Combining the existing spectral indices and landscape indices can indicate the changes in cultivated land quality and achieve timely and efficient supervision of cultivated land resources.

[0007] The specific technical solution of the present invention is as follows:

[0008] A method for evaluating the natural quality of cultivated land driven by all-remote sensing data, the specific steps are as follows:

[0009] Step 1: Construct a remote sensing index system for cultivated land quality:

[0010] Based on the consideration and analysis of the connotation of cultivated land quality, following the unity principle, dominance principle, sensitivity principle, practicality principle, independence principle, and stability principle for selecting evaluation indicators, and considering the availability of remote sensing evaluation indicators, the remote sensing evaluation indicators are divided into two categories: one is to directly use multi-source remote sensing to obtain cultivated land quality monitoring and evaluation indicators; the other is to indirectly reflect the cultivated land quality status by remotely sensing the growth of crops.

[0011] Step 2: Determine the evaluation unit of cultivated land quality:

[0012] The evaluation unit of cultivated land quality is the basic space and the smallest unit of cultivated land quality evaluation, and it is the basis for dividing the cultivated land quality grades in the project area; in the evaluation of cultivated land quality, due to the diversity of data sources of various indicators, the smallest unit divided is not unified; the present invention uses the grid method to divide the evaluation unit, and the grid area should be smaller than the average area of the units divided in the cultivated land quality grade map of the study area to achieve the unity of different evaluation units.

[0013] Furthermore, the grid area is equal to the minimum resolution of the obtained remote sensing data to achieve high evaluation accuracy.

[0014] Step 3: Obtain evaluation index data:

[0015] Use remote sensing technology to obtain evaluation index data.

[0016] Step 4: Standardize the evaluation indicators:

[0017] Indicators of different categories have different dimensions. By performing standardization processing, the influence of dimensions among various remote sensing data obtained in Step 3 is eliminated to obtain training data, enabling the evaluation process to be carried out under the same standard;

[0018] Step 5, Detection of data outliers:

[0019] If there is a problem of class imbalance in the training data obtained in Step 4, it will seriously affect the accuracy of the model. Therefore, it is necessary to detect the outlier samples in the training data obtained in Step 4 and mark the outliers as a class.

[0020] Step 6, Construct a competitive neural network model:

[0021] 1) Initialize the network.

[0022] Perform normalization processing on the training data obtained in Step 4. Let the training data be x ij (i = 1, 2,..., m; j = 1, 2,..., n), and the normalization formula is as follows:

[0023]

[0024] In the formula, m is the number of evaluation units, n is the number of evaluation indicators, x j(min) is the minimum value of the j-th evaluation indicator among m evaluation units, x j(max) is the maximum value of the j-th evaluation indicator among m evaluation units, and X ij corresponds to an n-dimensional vector representing the cultivated land quality.

[0025] Generate a random number in the interval [0, 1] as the initial weight vector w of the network.

[0026] 2) Obtain the winning neuron.

[0027] For an input pattern X = {x1, x2,..., x n} input into the network, the weight vectors w corresponding to all neurons in the competitive layer are subjected to similarity tests with X, and the w most similar to X is determined as the winning neuron. The link distance is calculated using the linkdist function as a measure of similarity. The link distance L r between the n-dimensional vector X s and X rs is calculated using the following formula:

[0028] If i = j, then L rs = 0;

[0029] If the Euclidean distance e between X r and X s is <= 1, then L rs = 1;

[0030] If there exist \(k_1,k_2\in n\) such that then \(L\) rs \(= 3\);

[0031] If there exist \(k_1,\cdots,k\) N \(\in n\) such that then \(L\) rs \(= N\);

[0032] If the above situations are not satisfied, then \(L\) rs \(= n\).

[0033] 3) Weight adjustment.

[0034] The competitive neural network adjusts the weights of the winning neuron according to the Kohonen learning rule. For the winning neuron \(v\), the corresponding weight \(w\) vj The change amount \(\Delta w\) vj is defined as follows:

[0035]

[0036] \(\eta\) is the learning rate, defaulting to 0.01;

[0037] Through learning, the winning neuron will have a greater chance of winning when a similar input vector appears next time, while the neurons that lost the competition in the previous round will have a smaller chance of winning.

[0038] 4) Network output.

[0039] Repeat steps 2) to 3) until the number of units in each category no longer changes, and output the result.

[0040] Step 7: Calculate the natural quality score.

[0041] According to the "Regulations for the Classification of Agricultural Land Quality" (GB / T 28407 - 2012), use the geometric mean method to calculate the natural quality score \(C\) of the evaluation unit Li , and the formula is as follows:

[0042]

[0043] Where: \(C\) Li is the natural quality score of the agricultural land of the classification unit; \(i\) is the evaluation unit number; \(j\) is the evaluation index number; \(m\) is the number of evaluation indicators; \(f\) i is the score of the \(j\)th evaluation indicator in the \(i\)th evaluation unit, and the value range is \((0 - 100]\).

[0044] Step 8: Category evaluation.

[0045] Statistically analyze the average natural quality scores of evaluation units under each category obtained in steps 5 and 6. Based on the magnitudes of the average values, conduct grade evaluations and perform statistical analyses for each evaluation grade.

[0046] The present invention combines remote sensing monitoring technology for cultivated land quality evaluation, which can frequently and persistently provide surface information, and has the advantage of monitoring the dynamic changes in cultivated land quality on a long - time scale. Considering that unsupervised machine learning has strong objectivity and does not require human intervention, it automatically discovers the inherent laws of cultivated land quality within evaluation units and the complex non - linear relationships with various evaluation indicators, thus realizing a cultivated land quality evaluation method driven by remote sensing indicators based on a competitive neural network. The present invention overcomes the limitations of traditional cultivated land quality evaluation that relies on expert experience and the acquisition of evaluation indicators that depends on ground surveys. The introduction of remote sensing indicators greatly reduces the huge consumption of human and material resources in traditional cultivated land quality evaluation. With its long - time and multi - scale monitoring advantages, it realizes the timely and efficient supervision of cultivated land resources. Brief Description of the Drawings

[0047] Figure 1 is a flowchart of an embodiment of the present invention;

[0048] Figure 2 is a table of the evaluation index system for the embodiment;

[0049] Figure 3 is a graph of the clustering results of the competitive neural network;

[0050] Figure 4 is a schematic diagram of the structure of the competitive neural network model of the present invention;

[0051] Figure 5 is a schematic diagram of the category evaluation results for the embodiment;

[0052] Figure 6 is a map of the cultivated land quality grades at the county level of Luxian County in 2018 for the embodiment;

[0053] Figure 7 is a schematic diagram of the evaluation accuracy of cultivated land quality for the embodiment. Detailed Embodiment

[0054] Next, the technical solutions of the present invention will be clearly and completely described in conjunction with the embodiments and the drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of them. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the scope of protection of the present invention.

[0055] A method for evaluating the natural quality of cultivated land driven by all - remote - sensing data, as Figure 1 shown, the specific steps are as follows:

[0056] Step 1: Establish a remote sensing index system for cultivated land quality:

[0057] Based on the thinking and analysis of the connotation of cultivated land quality, following the principles of unity, dominance, sensitivity, practicality, independence, and stability in the selection of evaluation indicators, and considering the availability of remote sensing evaluation indicators, the remote sensing evaluation indicators are divided into two categories: one is to directly use multi-source remote sensing to obtain cultivated land quality monitoring and evaluation indicators; the other is to indirectly reflect the cultivated land quality status by remotely sensing the growth of crops.

[0058] In this embodiment, 4 evaluation indicators are optimized from two aspects of natural quality and utilization quality, including slope, anomaly percentage of normalized difference vegetation index, irrigation condition, and fragmentation degree;

[0059] Slope: The slope is obtained after secondary processing of the digital elevation model (DEM), and the DEM is generated by calculating stereo image pairs.

[0060] The slope is represented by the degree α and is calculated using the inverse trigonometric function. The formula is as follows:

[0061]

[0062] where h is the elevation difference and l is the horizontal distance or grid resolution;

[0063] Irrigation condition: The remote sensing image is classified by the object-oriented method to achieve the extraction of different cultivated land types, and then the classification result is corrected by the visual interpretation method. In this embodiment, the cultivated land type result map in the cultivated land quality evaluation result of Luxian County in 2019 by the Natural Resources Department is used. The cultivated land types are divided into three categories: paddy fields, irrigated land, and dry land, and the irrigation conditions decrease in turn;

[0064] Fragmentation degree: Based on high-resolution remote sensing images, the object-oriented method and other classification methods are used to extract plots as the units for monitoring and evaluation. The fragmentation degree of the plots is reflected by the landscape shape index. The calculation formula is as follows:

[0065] Landscape fractal dimension index:

[0066] In the formula, D represents the fractal dimension; P is the patch perimeter; A is the patch area. The larger the D value, the more complex the patch shape, which can be used to describe the fragmentation degree of the plot;

[0067] Anomaly percentage of normalized difference vegetation index: The normalized difference vegetation index (NDVI) is constructed by the reflection spectral information of the visible and near-infrared bands sensitive to vegetation by the sensor, and can reflect the crop growth situation and the soil organic matter content under specific conditions. The formula is as follows:

[0068]

[0069] Where NIR is the reflection value in the near-infrared band and R is the reflection value in the red band;

[0070] Averaging the NDVI values over the same period of multiple years can reflect the long-term productivity level of farmland. Considering the influence brought by different vegetation types, we use the anomaly analysis method to process the average value of the vegetation index over multiple years. The formula is as follows:

[0071]

[0072] In the formula, KL is the average value of the vegetation index over multiple years, is the average value of the vegetation index of the same vegetation type layer, and AP is the anomaly percentage.

[0073] Step 2: Determine the cultivated land quality evaluation unit:

[0074] The cultivated land quality evaluation unit is the basic space and the smallest unit for cultivated land quality evaluation, and it is the basis for dividing the cultivated land quality grades in the project area; in cultivated land quality evaluation, due to the diversity of data sources for various indicators, the smallest unit divided is not unified; the present invention uses the grid method to divide the evaluation unit to achieve the unity of different evaluation units.

[0075] The research area of this embodiment is the Sichuan hilly area. Considering the cultivated land distribution and area in the Sichuan hilly area, a 30×30m grid is used as the evaluation unit here.

[0076] Step 3: Obtain evaluation index data:

[0077] In this embodiment, the research data includes the DEM image with a resolution of 30m downloaded from the Geospatial Data Cloud Platform http: / / www.gscloud.cn / ; the ground survey data of the Ministry of Agriculture in 2017, including cultivated layer thickness, soil texture, soil organic matter and other cultivated land attributes; the cultivated land quality evaluation result maps of Luxian County, Sichuan Province in 2018, including cultivated land types, cultivated land quality grades, etc.; the remote sensing images with a resolution of 10m of Sentinel2 in Luxian County in 2015, 2016 and 2017 downloaded from the Copernicus Data Center of the European Space Agency https: / / scihub.copernicus.eu / dhus / # / home.

[0078] Among them, the reception time of the remote sensing image data is from June to July, and data with no clouds or few clouds and high image quality are selected. The preprocessing of the image data includes radiometric calibration, atmospheric correction, and image mosaicking and cropping, etc.

[0079] Step 4: Standardize the evaluation indicators:

[0080] Indicators of different categories have different dimensions. By means of standardization, the influence of dimensions among various data is eliminated, enabling the evaluation process to be carried out under the same standard. In order to quantify the indicators and make the evaluation process carried out under the same standard for better statistics, this embodiment selects the hundred-mark system [0, 100] for score assignment, and the scores of each indicator are determined according to the attenuation degree of its impact on cultivated land quality and the Delphi method.

[0081] Among them, the classification criteria for slope and irrigation conditions refer to the "Regulations for the Classification of Agricultural Land Quality" (GB / T 28407-2012). The anomaly vegetation index is divided by the natural breakpoint method, and the fragmentation index is normalized to [0, 100] according to the upper and lower limits of its function. The specific evaluation index system is as Figure 2 shown.

[0082] Step 5: Detection of data outliers: Detect and mark the outliers in the data standardized in Step 4:

[0083] If there is a problem of class imbalance in the training samples, it will seriously affect the accuracy of the model. Therefore, it is necessary to detect these outlier samples, use the method of Mahalanobis distance to detect the outliers, and mark the outliers as a class.

[0084] Step 6: Construct a competitive neural network model:

[0085] Input the data standardized in Step 4 into the network for model training, and then output the clustering results. As Figure 3 shown, the structure of the competitive neural network model in this embodiment is as Figure 4 shown.

[0086] Step 7: Calculate the natural quality score.

[0087] Calculate the natural quality scores of each evaluation unit.

[0088] According to the "Regulations for the Classification of Agricultural Land Quality" (GB / T 28407-2012), use the geometric mean method to calculate the natural quality score C of the evaluation unit Li .

[0089] Step 8: Category evaluation.

[0090] Statistically analyze the average value of the natural quality scores of the evaluation units under each category obtained in Step 5 and Step 6. According to the size of the average value, conduct grade evaluation and statistical analysis for each evaluation grade. The cultivated land quality scores of each category are a = 86.1616, b = 78.0951, c = 66.3721, d = 54.3833, corresponding to the 4th grade, 5th grade, 6th grade, and 7th grade in the cultivated land grade results respectively. The results are as Figure 5 shown.

[0091] Step 9, Precision evaluation.

[0092] Combined with the results of the county-level cultivated land quality grades in Luxian County, Sichuan Natural Resources Department in 2018, as Figure 6 shown, the evaluation results were verified pixel by pixel using the confusion matrix method, and the results are as Figure 7 shown.

[0093] Through Figure 7 it can be seen that the overall precision of the evaluation in this embodiment is 85.5410%, and the kappa coefficient is 0.7111, which exceeds 0.7. The verification results indicate that it is feasible to use remote sensing data for cultivated land quality evaluation in the present invention, and the unsupervised competitive neural network does not require artificial determination of weights, simplifies the evaluation process, and the results are more objective and scientific. Generally speaking, it is reasonable and reliable to use remote sensing data and competitive neural network for cultivated land quality evaluation in the present invention.

Claims

1. A method for evaluating the natural quality of cultivated land based on all-remote sensing data-driven, characterized in that, It includes the following steps: Step 1. Construct a remote sensing index system for cultivated land quality: Based on the consideration and analysis of the connotation of cultivated land quality, following the principles of unity, dominance, sensitivity, practicality, independence, and stability in selecting evaluation indicators, and considering the availability of remote sensing evaluation indicators, the remote sensing evaluation indicators are divided into two categories: one is to directly use multi-source remote sensing to obtain cultivated land quality monitoring and evaluation indicators; the other is to indirectly reflect the cultivated land quality status by remotely sensing the growth of crops. Specifically, from two aspects of natural quality and utilization quality, it includes 4 evaluation indicators: slope, anomaly percentage of normalized difference vegetation index, irrigation condition, and fragmentation degree; Step 2. Determine the evaluation unit of cultivated land quality: The grid method is used to divide the evaluation unit, and the grid area should be smaller than the average area of the units divided in the cultivated land quality grade map of the study area to achieve the unity of different evaluation units; Step 3. Obtain evaluation index data: Use remote sensing technology to obtain evaluation index data; Step 4. Standardize the evaluation indicators: Through standardization processing, the influence of the dimension of various remote sensing data obtained in Step 3 is eliminated to obtain training data, so that the evaluation process is carried out under the same standard; Step 5. Detect data outliers: Detect the outlier samples in the training data obtained in Step 4 and mark the outliers as one category; Step 6. Construct a competitive neural network model: 1) Initialize the network; Normalize the training data obtained in step 4, and let the training data be x ij (i = 1, 2, …, m; j = 1, 2, …, n), and the normalization formula is as follows: where m is the number of evaluation units, n is the number of evaluation indicators, and x j(min) is the minimum value of the j-th evaluation indicator among the m evaluation units, and x j(max) is the maximum value of the j-th evaluation indicator among the m evaluation units. X ij corresponds to an n-dimensional vector representing the cultivated land quality; Generate random numbers in the interval [0, 1] as the initial weight vector w of the network; 2) Obtain the winning neuron; For an input pattern X = {x1, x2, …, x n} of the input network, similarity tests are performed between the weight vectors w corresponding to all neurons in the competition layer and X, and the w most similar to X is determined as the winning neuron; the link distance is calculated using the linkdist function as a measure of similarity, and the link distance L r between the n-dimensional vector X s and X rs is calculated as follows: If i = j, then L rs = 0; If X r and X s the Euclidean distance e <= 1, then L rs = 1; If there exist k1, k2 belonging to n such that then L rs = 3; If there exist k1, …, k N belonging to n such that then L rs = N; If none of the above conditions are met, then L rs = n; 3) Adjust the weights; The competitive neural network adjusts the weights of the winning neuron according to the Kohonen learning rule. For the winning neuron v, the corresponding weight w vj The change amount Δw vj is defined as follows: η is the learning rate, defaulting to 0.01; Through learning, the winning neuron will have a greater chance of winning when the next similar input vector appears, while the neuron that lost the competition in the previous round will have a smaller chance of winning; 4) Network output; Repeat Step 2) to 3) until the number of units in each category no longer changes, and output the result; Step 7. Calculate the natural quality score; According to the "Regulations for the Classification of Agricultural Land Quality" (GB / T 28407-2012), the natural quality score C of the evaluation unit is calculated using the geometric mean method Li , and the formula is as follows: Where: C Li is the natural quality score of agricultural land for the grading unit; i is the evaluation unit number; j is the evaluation index number; m is the number of evaluation indexes; f i is the score of the j-th evaluation index in the i-th evaluation unit, and the value range is (0 to 100]; Step 8. Category evaluation; Statistically calculate the average value of the natural quality scores of the evaluation units in each category obtained in Step 5 and Step 6. According to the size of the average value, conduct grade evaluation and statistical analysis for each evaluation grade.

2. The cultivated land natural quality evaluation method based on all-remote sensing data drive according to claim 1, characterized in that: When dividing the evaluation unit by the grid method in Step 2, the grid area is equal to the minimum resolution of the obtained remote sensing data.

Images