Dynamic evaluation method for regional soil environment quality

Abstract

The invention discloses a dynamic evaluation method for regional soil environment quality. According to the method, multi-temporal soil sample actual measurement data, remote sensing image data, land utilization data and environmental assistance data of a target area are acquired, the multi-source data are preprocessed and fused, and a time sequence Kriging algorithm interpolation is combined with a machine learning regression model; and generating a soil environment index raster data set with high temporal-spatial resolution. And a mixed deep learning model combining a long-short-term memory network and a convolutional neural network is constructed to learn a space-time evolution rule of the soil environment index. And calculating a soil environment quality comprehensive index by using the model, generating a quality grade spatial distribution diagram, and analyzing a change trend. Furthermore, the prediction of the regional soil environment quality can be realized by inputting future scene parameters. According to the method, dynamic evaluation and prediction of the regional soil environment quality are realized, the time-space continuity and precision of evaluation are improved, and the method has early warning capability.

Description

Technical Field

[0001] This invention relates to the field of environmental monitoring and geographic information technology, and in particular to a method for dynamically evaluating and predicting changes in regional soil environmental quality by combining multi-source data fusion, spatiotemporal modeling and deep learning technologies. Background Technology

[0002] Soil is a vital component of the ecosystem, and its quality directly impacts agricultural product safety, human health, and ecological balance. Accurately understanding the spatiotemporal patterns of regional soil environmental quality is crucial for implementing precise soil pollution control and sustainable land use.

[0003] Currently, traditional methods for assessing regional soil environmental quality mainly rely on field sampling and laboratory analysis. As Li Chunhui et al. pointed out in "Research Progress on Soil Environmental Quality Assessment Methods [J]" Soil, 2019, 51(1): 1-10, although this method is accurate, it has limitations such as high cost, long cycle, and discrete sampling points, making it difficult to achieve large-scale, high-frequency dynamic monitoring, and the evaluation results are mostly static or phased.

[0004] To overcome the aforementioned problems, related research has begun to explore technical approaches that integrate remote sensing and models. For example, patent CN112270668A, "A Soil Property Inversion Method Based on Multi-Source Remote Sensing," discloses a scheme for inverting soil properties using multi-source remote sensing data. However, it relies on a single remote sensing inversion model, making it susceptible to interference from vegetation cover and atmospheric conditions, resulting in limited inversion accuracy and reliability. Another patent, CN113657876A, "A Soil Heavy Metal Prediction Method Combining Spatiotemporal Interpolation and Machine Learning," proposes combining Kriging interpolation with a random forest model for soil heavy metal prediction. However, its technical solution focuses on static spatial distribution prediction and does not involve dynamic modeling of time series data or prediction of future scenarios.

[0005] In the area of ​​dynamic modeling, academic research such as "A ConvLSTM-based model for spatiotemporal prediction of soil moisture[J]" Agricultural Water Management, 2022, 273:107896 explored the use of ConvLSTM models for soil moisture prediction, verifying the potential of deep learning in the spatiotemporal prediction of soil parameters. However, this research is limited to a single soil parameter and has not constructed a complete technical system integrating multi-source data fusion, dynamic evaluation, and scenario prediction to solve the comprehensive evaluation problem of regional soil environmental quality.

[0006] In summary, existing technologies have not yet solved the technical problem of how to effectively integrate point-based measured data with area-based remote sensing data, and on this basis, construct a comprehensive evaluation model that can simultaneously characterize spatiotemporal dynamic changes and has the ability to predict future scenarios. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for evaluating changes in regional soil environmental quality. This method can achieve low-cost, high-precision, dynamic and predictable regional soil environmental quality evaluation.

[0008] To achieve the above objectives, the present invention adopts the following technical solution.

[0009] A method for evaluating changes in regional soil environmental quality, characterized by comprising the following steps.

[0010] S1, acquire multi-source data of the target area at multiple time points. The multi-source data includes at least measured soil sample data, remote sensing image data, land use data, and environmental auxiliary data. For example... Figure 1 As indicated by label ①, this step involves inputting the target region and time series, and acquiring the above four types of multi-source data in parallel.

[0011] S2, preprocessing and fusing the multi-source data to generate a multi-temporal soil environmental index raster dataset covering the target area. The fusion process utilizes spatiotemporal interpolation algorithms and machine learning regression models to correlate and correct the point-like measured soil sample data with the area-like remote sensing image data. Specifically, as... Figure 2 As shown, the measured soil sample data are first preprocessed, such as logarithmic transformation, to satisfy the normal distribution assumption. Then, time-series Kriging is used for interpolation to obtain a preliminary raster field. Simultaneously, features such as normalized vegetation index and building index are extracted from the preprocessed remote sensing imagery, and a feature variable set is constructed by combining land use type and environmental auxiliary data. Then, using the interpolated preliminary raster field as training labels, machine learning regression models such as random forest, gradient boosting tree, or support vector machine are employed to learn the nonlinear relationship between remote sensing features and soil environmental indicators. This allows for the inference of areas without samples across the entire region, ultimately generating a high spatiotemporal resolution raster dataset of soil environmental indicators, including heavy metal content, soil pH, and organic matter content. The accuracy of this process is verified using reserved validation points, with a root mean square error of 0.08 mg / kg and a coefficient of determination of 0.89.

[0012] S3. Construct a dynamic evaluation model for soil environmental quality. This model is based on time series analysis and spatial deep learning networks to learn the evolution characteristics of soil environmental quality in the spatiotemporal dimensions. Specifically, the dynamic evaluation model for soil environmental quality is a hybrid model combining long short-term memory networks and convolutional neural networks. Figure 3 As shown, the model input is a spatiotemporal data cube composed of multi-temporal soil indicator raster data. A Long Short-Term Memory (LSTM) network is used to extract the long-term dependencies of soil environmental indicators' time series at each spatial location. A Convolutional Neural Network (CNN) extracts the correlation features of the soil environmental indicators in their spatial neighborhood using 3x3 convolutional kernels. The output layer fuses spatiotemporal features for dynamic evaluation. The model training uses the Adam optimizer with an initial learning rate of 0.001, employs the Huber loss function, and sets the batch size to 32. Testing shows that the model's mean absolute error for predicting the next step is less than 0.05.

[0013] S4. Input the soil environmental index raster dataset into the trained dynamic evaluation model to calculate the spatial distribution map of the comprehensive soil environmental quality index of the target area at different time points, and perform quality level classification and trend analysis accordingly. The comprehensive soil environmental quality index is calculated using the Nemerow comprehensive index method or the comprehensive index method based on the entropy weight method, and the quality level is classified according to the national standards "Soil Environmental Quality Standard for Agricultural Land Soil Pollution Risk Control (Trial)" (GB 15618-2018) or "Soil Pollution Risk Control Standard for Construction Land (Trial)" (GB 36600-2018).

[0014] S5. Based on the dynamic evaluation model, input future scenario parameters to predict the soil environmental quality status of the target area at future points in time. The future scenario parameters include different land use change plans, pollutant emission reduction schemes, and climate change prediction data. Normalization processing is required before input to match the model input scale.

[0015] Preferably, the method further includes step S6, which involves constructing a visual interactive platform to display the spatial distribution map of the comprehensive soil environmental quality index, the quality grade classification results, the change trend analysis report and the future prediction scenario, and to provide spatial query and timeline backtracking functions.

[0016] The beneficial effects of this invention are as follows.

[0017] By fusing multi-source data, the accuracy of point-based measured data and the coverage advantage of area-based remote sensing data were combined to generate a high spatiotemporal resolution full-coverage raster dataset, which improved the spatiotemporal continuity and accuracy of the evaluation.

[0018] By constructing a dynamic evaluation model that combines long short-term memory networks and convolutional neural networks, we have achieved in-depth exploration of the spatiotemporal evolution of soil environmental quality, completing the leap from static description to dynamic simulation.

[0019] By inputting future scenario parameters into a trained dynamic evaluation model, a forward-looking prediction of changes in regional soil environmental quality is achieved, thus enabling early warning capabilities.

[0020] The entire methodology is highly automated and intelligent, and provides intuitive and scientific decision support for soil pollution prevention and land resource management through a visual interactive platform. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the overall process of the method of the present invention. Figure 2 A schematic diagram illustrating the generation of a soil index raster dataset through multi-source data fusion; Figure 3 This is a schematic diagram of the structure of a spatiotemporal hybrid deep learning model; Figure 4 Example diagram showing the spatiotemporal variation of regional soil environmental quality levels. Detailed Implementation

[0022] The present invention will be further described below with reference to the accompanying drawings and embodiments. These embodiments are implemented based on the technical solution of the present invention, providing detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0023] Example

[0024] An area of ​​approximately 500 square kilometers in the suburbs of a certain city was selected as the implementation target. The soil environmental quality of the area in 2010, 2015 and 2020 was dynamically evaluated and predicted.

[0025] S1, Data Acquisition and Preprocessing.

[0026] Multi-source data for the target area at three time points were acquired. Soil sample data were obtained from 150 evenly distributed sampling points within the area, with cadmium, arsenic, lead, and mercury content measured at each point. Remote sensing imagery data used was Landsat satellite imagery from the same period, with a spatial resolution of 30 meters. Environmental auxiliary data included annual land use data, distribution data of key industrial enterprises, and average annual precipitation data. All data were preprocessed using the CGCS2000 coordinate system and a unified spatial resolution of 30 meters.

[0027] S2, multi-source data fusion and raster dataset generation.

[0028] This step corresponds to Figure 1 Label ② and Figure 2 The process is illustrated below. First, the soil heavy metal content data is logarithmically transformed to satisfy the normal distribution assumption of the time-series kriging method, and interpolation is performed on discrete sample points to obtain a preliminary raster field of heavy metal content for each year. Then, normalized vegetation index and building index are extracted from the preprocessed remote sensing image, and a feature variable set is constructed by combining land use type and Euclidean distance of raster cells from the pollution source. Using the time-series kriging interpolation results as the training objective, a random forest regression model is used to learn the nonlinear relationship between remote sensing features and soil heavy metal content, thereby inferring the heavy metal content in areas without samples across the entire region. Finally, four heavy metal content raster datasets covering the entire region, with a spatial resolution of 30 meters and containing three time nodes are generated. Validated using 30 reserved validation points, the root mean square error of cadmium content prediction is 0.08 mg / kg, and the coefficient of determination reaches 0.89, significantly outperforming the results obtained by using the time-series kriging method alone.

[0029] S3, Dynamic Evaluation Model Construction.

[0030] This step corresponds to Figure 1 Label ③ and Figure 3 The model structure is shown. A convolutional long short-term memory network is constructed as the dynamic evaluation model for this implementation. The model input is a multi-channel spatiotemporal data cube consisting of raster data of four heavy metal contents at three time points and land use type data encoded using one-heat encoding. In the model structure, the convolutional layers use 3x3 convolutional kernels to extract spatial features, the long short-term memory network layer is used to learn long-term dependencies in the temporal dimension, and the fully connected output layer is used to output the comprehensive soil environmental quality index of each raster unit. During the model training phase, the data from the first two time points are used as input, and the data from the third time point is used as the prediction target. The Adam optimizer is used, with an initial learning rate of 0.001, the Huber loss function is used, and the batch size is set to 32. After testing, the average absolute error of the model for predicting the next step is less than 0.05.

[0031] S4, Dynamic Evaluation and Visualization of Soil Environmental Quality.

[0032] This step corresponds to Figure 1 Label ④ and Figure 4 The visualization effect is shown. The generated raster datasets from the three time points are input into a fully trained dynamic evaluation model, and the Nemerow composite index method is used to calculate the comprehensive soil pollution index for each raster cell. Based on the "Soil Environmental Quality Agricultural Land Soil Pollution Risk Control Standard (Trial Implementation)," the composite index is divided into five risk levels: safe, warning, slightly polluted, moderately polluted, and heavily polluted. Soil environmental quality risk level distribution maps for the three years are generated and dynamically displayed through a geographic information system platform. Figure 4As shown, comparative analysis clearly reveals the diffusion trend of heavy metal pollution in the soil of this region from the urban industrial area to the southeastern suburbs between 2010 and 2020, with a diffusion rate of approximately 150 meters per year.

[0033] S5, Future Scenario Prediction.

[0034] This step corresponds to Figure 1 Label ⑤. Assuming that a new industrial zone will be added to the eastern part of the region in 2025 according to urban planning, this land use change scenario is digitized and normalized along with other environmental auxiliary data to serve as parameters for the future scenario. The processed parameters are then input into a pre-trained convolutional long short-term memory network model to predict the soil environmental quality status of the target area in 2025. The prediction results indicate that, compared to 2020, the cadmium pollution level in the soil within approximately three kilometers downwind of the new industrial zone is expected to increase by one level. This prediction provides clear risk warnings and decision-making basis for regional planning and environmental protection departments.

[0035] S6, a visual interactive platform.

[0036] This step corresponds to Figure 1 Label ⑥. A visual interactive platform is constructed based on network geographic information system technology. This platform integrates all the above-mentioned spatiotemporal distribution maps, classification results, and future prediction scenarios. Users can use the platform to perform spatial range queries, switch between multiple temporal layers, and perform timeline backtracking operations, realizing dynamic demonstration and analysis of soil environmental quality.

[0037] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for dynamic evaluation of regional soil environmental quality, characterized in that, Includes the following steps: S1: Acquire multi-source data of the target area at multiple time points, including measured soil sample data, remote sensing image data, land use data, and environmental auxiliary data; S2: The multi-source data is preprocessed and fused to generate a multi-temporal soil environmental index raster dataset covering the target area; wherein, the fusion process first uses the temporal kriging algorithm to interpolate the measured data of soil samples, and then uses the interpolation results as training labels to use a machine learning regression model to associate remote sensing image data features to generate a high spatiotemporal resolution raster dataset. S3: Construct a dynamic evaluation model for soil environmental quality. The model is a hybrid structure combining long short-term memory network and convolutional neural network, used to learn the long-term dependence of soil environmental indicators in the time dimension and the neighborhood correlation features in the spatial dimension. S4: Input the raster dataset into the trained dynamic evaluation model, calculate the comprehensive index of soil environmental quality, generate a spatial distribution map, and perform quality level classification and trend analysis. S5: Based on the dynamic evaluation model, input future scenario parameters to predict the soil environmental quality status of the target area at future time points.

2. The method according to claim 1, characterized in that, In step S2, the time-space kriging algorithm is applicable to non-stationary spatiotemporal processes; the machine learning regression model is a random forest, gradient boosting tree, or support vector machine; the soil environmental indicators include heavy metal content, soil pH value, and organic matter content.

3. The method according to claim 1, characterized in that, In step S3, the long short-term memory network extracts the time-series features of each grid cell, the convolutional neural network extracts spatial features through 3×3 convolutional kernels, and the output layer fuses spatiotemporal features for dynamic evaluation.

4. The method according to claim 1, characterized in that, In step S4, the comprehensive soil environmental quality index is calculated using the Nemerow comprehensive index method or the entropy weight method, and the quality grade classification is based on GB 15618-2018 or GB 36600-2018 standards.

5. The method according to claim 1, characterized in that, In step S5, the future scenario parameters include land use change planning maps, pollutant emission reduction scheme indicators, and climate change prediction data, which need to be normalized before input to match the model input scale.

6. The method according to claim 1, characterized in that, The method further includes: S6: Construct a visual interactive platform that integrates the spatial distribution map, quality level classification results, and future prediction scenarios, and provides functions such as spatial range query, multi-temporal layer switching, and timeline backtracking.

Images