CN-121983168-A - Multi-index synchronous monitoring method for soil pollution of farmland in plateau lake basin

CN121983168ACN 121983168 ACN121983168 ACN 121983168ACN-121983168-A

Abstract

The invention relates to the technical field of environmental monitoring, and discloses a multi-index synchronous monitoring method for soil pollution of farmland in a plateau lake basin. The method is applicable to the farmland of the plateau lake river basin, under the conditions that the measured data of the farmland soil of the plateau lake river basin are high in acquisition cost and difficult to acquire in large quantity, the satellite remote sensing means is used for assisting the limited field monitoring data, the spatial distribution map of the soil surface source pollution source load index of the farmland is acquired in a large scale and low cost and high efficiency, the visualization, quantification, traceability and prevention and control of the surface source pollution are realized through the long-term soil pollution index content, variable quantity monitoring and risk level evaluation, the limitation of the conventional point position measurement on the surface of the point is overcome, and the method is the most efficient method for evaluating the potential risk of the agricultural surface source pollution.

Inventors

TANG BOHUI
JI XINRAN
DENG CHANGJUN
HUANG LIANG
CHEN JUNYI
ZHANG ZHEN
LI MENGHUA
ZHU BAI

Assignees

昆明理工大学

Dates

Publication Date: 20260505
Application Date: 20260104

Claims (8)

1. The multi-index synchronous monitoring method for soil pollution of the farmland in the plateau lake basin is characterized by comprising the following steps of: s1, acquiring actual measurement data of soil pollution detection indexes of farmland in a plateau lake basin; S2, acquiring soil forming factor data synchronous with measured data of soil pollution detection indexes of farmland in a plateau lake basin; S3, acquiring altitude lake basin cultivated land extraction data; S4, constructing a multi-task prediction model of the soil pollution detection index of the farmland in the plateau lake basin based on the actual measurement data of the soil pollution detection index of the farmland in the plateau lake basin and the synchronous soil forming factor data; s5, inputting the data extracted from the highland lake basin cultivated land to a multitasking prediction model of the highland lake basin cultivated land soil pollution detection index, and obtaining a spatial distribution diagram of key soil pollution indexes of the highland lake basin cultivated land, an annual change diagram of key soil pollution indexes of the highland lake basin cultivated land and a source pollution risk level distribution diagram of the highland lake basin cultivated land.
2. The method for synchronously monitoring the soil pollution of the plateau, lake and river basin cultivated land with multiple indexes is characterized in that the step S1 is to acquire actual measurement data of soil pollution detection indexes of the plateau, lake and river basin cultivated land, wherein the actual measurement data of the soil pollution detection indexes comprise the contents of organic matters, nitrogen, phosphorus and potassium in the soil after the soil samples are air-dried.
3. The multi-index synchronous monitoring method for soil pollution of the plateau lake basin cultivated land is characterized in that soil forming factor data synchronous with measured data of soil pollution detection indexes of the plateau lake basin cultivated land is obtained in the step S2, and the soil forming factor data comprises soil matrix data, soil property data, annual average surface temperature data, annual average precipitation data, topographic variable data and vegetation information; the soil matrix data and the soil property data have the following expression: The soil property data comprise NDVI, NBR2, NDMI, CMSI, CI, near infrared wave band in NIR multispectral remote sensing image data, red light wave band, SWIR1, SWIR2, blue light wave band, green light wave band, wherein the NDVI, the NBR2, BSI, SBI, NDMI and the CMSI are respectively represented by the following formula, the NDVI is a normalized vegetation index, the BSI is a bare soil index, the SBI is a surface brightness index, the NDMI is a normalized water index, the CMSI is a crop water stress index, the CI is a clay index, the near infrared wave band in the NIR multispectral remote sensing image data is represented by the Red light wave band, the SWIR1 is a short wave infrared 1 wave band, the SWIR2 is a short wave infrared 2 wave band, the Blue light wave band and the Green is a Green light wave band; The annual average earth surface temperature data are firstly obtained, the annual average earth surface temperature is processed into orthorectified earth surface temperature, images with annual cloud cover less than a set value in a sampling year are screened, and the images are converted into earth surface temperature data in the unit of degrees centigrade to obtain annual average earth surface temperature data; The annual average precipitation data are obtained by acquiring a annual month precipitation data set of sampling years to obtain annual average precipitation data; The terrain variable data comprise gradient, vector bumpiness, water flow power index, terrain bumpiness index, composite terrain index, roughness, terrain position index and elevation; the vegetation information comprises EVI and LSP, wherein EVI is an annual enhancement type vegetation index, and LSP represents a vegetation weather index.
4. The method for synchronously monitoring soil pollution of the plateau lake basin cultivated land according to claim 1, wherein the S4 multitask prediction model of the soil pollution detection index of the plateau lake basin cultivated land comprises a CNN branch and an LSTM branch, and the output of the two branches is fused and predicted to obtain a predicted value, wherein a channel attention mechanism is introduced into the CNN branch and a time step attention mechanism is introduced into the LSTM branch; the multi-task prediction function of the multi-task prediction model of the plateau lake basin cultivated land soil pollution detection index is as follows: Wherein, the Represents the soil pollution detection index of any cultivated land, () Is a function which is composed of 7 soil forming factors and represents the soil pollution detection index of any cultivated land, Data representative of the nature of the soil, Representing climate, including annual average surface temperature data and annual average precipitation data, Representing vegetation information is provided which is representative of the vegetation information, Representing the data of the topographical variable, Representing the data of the matrix of the soil, The time is represented by the time period of the time, The seven soil forming factors are used for comprehensively representing the soil forming environment of the research area; Inputting x_cnn_common to the CNN branch, outputting f_cnn, inputting x_ts_ evi _lsp to the LSTM, outputting f_ LSTM, and splicing the output f_cnn of the CNN branch and the output f_ LSTM of the LSTM branch to form a 32-dimensional fusion feature The specific expression is as follows: f=[f cnn ; f lstm ]∈R 32 Wherein, the Representing 32-dimensional fusion characteristics, [ ] represents vector splicing operation, f cnn is the output of the CNN branch, and f lstm is the output of the LSTM branch; Entering a final full-connection regression layer, and outputting a predicted value 。
5. The method for synchronously monitoring soil pollution of the plateau lake basin cultivated land according to claim 4, wherein the CNN branch extracts soil property data, soil matrix data, annual average surface temperature data, annual average precipitation data and topography variable data as x_cnn_common, and inputs the x_cnn_common into the CNN branch to obtain output data 16-dimensional representation f_cnn, and the method comprises the following specific steps: S4a.1, x_cnn_common is taken as input, and is subjected to two-layer convolution and ReLU operation, and then subjected to 2X 2 max pooling operation; s4a.2, connect SEBlock channel attention; s4a.3, flattening and then obtaining 16-dimensional characterization f_cnn E R 16 through one-layer full connection; the x_cnn_common shape is: (B, E, H, W) Wherein E is the number of image channels, B is the batch size, and H and W are the height and width of the feature map respectively.
6. The multi-index synchronous monitoring method for soil pollution of plateau lake basin cultivated land of claim 4, wherein a channel attention mechanism is introduced into the CNN branch, and a global average pooling operation is performed on the channels in the CNN branch: In the formula, Is a channel Is used to determine the global average value of (c), And Is the spatial height and width of the feature map, Representation channel In position Pixel values of (2); For the characteristic diagram Summing the heights; For the characteristic diagram Summing the heights; The dependence and output weight between two layers of full-connection modeling channels: In the formula, E R C is a descriptor of all channels, And Is a weight matrix of a full connection layer, For the function to be activated by the ReLU, The function is activated for Sigmoid, E (0, 1) C represents the attention weight of each channel; the feature map for each channel is scaled by weight: In the formula, Representative channel The re-weighted feature map is used to determine, Representative channel Updated attention weights; Is a channel The feature map before weighting.
7. The method for synchronously monitoring soil pollution of farmland in plateau, lake and river basin according to claim 4, wherein the LSTM branch extracts time characteristics of time sequence EVI and LSP dynamic variable as x_ts_ EVI _lsp, and inputs x_ts_ EVI _lsp into the LSTM branch, and the specific steps of obtaining the 16-dimensional characterization f_ LSTM are as follows: s4b.1, inputting and sending LSTM with the time sequence of EVI and LSP; S4b.2, receiving the attention of the time step, and carrying out weighted summation on hidden states of all the time steps to obtain a context vector; S4b.3, obtaining 16-dimensional characterization f_ lstm epsilon R 16 through one-layer full connection; the x_ts_ evi the_lsp shape is: (B, T, D) Wherein, the In order to make the number of time steps, Dimension size for the input feature; B is batch size; = + Wherein, the An EVI feature dimension for input; is the input LSP feature dimension.
8. The method for synchronously monitoring soil pollution of farmland in plateau and lake basin according to claim 4, wherein the LSTM branch is characterized in that the time steps are scored in the LSTM branch: In the formula, Representing time steps Is used to calculate the score of (a), Representing LSTM at time step Is used to determine the hidden state vector of (1), Representing a learnable attention weight vector; softmax normalization converts the original scoring into probability distributions: In the formula, Representing the attention weight of the time step t, A score representing time step t; Representing a time step index; for all time steps Is the sum of (3); is an exponential function; The weighted convergence is a context vector: In the formula, The context vector is represented by a vector of the context, Representing time steps Is used for the concentration weight of the person, A hidden state vector representing LSTM at time step t; Representing the number of time steps; For all steps of time from 1 to T Is a sum of (a) and (b).

Description

Multi-index synchronous monitoring method for soil pollution of farmland in plateau lake basin Technical Field The invention relates to the technical field of environmental monitoring, in particular to a multi-index synchronous monitoring method for soil pollution of farmland in a plateau lake basin. Background The importance of farmland soil pollution index monitoring and the bottleneck of the prior art are that the farmland soil health is ensured and the agricultural environmental pollution is prevented. Agricultural pollution mainly comprises point source pollution and non-point source pollution, wherein the non-point source pollution such as eutrophication caused by the runoff loss of chemical fertilizers and pesticides has the characteristics of dispersed emission, strong concealment, high randomness and the like, and the monitoring and treatment difficulty is far higher than that of the point source pollution. Currently, monitoring of soil pollution indicators in arable land and pollution risk assessment derived therefrom rely mainly on in-situ sampling and laboratory chemical analysis. Although the method is high in accuracy, the following inherent defects exist: 1) The cost is high, the time and the effort are wasted, a large amount of manpower, material resources and time are required for sample collection and laboratory analysis, and the frequent and large-scale development is difficult; 2) The method takes points as a surface and has limited representativeness, wherein discrete sampling points are difficult to comprehensively and continuously reflect the spatial heterogeneity of soil nutrients and pollutants in the whole area, and the distribution range and the degree of agricultural non-point source pollution cannot be accurately depicted; 3) The timeliness is poor, namely the period from sampling to obtaining a result is long, so that monitoring data is seriously lagged, and real-time data support cannot be provided for rapid early warning and accurate planning of environmental pollution; The development and the existing challenges of the remote sensing technology are that in order to break through the limitations of the traditional method, the remote sensing technology gradually becomes a main stream means for monitoring the physical and chemical properties of soil in a long-term and large range due to the advantages of macroscopic, rapid and dynamic monitoring, but the remote sensing technology still has significant defects: 1) The traditional method depends on single spectrum information of a single satellite platform, is difficult to capture deep and nonlinear association between physical and chemical properties of soil and complex spectrum characteristics, and has limited accuracy and applicability; 2) Optical remote sensing, such as Landsat, sentinel-2, is easily interfered by cloud rain, and data loss is serious due to the cloudy climate of the river basin of the plateau lake; 3) Radar remote sensing, such as Sentinel-1, has penetration capability, but cannot independently characterize the biochemical characteristics of soil carbon; 4) The model has insufficient adaptability, and is characterized in that a geostatistical method, such as a Kriging interpolation method, depends on sampling point density, and a sparse sample leads to error amplification; 5) The special problem of the plateau area is that the plateau lake basin is broken in topography, the cultivated land is dispersed, the soil types are various, such as brown calcium soil and chestnut calcium soil, and the suitability of the area of the existing model is poor; the development and limitation of the deep learning application are that the development of the deep learning improves the remote sensing estimation precision of the physical and chemical properties of the soil, but the cooperative optimization of a plurality of methods is still insufficient: 1) Single model dominant Random Forest (RF), support Vector Machine (SVM) and the like are widely adopted, such as SOC prediction in black soil area of Jilin province, but characteristic extraction relies on manual design, and multi-source data deep association is difficult to capture; 2) The fusion of visible-near infrared (VNIR) and hyperspectral imaging (HSI) can improve the precision, but the ecological process interpretation ability is weak due to the fact that the carbon circulation mechanism parameters such as the soil respiration rate are not integrated, and the suitability of the ecological process in a plateau lake basin is yet to be explored; 3) The existing method has the defects of multi-focus static estimation and timing modeling capability, particularly difficulty in synchronously predicting various indexes, and incapability of directly serving comprehensive evaluation of pollution risks; Particular challenges of plateau lake basin long-term and large-scale monitoring of soil pollution index of farmland in plateau lake basin faces unique challenges: 1) The data acq