Water and soil loss dynamic identification method and system based on deep learning and multi-source remote sensing data

Abstract

The invention discloses a water and soil loss dynamic identification method and system based on deep learning and multi-source remote sensing data, and belongs to the technical field of water and soil loss dynamic identification. The method comprises the following steps: S1, acquiring multi-source remote sensing data of a to-be-monitored area; s2, a dynamic adaptive feature extraction network DAFEN is constructed, the network comprises a plurality of feature recognition layers, each layer dynamically generates a recognition unit to adapt to spatial distribution and feature complexity of input data, and feature optimization is performed through bidirectional information flow; s3, performing feature extraction on the multi-source remote sensing data by using DAFEN, and outputting a multi-dimensional feature map including land utilization types, vegetation coverage and terrain factors; and S4, based on the multi-dimensional feature map, in combination with an improved soil erosion intensity calculation model, calculating the soil erosion intensity grade of each pixel at the time t, and generating a water and soil loss dynamic monitoring result map, thereby significantly improving the accuracy and robustness of monitoring.

Description

Technical Field

[0001] This invention relates to the field of dynamic soil erosion identification technology, and more specifically to a method and system for dynamic soil erosion identification based on deep learning and multi-source remote sensing data. Background Technology

[0002] Soil erosion is a global ecological and environmental problem that not only leads to land degradation and decreased productivity, but also triggers a chain reaction of siltation and water pollution, posing a serious threat to regional sustainable development. Timely and accurate understanding of the status, dynamic trends, and spatial distribution characteristics of soil erosion is fundamental to the scientific implementation of soil and water conservation work.

[0003] Traditional methods for monitoring soil erosion mainly rely on field measurements and manual surveys. While these methods offer high accuracy, they are costly, inefficient, and difficult to implement for large-scale, high-frequency dynamic monitoring. With the development of remote sensing technology, image-based methods for monitoring soil erosion have gradually become mainstream. These methods primarily extract key parameters such as vegetation cover, land use type, and topographic factors, and then combine them with soil erosion models (such as the Universal Soil Loss Equation (USLE) and its revisions) for evaluation.

[0004] However, a deep learning-based method for monitoring regional soil erosion dynamics, proposed in CN111738561B, uses the HRNet network for automatic land use classification and combines factors such as slope and vegetation cover to determine soil erosion intensity. While this method has made some progress in automation, its feature extraction network structure is fixed and difficult to adapt to the differences in data distribution across different regions and time periods. Furthermore, its soil erosion model still uses a static judgment table and lacks quantitative consideration of dynamic factors.

[0005] Therefore, how to provide a method and system for identifying soil erosion that can adapt to multi-source data features and achieve fully automatic dynamic monitoring is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] In view of this, the present invention provides a method and system for dynamic identification of soil erosion based on deep learning and multi-source remote sensing data, in order to meet the urgent needs of current ecological environment management.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: A method for dynamic identification of soil erosion based on deep learning and multi-source remote sensing data includes: S1: Acquire multi-source remote sensing data of the area to be monitored; S2: Construct a dynamic adaptive feature extraction network DAFEN, which includes multiple layers of feature recognition layers. Each layer dynamically generates recognition units to adapt to the spatial distribution and feature complexity of the input data, and performs feature optimization through bidirectional information flow. S3: Use the dynamic adaptive feature extraction network DAFEN to extract features from multi-source remote sensing data and output a multi-dimensional feature map including land use type, vegetation cover, and topographic factors. S4: Based on the multi-dimensional feature map and combined with the improved soil erosion intensity calculation model, calculate the soil erosion intensity level of each pixel at time t, and generate a dynamic monitoring result map of soil and water loss.

[0008] Preferably, the specific steps for constructing the dynamic adaptive feature extraction network DAFEN in S2 include: S21: Design a network hierarchy, including an input layer, a feature extraction layer, and an output layer. The feature extraction layer consists of multiple sub-networks, each of which is responsible for extracting features at different levels of abstraction. S22: During the training process, recognition units are dynamically generated based on the distribution of input data, and parameters are optimized through top-down and bottom-up information flows; S23: Train the network using multi-source remote sensing data, and adjust the network parameters using loss function and accuracy evaluation index until the network converges; The dynamic adaptive feature extraction network DAFEN uses HRNet as its basic architecture and introduces a dynamic unit generation mechanism in the feature extraction path.

[0009] Preferably, the calculation formula of the improved soil erosion intensity calculation model is as follows: ; Where t represents time, R(t) is the rainfall erosivity factor, K is the soil erodibility factor, LS(t) is the slope and slope length factor, C(t) is the vegetation cover and management factor, P(t) is the soil and water conservation measures factor, and Φ(t) is the dynamic adjustment factor.

[0010] Preferably, the slope length factor LS(t) is calculated as follows: ; Where λ(t) is the slope length at time t; θ(t) is the slope at time t; m(t) is the slope length exponent at time t; and n is the slope exponent.

[0011] Preferably, the vegetation cover and management factor C(t) are calculated as follows: ; Where FVC(t) is the vegetation cover at time t.

[0012] Preferably, the dynamic adjustment factor Φ(t) is calculated as follows: ; Among them, F( t () represents the feature vector extracted by DAFEN at time t; δ is the Euclidean norm; δ is the sensitivity coefficient; ε is a small constant.

[0013] A dynamic soil erosion identification system based on deep learning and multi-source remote sensing data includes: The data acquisition module acquires multi-source remote sensing data of the area to be monitored; The model building module constructs a dynamic adaptive feature extraction network (DAFEN), which includes multiple layers of feature recognition layers. Each layer dynamically generates recognition units to adapt to the spatial distribution and feature complexity of the input data, and performs feature optimization through bidirectional information flow. The feature extraction module uses the dynamic adaptive feature extraction network DAFEN to extract features from multi-source remote sensing data and outputs a multi-dimensional feature map including land use type, vegetation cover, and topographic factors. The erosion assessment module, based on the multi-dimensional feature map and combined with the improved soil erosion intensity calculation model, calculates the soil erosion intensity level of each pixel at time t, and generates a dynamic monitoring result map of soil and water loss.

[0014] As can be seen from the above technical solution, compared with the prior art, this invention discloses a method and system for dynamic identification of soil erosion based on deep learning and multi-source remote sensing data. By constructing a dynamic adaptive feature extraction network DAFEN, it automatically extracts key parameters such as land use, vegetation cover, and topographic factors from multi-source remote sensing data. Combined with an improved soil erosion intensity calculation model, it achieves quantitative assessment and temporal dynamic monitoring of soil erosion. The improved model introduces a time-dynamic adjustment factor, considering the feature change rate, which significantly improves the accuracy and robustness of monitoring. This invention has the advantages of high automation, strong adaptability, and high accuracy, and can be widely applied to soil and water conservation work in fields such as land resources, water conservancy, and environmental protection. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0016] Figure 1 This is a schematic diagram of the method flow provided by the present invention; Figure 2 This is a schematic diagram of the system structure provided by the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] See Figure 1 This invention discloses a method for dynamic identification of soil erosion based on deep learning and multi-source remote sensing data, comprising: S1: Acquire multi-source remote sensing data of the area to be monitored; S2: Construct a dynamic adaptive feature extraction network DAFEN, which includes multiple layers of feature recognition layers. Each layer dynamically generates recognition units to adapt to the spatial distribution and feature complexity of the input data, and performs feature optimization through bidirectional information flow. S3: Use the dynamic adaptive feature extraction network DAFEN to extract features from multi-source remote sensing data and output a multi-dimensional feature map including land use type, vegetation cover, and topographic factors. S4: Based on the multi-dimensional feature map and combined with the improved soil erosion intensity calculation model, calculate the soil erosion intensity level of each pixel at time t, and generate a dynamic monitoring result map of soil and water loss.

[0019] In one specific embodiment, the specific steps for constructing the dynamic adaptive feature extraction network DAFEN in S2 include: S21: Design a network hierarchy, including an input layer, a feature extraction layer, and an output layer. The feature extraction layer consists of multiple sub-networks, each of which is responsible for extracting features at different levels of abstraction. S22: During the training process, recognition units are dynamically generated based on the distribution of input data, and parameters are optimized through top-down and bottom-up information flows; S23: Train the network using multi-source remote sensing data, and adjust the network parameters using loss function and accuracy evaluation index until the network converges; Among them, the dynamic adaptive feature extraction network DAFEN uses HRNet as its basic architecture and introduces a dynamic unit generation mechanism in the feature extraction path.

[0020] In one specific embodiment, the calculation formula of the improved soil erosion intensity calculation model is as follows: ; Where t represents time, R(t) is the rainfall erosivity factor, K is the soil erodibility factor, LS(t) is the slope and slope length factor, C(t) is the vegetation cover and management factor, P(t) is the soil and water conservation measures factor, and Φ(t) is the dynamic adjustment factor.

[0021] Specifically, the slope length factor LS(t) is calculated as follows: ; Wherein, λ(t) is the slope length at time t, in meters, extracted from DEM data; θ(t) is the slope at time t, in degrees, extracted from DEM data; m(t) is the slope length exponent at time t, dynamically adjusted according to the slope; and n is the slope exponent, with a value of 1.35.

[0022] Specifically, the value of m(t) is dynamically adjusted according to the slope θ(t), as follows: When θ(t) ≤ 5°, m(t) = 0.15; When 5° < θ(t) ≤ 12°, m(t) = 0.2; When 12° < θ(t) ≤ 22°, m(t) = 0.35; When 22° < θ(t) < 35°, m(t) = 0.45.

[0023] Specifically, the vegetation cover and management factor C(t) are calculated as follows: ; Wherein, FVC(t) is the vegetation cover at time t, which is calculated by using the red and near-infrared bands of remote sensing images and obtained based on a pixel-division model.

[0024] Specifically, the dynamic adjustment factor Φ(t) is calculated as follows: ; Among them, F( t () is the feature vector extracted by DAFEN at time t, including normalized land use type, vegetation index and topographic features; δ is the Euclidean norm; δ is the sensitivity coefficient, ranging from 0.05 to 0.2; ε is a small constant, with a value of 1 × 10⁻⁶. -6 .

[0025] Specifically, the soil and water conservation measures factor P(t) is assigned a value based on land use type, including: Woodland, grassland, undisturbed land: P(t) = 1; Cultivated land: P(t) = 0.35; Construction land, transportation land, and water area: P(t) = 0; Disturbance land use: P(t) = 0.5; Land use types are automatically classified using the DAFEN network.

[0026] Specifically, the Dynamic Adaptive Feature Extraction Network (DAFEN) proposed in this invention is an end-to-end deep learning framework. Its core innovation lies in its ability to directly, automatically, and in parallel extract all key parameters for soil erosion calculation from raw multi-source remote sensing data (in this embodiment, optical imagery and DEM), including land use type, vegetation cover FVC(t), slope θ(t), and slope length λ(t), and generate a deep feature vector F(t) for calculating the dynamic adjustment factor Φ(t). The specific data processing flow is as follows: Step 1: Network Input Input optical remote sensing images (multi-band, such as blue, green, red, and near-infrared) to identify the spectral characteristics of ground features.

[0027] Input a digital elevation model (DEM) to provide terrain elevation information.

[0028] Optionally, a multi-temporal data stack can be input for direct learning of temporal variation characteristics.

[0029] Second step of preprocessing: Radiometric calibration, atmospheric correction (such as COST model), and geometric registration are performed on the input images to ensure data quality and spatial consistency.

[0030] The third step of DAFEN employs a multi-branch encoder structure, rather than a single feature extraction stream. These branches process different types of data independently in the early stages, and then perform feature fusion at a deeper level to achieve information complementarity.

[0031] Part 1: High-resolution spatial detail preservation branch; The HRNet backbone network maintains high-resolution feature maps throughout the entire process.

[0032] This branch focuses on extracting fine spatial information such as texture and boundaries of land use and vegetation. Its high-resolution characteristics are crucial for accurately distinguishing land types such as "arable land," "forest land," and "disturbed land."

[0033] Part Two: Terrain Feature Extraction Branch; A convolutional network that takes a DEM (Digital Elevation Model) as its primary input. This network is designed to learn micro-topographic features such as slope, aspect, and curvature, as well as macro-topographic features such as topographic relief and surface incision depth.

[0034] Instead of using the traditional neighborhood window method, the slope θ(t) and slope length λ(t) of each pixel are regressed directly from the DEM. This approach better considers the overall terrain context and reduces local noise interference.

[0035] Part Three: Spectroscopic-Time Series Analysis Branch; A module that focuses on channel dimension and time series analysis (may include attention mechanism).

[0036] This branch delves into the spectral characteristics of optical images, particularly the red and near-infrared bands, to accurately calculate NDVI and further implicitly solves for vegetation cover FVC(t) using a built-in pixel bisection model. When multi-temporal data is input, this branch can also automatically capture the seasonal variation patterns of vegetation.

[0037] Step 4: Dynamic Adaptation and Feature Fusion Dynamic unit generation: During network training and inference, not all processing units are predefined. DAFEN dynamically activates or generates the weights of convolutional kernels based on the local features of the input data (such as texture complexity and gradient magnitude), especially investing more computational attention in key areas such as feature boundaries and terrain abrupt changes, thereby achieving on-demand allocation of computational resources and improving adaptability to sparse and heterogeneous data.

[0038] The network ultimately generates the required parameters in parallel through multiple dedicated output heads: Land use classification head: A softmax classification layer that outputs the probability distribution of land use type for each pixel, and finally takes the class with the highest probability as the result. This result is directly used for querying and assigning the P(t) factor.

[0039] Vegetation cover regression head: A sigmoid-activated regression layer that directly outputs the FVC(t) value (between 0 and 1) for each pixel. This value is directly substituted into the formula in claim 4 to calculate the C(t) factor.

[0040] Terrain parameter regression head: directly outputs the slope θ(t) (in degrees) and slope length λ(t) (in meters) for each pixel. These values ​​are directly used for the LS(t) factor calculation in claims 2 and 3.

[0041] Deep feature output: The feature map is directly extracted from the end of the network fusion layer and pooled into a feature vector F(t) for each pixel. This vector encapsulates the spectral, spatial, topographical, and temporal context information of the pixel and is used to calculate the dynamic adjustment factor Φ(t).

[0042] Step 5: Training and Optimization The network is jointly trained using a large number of labeled land use datasets, as well as ground truth slope / slope length data calculated from high-precision DEMs using traditional GIS methods and ground truth vegetation cover data obtained through remote sensing inversion.

[0043] The total loss function is a weighted sum of the losses from multiple tasks.

[0044] DAFEN enables end-to-end learning. It takes raw remote sensing images and DEMs as input and outputs all the necessary, optimized parameters directly for subsequent soil erosion model calculations, greatly reducing the error accumulation and manual intervention caused by step-by-step processing in traditional methods.

[0045] Multi-level feature fusion: The three branches mentioned above perform cross-connection and feature fusion at different depths of the network. For example, fusing the slope features extracted by the terrain branch with the vegetation features extracted by the spectral branch can help the network understand the difference between "woodland on a steep slope" and "woodland on a gentle slope" in terms of soil erosion, thereby generating a more physically meaningful deep feature vector F(t).

[0046] On the other hand, see Figure 2 This invention also provides a dynamic soil erosion identification system based on deep learning and multi-source remote sensing data, comprising: The data acquisition module acquires multi-source remote sensing data of the area to be monitored; The model building module constructs a dynamic adaptive feature extraction network (DAFEN), which includes multiple layers of feature recognition layers. Each layer dynamically generates recognition units to adapt to the spatial distribution and feature complexity of the input data, and performs feature optimization through bidirectional information flow. The feature extraction module uses the dynamic adaptive feature extraction network DAFEN to extract features from multi-source remote sensing data and outputs a multi-dimensional feature map including land use type, vegetation cover, and topographic factors. The erosion assessment module, based on the multi-dimensional feature map and combined with the improved soil erosion intensity calculation model, calculates the soil erosion intensity level of each pixel at time t, and generates a dynamic monitoring result map of soil and water loss.

[0047] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0048] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for dynamic identification of soil erosion based on deep learning and multi-source remote sensing data, characterized in that, include: S1: Acquire multi-source remote sensing data of the area to be monitored; S2: Construct a dynamic adaptive feature extraction network DAFEN, which includes multiple layers of feature recognition layers. Each layer dynamically generates recognition units to adapt to the spatial distribution and feature complexity of the input data, and performs feature optimization through bidirectional information flow. S3: Use the dynamic adaptive feature extraction network DAFEN to extract features from multi-source remote sensing data and output a multi-dimensional feature map including land use type, vegetation cover, and topographic factors. S4: Based on the multi-dimensional feature map and combined with the improved soil erosion intensity calculation model, calculate the soil erosion intensity level of each pixel at time t, and generate a dynamic monitoring result map of soil and water loss.

2. The method for dynamic identification of soil erosion based on deep learning and multi-source remote sensing data according to claim 1, characterized in that, The specific steps for constructing the dynamic adaptive feature extraction network DAFEN in S2 include: S21: Design a network hierarchy, including an input layer, a feature extraction layer, and an output layer. The feature extraction layer consists of multiple sub-networks, each of which is responsible for extracting features at different levels of abstraction. S22: During the training process, recognition units are dynamically generated based on the distribution of input data, and parameters are optimized through top-down and bottom-up information flows; S23: Train the network using multi-source remote sensing data, and adjust the network parameters using loss function and accuracy evaluation index until the network converges; The dynamic adaptive feature extraction network DAFEN uses HRNet as its basic architecture and introduces a dynamic unit generation mechanism in the feature extraction path.

3. The method for dynamic identification of soil erosion based on deep learning and multi-source remote sensing data according to claim 1, characterized in that, The calculation formula for the improved soil erosion intensity calculation model is as follows: ； Where t represents time, R(t) is the rainfall erosivity factor, K is the soil erodibility factor, LS(t) is the slope and slope length factor, C(t) is the vegetation cover and management factor, P(t) is the soil and water conservation measures factor, and Φ(t) is the dynamic adjustment factor.

4. The method for dynamic identification of soil erosion based on deep learning and multi-source remote sensing data according to claim 3, characterized in that, The slope length factor LS(t) is calculated as follows: ； Where λ(t) is the slope length at time t; θ(t) is the slope at time t; m(t) is the slope length exponent at time t; and n is the slope exponent.

5. The method for dynamic identification of soil erosion based on deep learning and multi-source remote sensing data according to claim 3, characterized in that, The vegetation cover and management factor C(t) are calculated as follows: ； Where FVC(t) is the vegetation cover at time t.

6. The method for dynamic identification of soil erosion based on deep learning and multi-source remote sensing data according to claim 3, characterized in that, The dynamic adjustment factor Φ(t) is calculated as follows: ； Among them, F( t () represents the feature vector extracted by DAFEN at time t; δ is the Euclidean norm; δ is the sensitivity coefficient; ε is a small constant.

7. A system for implementing the dynamic identification method for soil erosion based on deep learning and multi-source remote sensing data as described in any one of claims 1-6, characterized in that, include: The data acquisition module acquires multi-source remote sensing data of the area to be monitored; The model building module constructs a dynamic adaptive feature extraction network (DAFEN), which includes multiple layers of feature recognition layers. Each layer dynamically generates recognition units to adapt to the spatial distribution and feature complexity of the input data, and performs feature optimization through bidirectional information flow. The feature extraction module uses the dynamic adaptive feature extraction network DAFEN to extract features from multi-source remote sensing data and outputs a multi-dimensional feature map including land use type, vegetation cover, and topographic factors. The erosion assessment module, based on the multi-dimensional feature map and combined with the improved soil erosion intensity calculation model, calculates the soil erosion intensity level of each pixel at time t, and generates a dynamic monitoring result map of soil and water loss.

Images