Water chlorophyll concentration inversion method and system based on multi-modal data and lightweight model

Abstract

The invention provides a water chlorophyll a concentration inversion method and system based on multi-modal data and a lightweight model, and relates to the technical field of water environment remote sensing evaluation. The method comprises the following steps: firstly, acquiring a Gaofeng No.5 satellite remote sensing image, a sentinel No.3 satellite image and ground actual measurement data, and completing image preprocessing and water body pixel extraction; constructing a hyperspectral index and an aquatic vegetation index, and fusing the hyperspectral index and the aquatic vegetation index with the water body temperature, the pH environmental factors and the spectral reflectivity to form a multi-dimensional feature sample set; a core feature subset is obtained through random forest and XGBoost coupling feature selection, and a lightweight student model is trained based on knowledge distillation; and constructing a to-be-predicted feature sample for the to-be-predicted time phase image and the environment factor, inputting the to-be-predicted feature sample into the lightweight student model to obtain a chlorophyll a concentration predicted value, and generating a spatial distribution map and a quality control map layer. According to the invention, high-precision, low-redundancy and efficient deployment chlorophyll a concentration inversion is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of water environment remote sensing evaluation, in particular to a water body chlorophyll-a concentration inversion method and system based on multi-modal data and a light-weight model. BACKGROUND

[0002] Water body chlorophyll-a (Chl-a) concentration is a key water quality parameter for representing water body eutrophication and algal growth state, and is commonly used for water environment monitoring and evaluation of lakes and nearshore waters. Among existing monitoring methods, remote sensing inversion is widely used for spatial distribution mapping and dynamic tracking of chlorophyll-a concentration due to its advantages of large-scale, periodicity and non-contact.

[0003] Existing chlorophyll-a remote sensing inversion relies on single multispectral or hyperspectral images to construct spectral features and train regression models, but the distribution of chlorophyll-a in complex water environments is not only affected by water optical properties, but also related to aquatic vegetation distribution and environmental factors such as water temperature and pH. Single optical data is difficult to quantify and integrate these indirect influences, resulting in insufficient inversion accuracy and stability in turbid water bodies, vegetation interference water bodies or areas with significant environmental gradients.

[0004] To improve accuracy, some methods introduce hyperspectral, multi-source satellite and environmental factors to form a high-dimensional feature system, but this leads to high feature dimension and strong information redundancy. Traditional complex models can achieve high accuracy, but have the drawbacks of model structure redundancy, large computational load and slow reasoning speed, making it difficult to meet the efficiency requirements of rapid, large-scale business deployment and edge-side applications. Exhaustive optimization of features and models also cannot fundamentally solve the redundancy and efficiency bottleneck. SUMMARY

[0005] To overcome the shortcomings of the prior art, the present application provides a water body chlorophyll-a concentration inversion method and system based on multi-modal data and a light-weight model, which constructs an efficient and stable chlorophyll-a inversion model through multi-modal feature fusion and knowledge distillation, solving the technical problems of unstable inversion, feature redundancy and difficulty in efficient deployment of traditional methods.

[0006] To achieve the above-mentioned purposes, the present application provides the following solutions:

[0007] A water body chlorophyll-a concentration inversion method based on multi-modal data and a light-weight model, comprising:

[0008] acquiring high-resolution five satellite remote sensing images covering the target water area, sentinel three satellite images, ground-truth chlorophyll-a concentration data, and contemporaneous water temperature and pH environmental factor data;

[0009] The high-resolution five satellite remote sensing image and the sentinel three satellite image are respectively preprocessed, and water body pixels are extracted to obtain a water body image subset;

[0010] A hyperspectral index is constructed based on the high-resolution five satellite remote sensing image, and a water vegetation index is constructed based on the sentinel three satellite image, and is fused with the water body temperature, the pH environmental factor data, and the spectral reflectivity of the high-resolution five satellite remote sensing image and the sentinel three satellite image in time and space to form a multi-dimensional feature sample set;

[0011] The multi-dimensional feature sample set is subjected to a feature selection strategy coupling random forest and XGBoost for redundancy optimization to obtain a core feature subset;

[0012] The teacher model is trained based on the core feature subset, and a lightweight student model is obtained through knowledge distillation training;

[0013] The high-resolution five satellite remote sensing image, the sentinel three satellite image, and the water body temperature and pH environmental factor data of the same period of the to-be-predicted time phase are formed into a to-be-predicted multi-dimensional feature sample in the construction mode of the multi-dimensional feature sample set, and the features corresponding to the core feature subset in the to-be-predicted multi-dimensional feature sample are input into the lightweight student model to output the predicted value of chlorophyll a concentration, and a chlorophyll a concentration spatial distribution map and a quality control layer are generated.

[0014] Preferably, the step of pre-processing the high-resolution five satellite remote sensing image and the sentinel three satellite image comprises:

[0015] Radiometric calibration is performed on the high-resolution five satellite remote sensing image to convert digital values into radiance values, and atmospheric correction is performed using a FLAASH or 6S atmospheric radiation transfer model to obtain an above-water reflectance image;

[0016] The sentinel three satellite image is subjected to water color atmospheric correction and solar glare correction using a C2RCC processor to obtain a water surface reflectance image, and is re-projected into a consistent coordinate system with the high-resolution five satellite remote sensing image.

[0017] Preferably, the step of extracting water body pixels comprises:

[0018] The normalized difference water index is calculated based on the above-water reflectance image of the high-resolution five satellite remote sensing image and the water surface reflectance image of the sentinel three satellite image, respectively;

[0019] The normalized difference water index is compared with a preset water body discrimination threshold, and a water body binary mask layer is generated according to the comparison result;

[0020] Water body region pixels are extracted according to the water body binary mask layer to form a water body image subset.

[0021] Preferably, the step of forming a multi-dimensional feature sample set comprises:

[0022] obtaining ground-truth chlorophyll-a concentration data corresponding to the satellite overpass time; the ground-truth chlorophyll-a concentration data at least includes sampling latitude and longitude coordinates, chlorophyll-a concentration value, water temperature and pH environmental factor data;

[0023] generating environmental factor layers consistent with the spatial resolution of the GF-5 satellite remote sensing image through spatial interpolation of the water temperature and the pH environmental factor data;

[0024] performing spatial registration based on the sampling latitude and longitude coordinates and a water body image subset of the GF-5 satellite remote sensing image, finding the nearest neighbor valid pixel and extracting the corresponding hyperspectral index;

[0025] performing spatial registration based on the sampling latitude and longitude coordinates and a water body image subset of the Sentinel-3 satellite image, finding the nearest neighbor valid pixel and extracting the corresponding aquatic vegetation index;

[0026] finding the pixel corresponding to the sampling latitude and longitude coordinates in the environmental factor layer and extracting the water temperature and pH environmental factor features;

[0027] pairing the hyperspectral index, the aquatic vegetation index and the environmental factor features with the corresponding chlorophyll-a concentration value to form a sample entry, and constructing the multi-dimensional feature sample set with all the sample entries.

[0028] Preferably, the hyperspectral index is at least one of the candidate hyperspectral indices sensitive to chlorophyll-a constructed based on the band ratio, derivative spectrum and three-band combination of the GF-5 satellite remote sensing image; the aquatic vegetation index is the normalized red edge index constructed based on the red edge band and red band of the Sentinel-3 satellite image.

[0029] Preferably, the feature selection strategy coupled with the random forest and XGBoost includes:

[0030] training a random forest regression model and an XGBoost regression model based on the multi-dimensional feature sample set respectively;

[0031] extracting a first feature importance list from the random forest regression model and a second feature importance list from the XGBoost regression model;

[0032] normalizing and then weightedly averaging the scores of the first feature importance list and the second feature importance list to obtain a coupled feature importance score;

[0033] According to the coupling feature importance scores, top-K features are sorted from high to low or features satisfying an importance threshold are filtered to obtain the core feature subset.

[0034] Preferably, the teacher model and the lightweight child model are both models for chlorophyll-a concentration inversion trained by taking the core feature subset as input, and the teacher model is a deep neural network model or an XGBoost regression model; and the lightweight child model is a neural network model, a pruned decision tree model or an XGBoost regression model.

[0035] Preferably, the lightweight child model is trained by knowledge distillation, including:

[0036] The core feature subset is input into the teacher model to obtain a corresponding teacher model prediction value, and the teacher model prediction value is taken as a soft label.

[0037] A ground-truth chlorophyll-a concentration value is taken as a hard label to construct a joint loss function that simultaneously constrains the error between the lightweight child model prediction value and the hard label and the error between the lightweight child model prediction value and the soft label.

[0038] The lightweight child model is trained according to the joint loss function to obtain the lightweight child model.

[0039] Preferably, the evaluation and deployment of the lightweight child model include:

[0040] The coefficient of determination, root mean square error and mean absolute error are used to evaluate the chlorophyll-a concentration inversion accuracy of the lightweight child model.

[0041] In the deployment application, the teacher model is discarded, and only the lightweight child model is used to predict the chlorophyll-a concentration of the to-be-predicted multi-dimensional feature sample to obtain a chlorophyll-a concentration prediction value.

[0042] According to the chlorophyll-a concentration prediction value, a chlorophyll-a concentration spatial distribution map and a quality control layer are generated.

[0043] A water body chlorophyll-a concentration inversion system based on multi-modal data and a lightweight model, including:

[0044] A multi-modal data acquisition unit is configured to acquire high-resolution five satellite remote sensing images, sentinel three satellite images, ground-truth chlorophyll-a concentration data and contemporaneous water temperature and pH environmental factor data covering a target water area.

[0045] An image preprocessing and water body pixel extraction unit is configured to preprocess the GF-5 satellite remote sensing image and the Sentinel-3 satellite image respectively, and extract water body pixels to obtain a water body image subset;

[0046] A multi-modal index construction and feature fusion unit is configured to construct a hyperspectral index based on the GF-5 satellite remote sensing image, construct a water vegetation index based on the Sentinel-3 satellite image, and fuse the water body temperature, the pH environmental factor data, and the spectral reflectance of the GF-5 satellite remote sensing image and the Sentinel-3 satellite image in space and time to form a multi-dimensional feature sample set;

[0047] A coupling feature selection and redundancy optimization unit is configured to perform redundancy optimization on the multi-dimensional feature sample set by using a feature selection strategy coupling random forest and XGBoost to obtain a core feature subset;

[0048] A teacher model training and knowledge distillation unit is configured to train a teacher model based on the core feature subset, and obtain a lightweight student model through knowledge distillation training;

[0049] A to-be-predicted feature generation and inversion output unit is configured to form a to-be-predicted multi-dimensional feature sample based on the GF-5 satellite remote sensing image, the Sentinel-3 satellite image, and the water body temperature and the pH environmental factor data of the same period according to the construction method of the multi-dimensional feature sample set, input the features corresponding to the core feature subset in the to-be-predicted multi-dimensional feature sample into the lightweight student model, output a predicted value of chlorophyll a concentration, and reconstruct a chlorophyll a concentration spatial distribution map and a quality control layer.

[0050] The present application discloses the following technical effects:

[0051] The present application realizes the joint representation of water body optical characteristics and key environmental characteristics by constructing a multi-modal feature system including hyperspectral indexes of GF-5 satellite remote sensing images, water vegetation indexes of Sentinel-3 satellite images, and water temperature and pH environmental factors. Compared with the prior art which only relies on single spectral information and is difficult to cover turbid water bodies, vegetation interference water bodies, and areas affected by environmental gradients, the present application can extract key features related to the change of chlorophyll a from multiple sources, multiple scales, and multiple physical dimensions, making the inversion results have stronger stability and applicability in complex water environments.

[0052] The application utilizes a coupling feature selection strategy of random forest and XGBoost to optimize the redundancy of multi-modal feature sample set, obtains a core feature subset with clear structure, and avoids the problem of relying on a large number of repeated training and manual screening in a high-dimensional feature space in the traditional method. The coupling strategy can extract feature importance from different model mechanisms, and realize multi-angle feature evaluation through weighted fusion, so that the final feature used for inversion has both physical interpretability and avoids the negative impact of redundant features on model calculation and generalization ability.

[0053] The application trains a teacher model with the core feature subset, and trains a lightweight student model using knowledge distillation, so that the student model greatly reduces the feature input size and reasoning overhead while maintaining the inversion accuracy of the teacher model. Compared with the contradiction in the background art that "large model has high precision but is difficult to deploy, and small model is fast but has insufficient precision", the application migrates the key mapping relationship to the lightweight student model through distillation, realizes a business inversion framework with simple model structure, fast reasoning speed and controllable precision, and is more suitable for the engineering requirements of large-scale, multi-temporal water quality monitoring.

[0054] The application uses the lightweight student model to reason the multi-dimensional feature sample to be predicted, and reconstructs the spatial distribution map of chlorophyll-a concentration based on the predicted value, and generates a quality control layer for identifying the integrity of the input feature, the matching condition of the environmental factor and the potential shielding area. This mechanism not only improves the spatial continuity and overall reliability of the inversion result, but also provides visual confidence auxiliary information for subsequent water environment supervision, solving the problem of "inversion result difficult to verify and lack of quantitative marking of reliability" in the prior art. BRIEF DESCRIPTION OF DRAWINGS

[0055] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments. Obviously, the drawings in the following description only constitute some embodiments of the application, and for those skilled in the art, other drawings can also be obtained from these drawings without creative labor.

[0056] Figure 1 The method flowchart provided for the embodiments of the application;

[0057] Figure 2 The system structure schematic diagram provided for the embodiments of the application. DETAILED DESCRIPTION

[0058] With reference to the accompanying drawings, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those ordinarily skilled in the art without creative effort belong to the scope of the present application.

[0059] The present application aims to provide a water body chlorophyll-a concentration inversion method and system based on multi-modal data and a light-weight model, which realizes high-precision, low-cost and business deployment of chlorophyll-a concentration rapid inversion based on multi-source data in complex water environments.

[0060] In order to make the above-mentioned purposes, features and advantages of the present application more apparent and easy to understand, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0061] Figure 1 The method flowchart provided by the embodiments of the present application is shown in Figure 1 The present application provides a water body chlorophyll-a concentration inversion method based on multi-modal data and a light-weight model, which includes:

[0062] Step 100: Obtain high-resolution five satellite remote sensing images, sentinel three satellite images, ground measured chlorophyll-a concentration data, and contemporaneous water temperature and pH environmental factor data covering the target water area;

[0063] Step 200: Preprocess the high-resolution five satellite remote sensing images and the sentinel three satellite images respectively, and extract water body pixels to obtain a water body image subset;

[0064] Step 300: Construct hyperspectral indices based on the high-resolution five satellite remote sensing images, construct aquatic vegetation indices based on the sentinel three satellite images, and fuse with the water temperature, pH environmental factor data, and the spectral reflectance spatiotemporal registration of the high-resolution five satellite remote sensing images and the sentinel three satellite images to form a multi-dimensional feature sample set;

[0065] Step 400: Perform redundancy optimization on the multi-dimensional feature sample set using a feature selection strategy coupling random forest and XGBoost to obtain a core feature subset;

[0066] Step 500: Train a teacher model with the core feature subset, and obtain a light-weight student model through knowledge distillation training;

[0067] Step 600: Forming a multi-dimensional feature sample to be predicted according to the construction mode of the multi-dimensional feature sample set for the GF-5 high-resolution satellite remote sensing image, the Sentinel-3 satellite image, and the synchronous water temperature and pH environmental factor data of the time phase to be predicted, and selecting a feature input lightweight student model corresponding to the core feature subset in the multi-dimensional feature sample to be predicted, outputting a chlorophyll a concentration prediction value, and reconstructing to generate a chlorophyll a concentration spatial distribution map and a quality control layer.

[0068] Specifically, the technical route of the embodiment is as follows:

[0069] (1) Multi-modal data acquisition. Obtain GF-5 high-resolution remote sensing images, Sentinel-3 OLCI images, ground-truth chlorophyll a concentration data, and synchronous water temperature (Temperature) and pH (pH) environmental factor data covering the target water area;

[0070] (2) Multi-modal data preprocessing. Perform radiation calibration and atmospheric correction on the GF-5 image; perform water color atmospheric correction (such as C2RCC) on the Sentinel-3 image; and perform water pixel extraction on all images to generate their respective water image subsets;

[0071] (3) Multi-dimensional feature construction and sample set generation. Construct hyperspectral indices (such as band ratio, derivative) based on the GF-5 image; construct aquatic vegetation indices (such as NDRE) based on the Sentinel-3 image; synchronize the water temperature, pH, and other environmental factors; and perform spatial and temporal registration of all features and ground-truth Chl-a concentration points to form a multi-dimensional feature sample set.

[0072] (4) Feature optimization and data redundancy elimination. Use a random forest (RF) and XGBoost coupled feature selection strategy to evaluate the feature importance and reduce the dimension of the multi-dimensional feature sample set, and select a core feature subset with the largest contribution and the lowest redundancy.

[0073] (5) Lightweight model construction. Construct a high-precision "teacher model" and a lightweight "student model". Use the "knowledge distillation" framework to migrate and compress the complex knowledge learned by the teacher model on the "core feature subset" into the student model.

[0074] (6) Concentration prediction. Deploy and apply the trained lightweight student model, input new multi-modal data, perform rapid concentration prediction, and generate a chlorophyll a concentration spatial distribution map and a supporting quality control layer.

[0075] Preferably, the embodiment performs preprocessing and water pixel extraction on the obtained original GF-5 and Sentinel-3 remote sensing images covering the target water area to obtain water image subsets, including:

[0076] In this embodiment, the original GF-5 remote sensing image covering the target water area is first acquired, and the digital values ​​(DN) of each band are converted into physical unit radiance values ​​using a radiometric calibration algorithm. Subsequently, the calibration results are atmospherically corrected using the FLAASH or 6S atmospheric radiative transfer model to obtain the true reflectance image above the water surface.

[0077] Simultaneously, Sentinel-3OLCIL1 level images were acquired, and water color atmospheric correction was performed using the C2RCC (Case2RegionalCoastColour) processor integrated in the ESASNAP software, and water surface reflectance (Rrs) was output.

[0078] Based on the obtained GF-5 and Sentinel-3 reflectance images, typical water body identification indices (such as the Normalized Difference Water Index, NDWI) were selected for calculation. The NDWI calculation formula is as follows:

[0079] NDWI=(Green-NIR) / (Green+NIR);

[0080] In the formula, Green represents the green light band (such as B20 of GF-5 and Oa06 of S3-OLCI), and NIR represents the near-infrared band (such as B45 of GF-5 and Oa17 of S3-OLCI).

[0081] The NDWI value is compared with a preset discrimination threshold (e.g., 0). Pixels with NDWI greater than the threshold are classified as water pixels (assigned a value of 1), and others are classified as non-water pixels (assigned a value of 0), generating a binary mask layer. This mask layer is then used to extract pixel data of the water area, forming a subset of the water image.

[0082] Furthermore, this embodiment utilizes the Gaofen-5 (GF-5) satellite, which carries the Advanced Hyperspectral Imager (AHSI). This camera has 330 spectral bands, covering a wide spectral range (approximately 400-2500 nm) from visible-near infrared (VNIR) to short-wave infrared (SWIR). Its spectral resolution is better than 5 nm in the VNIR band and better than 10 nm in the SWIR band, with a spatial resolution of 30 m.

[0083] Furthermore, this embodiment uses the Sentinel-3 satellite, which carries the Ocean and Land Colorimeter (OLCI). This instrument carries 21 spectral bands in the spectral range of 400 nm to 1020 nm, with a spatial resolution of 300 m.

[0084] Preferably, the steps for constructing multidimensional features and generating sample sets in this embodiment include:

[0085] 1. Feature Construction:

[0086] Hyperspectral indices (based on GF-5): From the 330 bands of GF-5, a series of candidate hyperspectral indices sensitive to chlorophyll a were constructed by means of band ratio, derivative spectrum, and three-band combination.

[0087] Aquatic vegetation index (based on Sentinel-3): To reflect the indirect influence of aquatic vegetation on chlorophyll a distribution, the normalized red edge index (NDRE) is constructed using the red-edge band (e.g., B11, ~709nm) and red band (e.g., B8, ~665nm) of Sentinel-3OLCI to quantify the distribution of vegetation in water bodies (especially submerged water).

[0088] Environmental factors: The measured water temperature and pH data are used to generate an environmental factor layer with the same spatial resolution (30m) as the GF-5 image by spatial interpolation (such as inverse distance weighting).

[0089] As an example, this embodiment takes into account a problem that objectively exists in lakes, reservoirs, and nearshore waters: water temperature and pH often exhibit spatial distribution characteristics of "non-stationary abrupt changes" and "relative continuity within the same water mass and abrupt changes between different water masses." For example, when affected by plumes flowing into the lake, shallow-deep channel boundaries along the shoreline, thermal stratification, nearshore tidal fronts, or local algal blooms, environmental factors within the same 30m pixel scale can change drastically over short distances. In this case, traditional inverse distance weighted interpolation, when extrapolating smoothly based solely on geometric distance, will forcibly mix measured points from different water masses or different optical types, resulting in an overly smoothed environmental factor layer and "flattened" boundaries. This leads to inconsistencies with the true physical partitioning of water pixels in the Gaofen-5 satellite remote sensing image, thereby weakening the effectiveness of environmental factors in subsequent multidimensional feature sample sets.

[0090] To address the aforementioned issues, this embodiment proposes a specific procedure for spectrally constrained non-stationary environmental factor interpolation, replacing conventional spatial interpolation. Specifically, after obtaining measured water temperature and pH environmental factor data during the same period as the satellite transit, the spectral feature vectors of the nearest neighbor effective pixels to each measured point are first extracted from the water image subset of the Gaofen-5 satellite remote sensing image. These spectral feature vectors consist of the multi-band spectral reflectance of the Gaofen-5 satellite remote sensing image at the target water pixel or a set of hyperspectral indices constructed from it. Then, when estimating environmental factors for any water pixel to be interpolated, the spatial distance similarity and spectral feature similarity between the pixel and each measured point are considered simultaneously. This suppresses the erroneous influence of measured points across water masses and optical types on the interpolation, thereby generating an environmental factor layer with consistent spatial resolution and stronger boundary preservation capabilities that matches the Gaofen-5 satellite remote sensing image, and further forming a multi-dimensional feature sample set.

[0091] In this embodiment, the spectral constrained interpolation calculates the water body pixels to be interpolated using the following formula. Environmental factor estimates:

[0092] ;

[0093] in, Water body pixels to be interpolated The estimated values ​​of environmental factors, wherein the environmental factors are water temperature or pH; For the first Measured values ​​of environmental factors corresponding to each measurement point; The number of measured points participating in the interpolation; Water body pixels to be interpolated With the The spatial distance between the measured points in the projection plane coordinate system; Water body pixels to be interpolated Spectral feature vectors extracted from a subset of water images in the Gaofen-5 satellite remote sensing imagery; For the first The spectral feature vector corresponding to the nearest neighbor effective pixel of each measured point; It is a 2-norm; This is a scale parameter for spectral feature similarity, used to measure the attenuation of interpolation weights due to spectral differences.

[0094] The spectral feature similarity scale parameter involved in the above formula refers to the scale quantity obtained by statistically analyzing the pairwise differences of the spectral feature vectors of all measured points within the same target water area. It is used to map the magnitude of spectral differences to the attenuation degree of interpolation weights. Its function is to adaptively characterize the optical partitions of different water masses, different turbidities, or different vegetation disturbance backgrounds in lakes, reservoirs, and nearshore waters, so that measured points with large spectral differences automatically obtain smaller weights in interpolation, thereby avoiding non-physical smoothing caused by cross-optical type extrapolation.

[0095] Furthermore, in this embodiment, From the day the satellite passes over (e.g.) Surface-measured water temperature and pH environmental factor data within 1 hour; The coordinates are calculated by converting the latitude and longitude coordinates of the water body pixels to be interpolated and the measured points into projected plane coordinates using the georegistration matrix of Gaofen-5 satellite remote sensing images. The unit is meters. and All are composed of multi-band spectral reflectance vectors or hyperspectral index vectors extracted from the water body image subsets of Gaofen-5 satellite remote sensing images at the corresponding pixels, and their dimensions are consistent with the construction of hyperspectral indices used for chlorophyll a inversion. The standard deviation of the differences in the spectral feature vectors of the nearest-neighbor effective pixels of all measured points is automatically calculated. For example, when the standard deviation of the differences in the pairwise L2 norms of the spectral features of measured points within a target water area is calculated as follows: If the spectral difference between a pixel to be interpolated and a measured point is... Then the exponent term is The weight of the measured point will be significantly reduced, thus ensuring that the interpolation results follow the actual optical zoning and non-stationary boundaries of the water body.

[0096] 2. Spatial registration and sample set generation:

[0097] Obtain ground-based measured data of the target water area on the day the satellite passes over (e.g., within ±24 hours), including sampling time, latitude and longitude coordinates, chlorophyll a concentration, water temperature, and pH value.

[0098] Based on the latitude and longitude coordinates of each measured point, spatial registration is performed with multimodal data to find the nearest neighbor valid pixels:

[0099] In the GF-5 water body subset (30m), the corresponding pixels are located and their hyperspectral indices are extracted.

[0100] In the Sentinel-3 water subset (300m), the corresponding pixel is located and its NDRE index is extracted.

[0101] In the environmental factor layer (30m), find the corresponding pixel and extract its water temperature and pH value.

[0102] All the features extracted above (hyperspectral index, NDRE, water temperature, pH) are used as input feature vectors and paired with the measured chlorophyll a concentration value at that point (as a label) to form a complete sample entry. Combining all sample entries constitutes a multidimensional feature sample set for subsequent model training.

[0103] Preferably, the feature optimization steps in this embodiment include:

[0104] To address the issues of high feature dimensionality and information redundancy in multidimensional feature sample sets, this embodiment employs a feature selection strategy coupled with Random Forest (RF) and XGBoost. The machine learning model library in this embodiment includes:

[0105] (1) Random Forest Model

[0106] Random Forest is a classic algorithm in Bagging. Its basic principle is to build multiple weak estimators in parallel and then combine their predictions to obtain the final output. Random Forest generates multiple sample subsets by sampling the training set with replacement, and independently builds a decision tree on each subset. At each node of the tree, the algorithm randomly selects a subset of features and searches for the optimal feature to split the node. Finally, the output of Random Forest is the ensemble of the predictions from all decision trees, expressed in the formula:

[0107]

[0108] In the formula: T is the number of decision trees, x is the independent variable, which comes from the training set, and DT is the output of the random forest in the training set.

[0109] (2) XGBoost model

[0110] XGBoost (ExtremeGradientBoosting) is a next-generation algorithm based on GBDT. Its base tree model can fit nonlinear data well, and the calculation formula is as follows:

[0111]

[0112] In the formula, This represents the model prediction value for the i-th sample. This represents the input data for the i-th sample. This represents the overall prediction model consisting of multiple trees, where K is the number of trees in the model. For the k-th tree in the sample The output at the point, F is the space of the tree (usually a CART tree).

[0113] The core idea of ​​XGBoost is the same as that of GBDT, but it achieves a balance between accuracy and complexity. This is mainly reflected in the fact that XGBoost adds a structural risk term to the loss function, which forms the objective function, and its expression is:

[0114]

[0115] In the formula, SRI represents the structural risk term. This change causes XGBoost to be trained in the direction of minimizing the objective function, rather than minimizing the loss function. The objective function value of XGBoost is composed of the loss term and the structural risk term. The actual observed values ​​of the sample (the actual chlorophyll a concentration or other target variables); This represents the model's predicted value for the sample (obtained by stacking multiple trees). This is the loss function used to measure the predicted value. Compared with the true value Errors between (e.g., squared error).

[0116] Preferably, to evaluate the inversion performance of the random forest model coupled with the XGBoost model on remote sensing images, this embodiment uses the measured chlorophyll a concentration to evaluate the accuracy of the inversion results. This embodiment uses the coefficient of determination R0. 2 The root mean square error (RMSE) and mean absolute error (MAE) are used as evaluation metrics to assess the accuracy of image inversion results, using the measured chlorophyll a concentration as the true value. 2 The formulas for calculating RMSE and MAE are as follows:

[0117] ;

[0118] ;

[0119] ;

[0120] In the formula, n is the number of samples. It is the i-th observation. It is the i-th predicted value. It is the mean of the observed values.

[0121] In this embodiment, to fully exploit the differences in feature sensitivity among different model structures from the multidimensional feature sample set, a parallel evaluation strategy is preferably adopted to construct two independent regression models. Specifically, the multidimensional feature sample set is input into a Random Forest (RF) regression model and an XGBoost regression model, respectively, and two feature learning models with differentiated decision-making mechanisms are trained independently. The Random Forest model obtains feature splitting contributions through ensemble learning of multiple decision trees, while the XGBoost model forms feature gain information based on an additive tree structure through a gradient boosting framework, providing an independent source of feature importance expression for subsequent coupled evaluation.

[0122] After the RF model and XGBoost model are trained, this embodiment further extracts the feature importance indices from the two models respectively. For the RF model, a feature importance list A (first feature importance list) based on "reduction of Gini impurity" is extracted from the node splitting of each decision tree; for the XGBoost model, a feature importance list B (second feature importance list) based on "average gain" is extracted from the splitting contribution of each tree. Since the two models have different importance scoring dimensions, this embodiment normalizes the scores in lists A and B, for example, by using max-min normalization or mean-variance normalization, to make the two scoring systems comparable and provide a consistent dimensional basis for subsequent calculation of coupled feature importance.

[0123] After obtaining the normalized lists of two feature importances, this embodiment employs a weighted fusion strategy to calculate the coupling feature importance score. The coupling score can be calculated using the formula... The values ​​are obtained, where α and β are the weighting coefficients of the RF importance score and the XGBoost importance score, respectively, reflecting the proportion of different models in the feature sensitivity evaluation. Subsequently, all features are sorted from high to low according to the coupled feature importance scores, and the Top-K features with the highest scores are selected, or features that meet the score requirements are selected according to a preset importance threshold. This embodiment thus obtains a subset of core features that significantly contribute to chlorophyll a concentration inversion and have the lowest redundancy, providing the most representative input features for subsequent teacher model training and lightweight student model distillation.

[0124] Specifically, in this embodiment, a knowledge distillation framework is preferably used to construct a lightweight model, enabling the model to maintain inversion accuracy while achieving higher inference speed. First, a teacher model is constructed based on a core feature subset obtained through feature optimization. The input to the teacher model is the core feature subset, and the model structure can be a deep neural network (DNN) model or a finely tuned XGBoost regression model with a large number of parameters. During the training process, the core feature subset and the corresponding measured chlorophyll a (Chl-a) concentration constitute the training samples. Iterative optimization is performed through supervised learning, with the training objective being to maximize inversion accuracy without considering inference computational costs. After training, the teacher model can output high-precision predicted values ​​corresponding to the input features.

[0125] After the teacher model is constructed, this embodiment further constructs a lightweight student model, whose input is also the aforementioned subset of core features. The student model employs a more concise structure, fewer parameters, and faster computation speed, such as a shallow neural network model, a pruned decision tree model, or a simplified XGBoost model with fewer trees and shallower depth. The design of this lightweight model aims to reduce the computational cost caused by multi-layered computational structures and excessive nodes, significantly improving operational efficiency during the model inference stage, thereby meeting the engineering deployment requirements of large-scale and high-time-sensitivity water chlorophyll a concentration inversion tasks.

[0126] In the training process of the lightweight student model, this embodiment uses a joint loss function for model learning. The joint loss function consists of hard-label loss and soft-label loss, and is defined as follows: ,in This is the balance coefficient between hard-label loss and soft-label loss. Hard-label loss This is used to calculate the error between the student model's predicted values ​​and the measured chlorophyll a concentration, such as the root mean square error (RMSE), to ensure that the student model can learn the true inversion results; soft label loss. The error between the student model's predictions and the teacher model's predictions is calculated using methods such as L1 or L2 loss. To obtain the soft labels, each core feature subset from the training set is input into the teacher model to obtain the corresponding teacher model prediction. By jointly using hard and soft labels during training, the student model learns not only from real observation data but also from the teacher model's feature mapping capabilities and generalization patterns based on a large-scale parameter structure. Ultimately, this achieves effective transfer and compression of teacher model knowledge to a lightweight model.

[0127] Preferably, the concentration prediction step in this embodiment includes:

[0128] After training is complete, discard the large and slow "teacher model" and deploy only the trained "lightweight student model".

[0129] After acquiring new GF-5, Sentinel-3, and other multimodal image data, perform the following operations sequentially for each pixel in the water body pixel subset:

[0130] We used a feature construction method to extract its hyperspectral index, NDRE index, and corresponding environmental factors.

[0131] By applying feature selection criteria, a "core feature subset" is selected from the extracted features.

[0132] The "core feature subset" is input into the lightweight student model to perform rapid concentration prediction and obtain the predicted chlorophyll a concentration value for that pixel.

[0133] The predicted values ​​of all water body pixels are reconstructed according to their spatial locations to generate a spatial distribution map of chlorophyll a concentration consistent with the GF-5 image resolution (30m). Simultaneously, information such as model response confidence level and input data integrity (e.g., missing S3 data, GF-5 cloud cover) is recorded to generate a corresponding quality control layer, which can be used in subsequent visualization, water quality assessment, or monitoring and early warning systems.

[0134] The beneficial effects of this invention are as follows:

[0135] This invention provides a method for retrieving chlorophyll a concentration in water bodies by fusing hyperspectral remote sensing imagery and ground-based measured data. It constructs a complete workflow encompassing multimodal data acquisition and preprocessing, spatial matching of measured data, construction of aquatic vegetation indices and hyperspectral indices, feature optimization, lightweight model construction, and concentration prediction. The method features clearly defined modules, explicit data transfer logic between steps, and a highly versatile and scalable overall structure.

[0136] The method provided by this invention, by fusing Sentinel-3 imagery, environmental factors such as water temperature and pH, and combining data redundancy optimization techniques, achieves the following technical effects:

[0137] Enhancing the integrity of the feature system: The introduction of aquatic vegetation indices and key environmental factors (temperature, pH) constructed using Sentinel-3 effectively quantifies the indirect effects of aquatic vegetation growth and the aquatic environment on Chl-a concentration, thereby improving the physical interpretability and robustness of the inversion model.

[0138] Achieving dataset optimization and lightweighting: By introducing random forest coupled with XGBoost dimensionality reduction technology into complex models, redundant information in multimodal data is reduced during the training phase, ensuring that the final model focuses on and learns the most discriminative core features.

[0139] High efficiency and strong engineering applicability: The complex knowledge of the model is efficiently compressed into a lightweight model, which maintains the inversion accuracy while significantly reducing the size of the model and the inference latency, greatly improving the computational efficiency of the model and the flexibility of business deployment.

[0140] High precision and adaptability: Deep features of multimodal data are mined through a high-precision model and then passed to a lightweight model, ensuring high precision of the inversion results.

[0141] Overall, the method of this invention has the advantages of complete closed-loop process, high model accuracy, flexible feature expression, accurate spatial positioning, and standardized result expression. It is suitable for water quality remote sensing inversion scenarios based on multimodal data and lightweight models, and has good prospects for promotion and application.

[0142] Corresponding to the above methods, such as Figure 2 As shown, this embodiment also provides a water chlorophyll a concentration inversion system based on multimodal data and a lightweight model, including:

[0143] The multimodal data acquisition unit is used to acquire remote sensing images from the Gaofen-5 satellite, Sentinel-3 satellite images, ground-measured chlorophyll a concentration data, and concurrent water temperature and pH environmental factor data covering the target water area.

[0144] The image preprocessing and water body pixel extraction unit is used to preprocess the Gaofen-5 satellite remote sensing images and the Sentinel-3 satellite images respectively, and extract water body pixels to obtain a subset of water body images.

[0145] The multimodal index construction and feature fusion unit is used to construct a hyperspectral index based on the Gaofen-5 satellite remote sensing image, construct an aquatic vegetation index based on the Sentinel-3 satellite image, and fuse it with the water temperature, pH environmental factor data, and the spatiotemporal registration of the spectral reflectance of the Gaofen-5 satellite remote sensing image and the Sentinel-3 satellite image to form a multidimensional feature sample set.

[0146] The coupled feature selection and redundancy optimization unit is used to perform redundancy optimization on the multidimensional feature sample set by using a feature selection strategy coupled with random forest and XGBoost to obtain a core feature subset.

[0147] The teacher model training and knowledge distillation unit is used to train the teacher model with the core feature subset and to obtain a lightweight student model through knowledge distillation training.

[0148] The predictable feature generation and inversion output unit is used to generate predictable multidimensional feature samples from the Gaofen-5 satellite remote sensing images, Sentinel-3 satellite images, and water temperature and pH environmental factor data of the same time phase according to the construction method of the multidimensional feature sample set. It selects the features corresponding to the core feature subset in the predictable multidimensional feature samples and inputs them into the lightweight student model, outputs the predicted value of chlorophyll a concentration, and reconstructs and generates the spatial distribution map of chlorophyll a concentration and the quality control layer.

[0149] 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 systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0150] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for inverting chlorophyll a concentration in water bodies based on multimodal data and a lightweight model, characterized in that, include: Acquire Gaofen-5 satellite remote sensing images, Sentinel-3 satellite images, ground-measured chlorophyll a concentration data, and concurrent water temperature and pH environmental factor data covering the target water area; The Gaofen-5 satellite remote sensing images and the Sentinel-3 satellite images were preprocessed and water body pixels were extracted to obtain a subset of water body images. A hyperspectral index is constructed based on the Gaofen-5 satellite remote sensing image, and an aquatic vegetation index is constructed based on the Sentinel-3 satellite image. These are then fused with the water temperature, pH environmental factor data, and the spectral reflectance data of the Gaofen-5 satellite remote sensing image and the Sentinel-3 satellite image through spatiotemporal registration to form a multidimensional feature sample set. The multidimensional feature sample set is subjected to a feature selection strategy that combines random forest and XGBoost for redundancy optimization to obtain a core feature subset; The teacher model is trained using the aforementioned subset of core features, and a lightweight student model is obtained through knowledge distillation. The Gaofen-5 satellite remote sensing images, Sentinel-3 satellite images, and water temperature and pH environmental factor data of the same time phase to be predicted are used to form multidimensional feature samples to be predicted according to the construction method of the multidimensional feature sample set. The features corresponding to the core feature subset in the multidimensional feature samples to be predicted are selected and input into the lightweight student model to output the predicted chlorophyll a concentration value. The spatial distribution map of chlorophyll a concentration and the quality control layer are reconstructed and generated.

2. The method for inverting chlorophyll a concentration in water bodies based on multimodal data and a lightweight model according to claim 1, characterized in that, The preprocessing steps for the Gaofen-5 satellite remote sensing images and the Sentinel-3 satellite images include: Radiometric calibration was performed on the Gaofen-5 satellite remote sensing image to convert digital values ​​into radiance values, and atmospheric correction was performed using the FLAASH or 6S atmospheric radiative transfer model to obtain a reflectance image above the water surface. The Sentinel-3 satellite imagery was subjected to water color and atmospheric correction and solar glare correction using a C2RCC processor to obtain a water surface reflectance image, which was then reprojected onto the Gaofen-5 satellite remote sensing imagery to a consistent coordinate system.

3. The method for inverting chlorophyll a concentration in water bodies based on multimodal data and a lightweight model according to claim 1, characterized in that, The steps for extracting water body pixels include: Normalized difference water index was calculated based on water surface reflectance images from Gaofen-5 satellite remote sensing imagery and water surface reflectance images from Sentinel-3 satellite imagery, respectively. The normalized difference water index is compared with a preset water body discrimination threshold, and a binary water body mask layer is generated based on the comparison result. Based on the aforementioned binary mask layer of the water body, pixels of the water body region are extracted to form a subset of the water body image.

4. The method for inverting chlorophyll a concentration in water bodies based on multimodal data and a lightweight model according to claim 1, characterized in that, The steps to form a multidimensional feature sample set include: Acquire ground-based measured chlorophyll a concentration data during the same period of satellite transit; the ground-based measured chlorophyll a concentration data shall include at least the sampling latitude and longitude coordinates, chlorophyll a concentration value, water temperature and pH environmental factor data; The water temperature and pH environmental factor data are spatially interpolated to generate an environmental factor layer with the same spatial resolution as the Gaofen-5 satellite remote sensing image. Spatial registration is performed based on the sampled latitude and longitude coordinates and the water image subset of the Gaofen-5 satellite remote sensing image, the nearest neighbor effective pixels are found and the corresponding hyperspectral indices are extracted; Spatial registration is performed based on the sampled latitude and longitude coordinates and the water image subset of the Sentinel-3 satellite imagery to find the nearest neighbor effective pixels and extract the corresponding aquatic vegetation index; Find the pixels in the environmental factor layer that correspond to the sampled latitude and longitude coordinates and extract the water temperature and pH environmental factor features; The hyperspectral index, the aquatic vegetation index, and the environmental factor features are paired with the corresponding chlorophyll a concentration values ​​to form sample entries, and the multidimensional feature sample set is constructed using all the sample entries.

5. The method for inverting chlorophyll a concentration in water bodies based on multimodal data and a lightweight model according to claim 1, characterized in that, The hyperspectral index is at least one of the candidate hyperspectral indices sensitive to chlorophyll a constructed based on the band ratio, derivative spectrum, and three-band combination of Gaofen-5 satellite remote sensing images; the aquatic vegetation index is the normalized red edge index constructed based on the red edge band and red light band of Sentinel-3 satellite images.

6. The method for inverting chlorophyll a concentration in water bodies based on multimodal data and a lightweight model according to claim 1, characterized in that, The feature selection strategy coupled with random forest and XGBoost includes: Based on the multidimensional feature sample set, a random forest regression model and an XGBoost regression model were trained respectively. Extract the first feature importance list from the random forest regression model, and extract the second feature importance list from the XGBoost regression model; The scores of the first feature importance list and the second feature importance list are normalized and then weighted to obtain the coupling feature importance score. The core feature subset is obtained by sorting the coupling features from high to low importance scores and selecting the Top-K features or selecting features that meet the importance threshold.

7. The method for inverting chlorophyll a concentration in water bodies based on multimodal data and a lightweight model according to claim 1, characterized in that, Both the teacher model and the lightweight student model are models for chlorophyll a concentration inversion trained with the core feature subset as input. The teacher model is a deep neural network model or an XGBoost regression model; the lightweight student model is a neural network model, a pruned decision tree model, or an XGBoost regression model.

8. The method for inverting chlorophyll a concentration in water bodies based on multimodal data and a lightweight model according to claim 1, characterized in that, A lightweight student model is obtained through knowledge distillation training, including: The core feature subset is input into the teacher model to obtain the corresponding teacher model prediction value, and the teacher model prediction value is used as a soft label. Using the measured chlorophyll a concentration on the ground as a hard label, a joint loss function is constructed that simultaneously constrains the error between the predicted value of the lightweight student model and the hard label, as well as the error between the predicted value of the lightweight student model and the soft label. The lightweight student model is trained based on the joint loss function to obtain the lightweight student model.

9. The method for inverting chlorophyll a concentration in water bodies based on multimodal data and a lightweight model according to claim 1, characterized in that, The evaluation and deployment of the lightweight student model include: The chlorophyll a concentration inversion accuracy of the lightweight student model was evaluated using the coefficient of determination, root mean square error, and mean absolute error. When deploying the application, the teacher model is discarded, and only the lightweight student model is used to predict the chlorophyll a concentration of the multidimensional feature sample to be predicted, so as to obtain the predicted chlorophyll a concentration value. Based on the predicted chlorophyll a concentration values, a spatial distribution map of chlorophyll a concentration and a quality control layer are reconstructed and generated.

10. A system for retrieving chlorophyll a concentration in water bodies based on multimodal data and a lightweight model, characterized in that, include: The multimodal data acquisition unit is used to acquire remote sensing images from the Gaofen-5 satellite, Sentinel-3 satellite images, ground-measured chlorophyll a concentration data, and concurrent water temperature and pH environmental factor data covering the target water area. The image preprocessing and water body pixel extraction unit is used to preprocess the Gaofen-5 satellite remote sensing images and the Sentinel-3 satellite images respectively, and extract water body pixels to obtain a subset of water body images. The multimodal index construction and feature fusion unit is used to construct a hyperspectral index based on the Gaofen-5 satellite remote sensing image, construct an aquatic vegetation index based on the Sentinel-3 satellite image, and fuse it with the water temperature, pH environmental factor data, and the spatiotemporal registration of the spectral reflectance of the Gaofen-5 satellite remote sensing image and the Sentinel-3 satellite image to form a multidimensional feature sample set. The coupled feature selection and redundancy optimization unit is used to perform redundancy optimization on the multidimensional feature sample set by using a feature selection strategy coupled with random forest and XGBoost to obtain a core feature subset. The teacher model training and knowledge distillation unit is used to train the teacher model with the core feature subset and to obtain a lightweight student model through knowledge distillation training. The predictable feature generation and inversion output unit is used to generate predictable multidimensional feature samples from the Gaofen-5 satellite remote sensing images, Sentinel-3 satellite images, and water temperature and pH environmental factor data of the same time phase according to the construction method of the multidimensional feature sample set. It selects the features corresponding to the core feature subset in the predictable multidimensional feature samples and inputs them into the lightweight student model, outputs the predicted value of chlorophyll a concentration, and reconstructs and generates the spatial distribution map of chlorophyll a concentration and the quality control layer.

Images