A method for downscaling ndvi based on random forest and nonlinear residual correction
By constructing an NDVI downscaling method based on random forest and nonlinear residual correction, the problems of insufficient nonlinear feature representation and inability to characterize spatial heterogeneity in existing technologies are solved, achieving accurate reconstruction of high-resolution NDVI and improving the accuracy and stability of prediction results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- JILIN UNIVERSITY
- Filing Date
- 2026-02-06
- Publication Date
- 2026-05-29
AI Technical Summary
Existing NDVI downscaling methods are unable to effectively characterize the complex nonlinear relationships between factors such as climate, hydrology, and topography and vegetation, resulting in limitations in high-resolution predictions, low accuracy of prediction results, and inability to fully characterize spatial heterogeneity.
A method based on random forest model and nonlinear residual correction is adopted. By constructing a random forest prediction model driven by multi-source data and combining it with a generalized additive model (GAM) for pixel-by-pixel nonlinear regression correction, systematic bias is captured and corrected, thus achieving high-resolution downscaling of NDVI.
It significantly improves the spatial consistency, temporal continuity and overall accuracy of downscaling results, and enhances the expressive power and robustness in complex ecological environments.
Smart Images
Figure CN122116168A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the technical field of ecological remote sensing and geoscience information, and in particular to an NDVI downscaling method based on random forest and nonlinear residual correction. Background Technology
[0002] Normalized Difference Vegetation Index (NDVI) is widely used for dynamic monitoring of vegetation cover due to its high sensitivity to changes in land surface vegetation. However, the inconsistent accuracy and resolution of different NDVI datasets, and the lack of long-term series data covering the period around 2000 with uniform resolution, make it difficult to meet the needs of regional ecological management and restoration. Therefore, fusing NDVI data from different sensors and spatially downscaling low-resolution NDVI data has become an important research direction in the fields of remote sensing and ecological geosciences.
[0003] Traditional NDVI downscaling methods mainly include statistical regression, interpolation reconstruction based on surface process models, and pixel-mixed decomposition. However, these methods generally struggle to effectively characterize the complex nonlinear relationships between factors such as climate, hydrology, and topography and vegetation, thus limiting their application in high-resolution prediction. In recent years, thanks to the rapid development of remote sensing technology and machine learning methods, these methods have gradually become important approaches to improve the accuracy of NDVI downscaling predictions and have been widely applied in fields such as ecological environment monitoring. Nevertheless, existing downscaling models still suffer from problems such as insufficient selection of feature variables and inaccurate characterization of key driving factors, making it difficult to accurately reconstruct the spatial pattern of high-resolution NDVI. Furthermore, they may exhibit local unevenness, scale inconsistencies, or accumulated biases, thus affecting subsequent analysis of the spatiotemporal variation patterns of vegetation. Summary of the Invention
[0004] The purpose of this invention is to provide an NDVI downscaling method based on random forest and nonlinear residual correction, which solves the technical problems of insufficient nonlinear feature representation, inability to fully characterize spatial heterogeneity, and low prediction accuracy in existing downscaling methods, and has strong expressive power and robustness.
[0005] To achieve the above objectives, this invention provides an NDVI downscaling method based on random forest and nonlinear residual correction, comprising the following steps: S1. Construct a random forest model; S2. Multi-source data preprocessing and feature raster construction to generate a total sample dataset for training and prediction of the random forest model; S3. Train the random forest model to obtain the optimal random forest model; S4, NDVI downscaling prediction: The trained random forest model is applied to a high-resolution feature variable dataset to obtain high-resolution annual average NDVI prediction results. S5. Nonlinear residual correction of the optimal random forest model based on GAM; S6. Apply the constructed nonlinear residual correction model to all downscaled NDVI data to obtain the corrected NDVI results, and evaluate the accuracy of the results using the coefficient of determination, mean absolute error, and root mean square error.
[0006] Preferably, in step S1, low-resolution land use type, temperature, precipitation, elevation, slope, aspect, population density, and groundwater depth are selected as feature variables of the model, and the low-resolution annual average NDVI value is used as the target variable.
[0007] Preferably, in step S2, the multi-source data preprocessing and feature raster construction specifically include the following steps: S21. Using low-resolution NDVI raster data as a spatial reference benchmark, extract its coordinate system, spatial resolution, number of rows and columns, and spatial range as a unified spatial framework for all subsequent feature variables. S22. Through reprojection, resampling and range clipping operations, land use type, temperature, precipitation, elevation, slope, aspect, population density and groundwater depth are aligned to the same spatial grid to achieve consistency of spatial scale and coordinate system and construct a feature grid with unified structure. S23. After completing spatial alignment, the land use classification data is discretized. By constructing a multi-category binary method, each land use type is transformed into an independent binary band. S24. Set the data format, storage type, invalid value encoding and compression method for all feature variables in a unified manner to generate feature raster data with standardized structure. S25. Extract the corresponding attribute values from each feature raster based on the spatial location of the sample points, match them with the target NDVI, and construct the total sample dataset for training and prediction of the random forest model.
[0008] Preferably, in step S3, training the random forest model specifically includes the following steps: S31. Randomly divide the total sample dataset into a training set and an independent validation set; S32. Within the training set, grid search combined with five-fold cross-validation is used to fine-tune the key parameters of the random forest model; S33. Evaluate whether the random forest model is overfitted or underfitted using an independent validation set. The optimal random forest model is obtained when the model fits appropriately.
[0009] Preferably, in step S32, the key parameters of the random forest model include the number of decision trees, the maximum depth of the trees, the minimum number of samples required for internal node partitioning, and the minimum number of samples required for leaf nodes.
[0010] Preferably, in step S4, the feature raster data of the target area is processed cyclically by year. For each year, the corresponding multi-band feature array is read, and invalid values are masked. Only valid pixels are input into the optimal random forest model for prediction.
[0011] Preferably, in step S5, a pixel-by-pixel GAM nonlinear regression model is constructed using the residual between the high-resolution downscaling results of overlapping years and the high-resolution observed NDVI. Subsequently, the GAM nonlinear regression model is independently validated using the high-resolution data of the remaining years that did not participate in the training, in order to correct the systematic bias of the random forest downscaling results.
[0012] Preferably, constructing and correcting a pixel-by-pixel GAM nonlinear regression model specifically includes the following: constructing time-series training samples for each pixel, calculating the residual between the predicted NDVI and observed NDVI of the training year; fitting a generalized additive model to each pixel individually to capture the nonlinear relationship between the predicted value and the residual of that pixel; after fitting, inputting the predicted NDVI of the target year into the GAM model of the corresponding pixel to obtain the predicted residual of that pixel, and adding it back to the original predicted value, thereby achieving pixel-level NDVI correction.
[0013] Preferably, in step S6, the specific formulas for the coefficient of determination, mean absolute error, and root mean square error are as follows: ; ; ; in, Represents the coefficient of determination. Represents the mean absolute error. Represents the root mean square error. Represents the number of samples. Represents the original NDVI value. This represents the downscaled NDVI value. This represents the original average NDVI value.
[0014] Therefore, the NDVI downscaling method based on random forest and nonlinear residual correction described above has the following advantages: (1) In this invention, by constructing a random forest prediction model driven by multi-source data, it is possible to effectively characterize the complex nonlinear relationship between multidimensional feature variables such as meteorology, hydrology, topography and land use and NDVI, and solve the technical problems of insufficient nonlinear feature expression and inability to fully characterize spatial heterogeneity in existing downscaling methods.
[0015] (2) In this invention, GAM (Generalized Additive Model) is used to perform nonlinear residual correction on the prediction results. By constructing a pixel-by-pixel nonlinear regression model to capture and correct systematic biases, the spatial consistency, temporal continuity and overall accuracy and stability of the downscaling results are significantly improved. It has stronger expressive power and robustness under complex ecological environment conditions.
[0016] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description
[0017] Figure 1 This is a flowchart illustrating the NDVI downscaling method based on random forest and nonlinear residual correction of the present invention. Figure 2 This is a diagram showing the spatial distribution ratio of regional NDVI in an embodiment of the present invention. Figure 3 This is an NDVI prediction-observation scatter plot of the resampling residual correction method in this embodiment of the invention; Figure 4 The NDVI prediction-observation scatter plot of the method of the present invention provided in the embodiments of the present invention. Detailed Implementation
[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Specific model specifications need to be selected and determined according to the actual specifications of the device, etc. The specific selection calculation method adopts existing technology in the art, and therefore will not be described in detail.
[0019] Example like Figure 1 As shown, this invention provides an NDVI downscaling method based on random forest and nonlinear residual correction, comprising the following steps: S1. Construct a random forest model, selecting low-resolution land use type, temperature, precipitation, elevation, slope, aspect, population density and groundwater depth as feature variables of the model, and using the low-resolution annual average NDVI as the target variable. S2. Multi-source data preprocessing and feature raster construction to generate a total sample dataset for training and prediction of the random forest model; S21. Using low-resolution NDVI raster data as a spatial reference benchmark, extract its coordinate system, spatial resolution, number of rows and columns, and spatial range as a unified spatial framework for all subsequent feature variables. S22. Through reprojection, resampling and range clipping operations, land use type, temperature, precipitation, elevation, slope, aspect, population density and groundwater depth are aligned to the same spatial grid to achieve consistency of spatial scale and coordinate system and construct a feature grid with unified structure. S23. After completing spatial alignment, the land use classification data is discretized. By constructing a multi-category binary method, each land use type is transformed into an independent binary band. S24. Set the data format, storage type, invalid value encoding and compression method for all feature variables in a unified manner to generate feature raster data with standardized structure. S25. Extract the corresponding attribute values from each feature grid according to the spatial location of the sample points, match them with the target NDVI, and construct the total sample dataset for training and prediction of the random forest model. S3. Train the random forest model to obtain the optimal random forest model; S31. Randomly divide the total sample dataset into a training set and an independent validation set; S32. Within the training set, grid search combined with five-fold cross-validation is used to fine-tune the key parameters of the random forest model. The key parameters include the number of decision trees, the maximum depth of the trees, the minimum number of samples required for internal node splitting, and the minimum number of samples required for leaf nodes. S33. Evaluate whether the random forest model is overfitted or underfitted through an independent validation set. When the model fits appropriately, the optimal random forest model is obtained. Specifically, calculate the determination coefficient, root mean square error, and mean absolute error for each fold, and calculate their average values to comprehensively evaluate the stability and generalization ability of the model. S4, NDVI downscaling prediction: The trained random forest model is applied to a high-resolution feature variable dataset to obtain high-resolution annual average NDVI prediction results. First, the pre-trained random forest model is loaded. Then, the feature raster data of the target area is processed cyclically by year. For each year, the corresponding multi-band feature array is read, and invalid values are masked, so only valid pixels are input into the model for prediction. At the same time, information such as coordinate reference system, spatial range and pixel resolution is obtained through reference raster to ensure that the prediction results have correct spatial geographic attributes. Finally, the NDVI raster predicted each year is saved to provide continuous high-resolution spatial data for subsequent analysis. S5. Nonlinear residual correction of the optimal random forest model based on GAM; By using the residuals between the high-resolution downscaling results of overlapping years and the high-resolution observed NDVI, a pixel-by-pixel GAM nonlinear regression model was constructed. Then, the GAM nonlinear regression model was independently validated using the high-resolution data of the remaining years that did not participate in the training, in order to correct the systematic bias of the random forest downscaling results. The construction and correction of a pixel-by-pixel GAM nonlinear regression model specifically includes the following: constructing time-series training samples for each pixel, calculating the residual between the predicted NDVI and observed NDVI of the training year; fitting a generalized additive model for each pixel individually to capture the nonlinear relationship between the predicted value and the residual of that pixel; after fitting, inputting the predicted NDVI of the target year into the GAM model of the corresponding pixel to obtain the predicted residual of that pixel, and adding it back to the original predicted value, thereby achieving pixel-level NDVI correction. S6. Apply the constructed nonlinear residual correction model to all downscaled NDVI data to obtain the corrected NDVI results. Evaluate the accuracy of the results using the coefficient of determination, mean absolute error, and root mean square error. The specific formulas for the coefficient of determination, mean absolute error, and root mean square error are as follows: ; ; ; in, Represents the coefficient of determination. Represents the mean absolute error. Represents the root mean square error. Represents the number of samples. Represents the original NDVI value. This represents the downscaled NDVI value. This represents the original average NDVI value.
[0020] In this embodiment, the study area is located in the Songnen Plain within Jilin Province, characterized by high elevations on the east and west sides and a lower elevation in the middle. The area belongs to the North Temperate Continental Monsoon Climate Zone, primarily exhibiting semi-arid characteristics. Soil salinization is a prominent issue, and the ecosystem structure is relatively fragile. This region is an important agricultural and pastoral base and an ecologically vulnerable area. Groundwater levels are generally shallower than 8 meters, falling within the usable range for groundwater-dependent ecosystems, which provide a vital water source for the region's ecosystem.
[0021] The NDVI data for the study area mainly came from multi-temporal normalized vegetation index products acquired by different sensors, covering the period from 1985 to 2022. Due to limitations in sensor conditions and data acquisition methods, the spatial resolution of NDVI data before 2000 was relatively low. Although it could reflect regional-scale vegetation cover change trends, it was difficult to characterize fine spatial differences. The spatial resolution of NDVI data after 2000 was higher, and the data quality and ability to express spatial details were significantly improved, making it suitable for high-resolution vegetation cover feature analysis.
[0022] First, using the annual mean GIMMS-NDVI data at 8km resolution from 1985 to 2013 as the target variable, and combining it with characteristic variables such as land use type, temperature, precipitation, elevation, slope, aspect, population density, and groundwater depth during the same period, a random forest model was constructed to learn the nonlinear mapping relationship between NDVI and multiple driving factors. The sample data was divided into training and validation sets in a 7:3 ratio, and the model parameters were optimized using cross-validation to obtain the optimal random forest model. After training the random forest model, it was applied to a 1km resolution dataset of characteristic variables to perform spatial downscaling prediction of NDVI from 1985 to 2013, obtaining initial results for the annual mean NDVI at 1km resolution. Since the random forest model may have systematic biases in different time periods, a nonlinear residual correction method based on GAM (Generalized Additive Model) was further introduced.
[0023] Residuals from 1km GIMMS-NDVI results obtained by downscaling random forest data between 2000 and 2010 were selected, and MODIS-NDVI observation data from the same period were used to construct a generalized additive model pixel-by-pixel to characterize the nonlinear relationship between the predicted results and the residuals. The residual correction model was independently validated using data from 2011 to 2013 to evaluate its generalization ability in years not involved in modeling. After verifying its stability, the residual correction model was further applied to downscaled NDVI results before 2000 to achieve systematic bias correction of historical NDVI data. Figure 2 As shown, an NDVI dataset with uniform spatial resolution (1km), good temporal continuity, and high consistency with MODIS-NDVI was constructed from 1985 to 2022, providing a reliable data foundation for regional-scale ecological monitoring and analysis.
[0024] To evaluate the applicability and effectiveness of the method proposed in this invention, two different NDVI downscaling and correction schemes were compared: one method involves directly resampling GIMMS-NDVI data to a 1km spatial resolution and then correcting its residuals; the other is the downscaling method proposed in this invention, based on a random forest model combined with nonlinear residual correction. The annual average NDVI results obtained by both methods were compared with MODIS-NDVI data, and the correction effect was evaluated using indicators such as the coefficient of determination, root mean square error, and mean absolute error. Figure 3 As shown, the resampling residual correction method has a certain fitting ability. The value is approximately 0.82; however, the method of this invention exhibits superior overall performance, such as... Figure 4 As shown, The accuracy was increased to approximately 0.91, an improvement of approximately 0.09, while RMSE and MAE decreased by approximately 0.008 and 0.007 respectively, significantly improving the accuracy and stability of the downscaling results.
[0025] Therefore, this invention adopts the NDVI downscaling method based on random forest and nonlinear residual correction to solve the technical problems of insufficient nonlinear feature expression, inability to fully characterize spatial heterogeneity, and low prediction accuracy in existing downscaling methods. It has strong expressive power and robustness.
[0026] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.
Claims
1. An NDVI downscaling method based on random forest and nonlinear residual correction, characterized by: Includes the following steps: S1. Construct a random forest model; S2. Multi-source data preprocessing and feature raster construction to generate a total sample dataset for training and prediction of the random forest model; S3. Train the random forest model to obtain the optimal random forest model; S4, NDVI downscaling prediction: The trained random forest model is applied to a high-resolution feature variable dataset to obtain high-resolution annual average NDVI prediction results. S5. Nonlinear residual correction of the optimal random forest model based on GAM; S6. Apply the constructed nonlinear residual correction model to all downscaled NDVI data to obtain the corrected NDVI results, and evaluate the accuracy of the results using the coefficient of determination, mean absolute error, and root mean square error.
2. The NDVI downscaling method based on random forest and nonlinear residual correction according to claim 1, characterized in that: In step S1, low-resolution land use type, temperature, precipitation, elevation, slope, aspect, population density, and groundwater depth are selected as feature variables of the model, and the low-resolution annual average NDVI value is used as the target variable.
3. The NDVI downscaling method based on random forest and nonlinear residual correction according to claim 2, characterized in that: In step S2, the multi-source data preprocessing and feature raster construction specifically include the following steps: S21. Using low-resolution NDVI raster data as a spatial reference benchmark, extract its coordinate system, spatial resolution, number of rows and columns, and spatial range as a unified spatial framework for all subsequent feature variables. S22. Through reprojection, resampling and range clipping operations, land use type, temperature, precipitation, elevation, slope, aspect, population density and groundwater depth are aligned to the same spatial grid to achieve consistency of spatial scale and coordinate system and construct a feature grid with unified structure. S23. After completing spatial alignment, the land use classification data is discretized. By constructing a multi-category binary method, each land use type is transformed into an independent binary band. S24. Set the data format, storage type, invalid value encoding and compression method for all feature variables in a unified manner to generate feature raster data with standardized structure. S25. Extract the corresponding attribute values from each feature raster based on the spatial location of the sample points, match them with the target NDVI, and construct the total sample dataset for training and prediction of the random forest model.
4. The NDVI downscaling method based on random forest and nonlinear residual correction according to claim 3, characterized in that: Step S3, training the random forest model specifically includes the following steps: S31. Randomly divide the total sample dataset into a training set and an independent validation set; S32. Within the training set, grid search combined with five-fold cross-validation is used to fine-tune the key parameters of the random forest model; S33. Evaluate whether the random forest model is overfitted or underfitted using an independent validation set. The optimal random forest model is obtained when the model fits appropriately.
5. The NDVI downscaling method based on random forest and nonlinear residual correction according to claim 4, characterized in that: In step S32, the key parameters of the random forest model include the number of decision trees, the maximum depth of the trees, the minimum number of samples required to split internal nodes, and the minimum number of samples required for leaf nodes.
6. The NDVI downscaling method based on random forest and nonlinear residual correction according to claim 5, characterized in that: In step S4, the feature raster data of the target area is processed cyclically by year. For each year, the corresponding multi-band feature array is read, and invalid values are masked. Only valid pixels are input into the optimal random forest model for prediction.
7. The NDVI downscaling method based on random forest and nonlinear residual correction according to claim 6, characterized in that: In step S5, a pixel-by-pixel GAM nonlinear regression model is constructed using the residual between the high-resolution downscaling results of overlapping years and the high-resolution observed NDVI. Then, the GAM nonlinear regression model is independently validated using the high-resolution data of the remaining years that did not participate in the training, in order to correct the systematic bias of the random forest downscaling results.
8. The NDVI downscaling method based on random forest and nonlinear residual correction according to claim 7, characterized in that: Constructing and correcting a pixel-by-pixel GAM nonlinear regression model specifically includes the following: For each pixel, a time series training sample is constructed, and the residual between the predicted NDVI and the observed NDVI of the training year is calculated. A generalized additive model is fitted to each pixel individually to capture the nonlinear relationship between the predicted value and the residual of that pixel. After fitting is complete, the predicted NDVI of the target year is input into the GAM model of the corresponding pixel to obtain the prediction residual of that pixel, and then added back to the original prediction value to achieve pixel-level NDVI correction.
9. The NDVI downscaling method based on random forest and nonlinear residual correction according to claim 8, characterized in that: In step S6, the specific formulas for the coefficient of determination, mean absolute error, and root mean square error are as follows: ； ； ； in, Represents the coefficient of determination. Represents the mean absolute error. Represents the root mean square error. Represents the number of samples. Represents the original NDVI value. This represents the downscaled NDVI value. This represents the original average NDVI value.