Abandoned farmland identification method based on multi-source remote sensing data and time series correction

Abstract

The present invention discloses a method for identifying abandoned farmland based on multi-source remote sensing data and time series correction, comprising the following steps: Step 1: Data acquisition and preprocessing; Step 2: Extraction of key phenological periods; Step 3: Construction of a feature library; Step 4: Training and prediction of a random forest model; Step 5: Post-processing and accuracy assessment. The present invention first integrates multi-source remote sensing image data and performs strict preprocessing on it to ensure the temporal and spatial consistency of the data. By in-depth analysis of the vegetation minimum nutrient period and nutrient peak period of the NDVI and NDSI time series, key phenological period information is accurately extracted. In the feature construction stage, multiple spectral indices are integrated, and principal component analysis and feature importance screening are used to optimize feature dimensions and improve model performance. A random forest model is used for training and prediction, combined with oversampling and category weight adjustment techniques, to ultimately generate a high-precision spatial distribution map of abandoned land.

Description

Technical Field

[0001] The present invention relates to the field of abandoned farmland identification technology, and in particular to a method for identifying abandoned farmland based on multi-source remote sensing data and time series correction. Background Art

[0002] As a core link in ensuring food security and ecological balance, the technological evolution of abandoned farmland monitoring has always faced multi-dimensional challenges. Traditional monitoring methods are highly dependent on manual statistics and field surveys, which are not only costly and time-limited, but also difficult to capture the spatiotemporal heterogeneity of cultivated land use. With the popularization of remote sensing technology, monitoring methods based on satellite images have gradually become mainstream, but their application efficiency is constrained by multiple technical bottlenecks. First, existing studies mostly rely on medium- and low-spatial resolution data. In hilly and mountainous areas where fragmented cultivated land is widely distributed, the mixed pixel effect leads to a significant reduction in the identification accuracy of small-scale abandoned land plots; second, although traditional time series analysis methods can characterize surface cover dynamics, the problem of accumulated classification errors caused by the lack of data continuity in cloudy and rainy climate zones has not been effectively solved; third, existing algorithms are still insufficient in exploring phenological characteristics and fail to fully integrate the differences in spectral-temporal responses of key phenological periods, which restricts the ability to distinguish abandoned land from fallow land. Fourth, the classification model driven by a single data source has poor generalization in complex surface cover scenarios, and is particularly prone to misjudgment in vegetation succession transition zones.

[0003] Current remote sensing monitoring technology for abandoned farmland is mainly based on the coordinated optimization of multi-source remote sensing data and algorithms. Specific technical paths can be divided into the following categories:

[0004] 1. Spectral analysis based on phenological characteristics. This method extracts differential NDVI and NDSI characteristics during key phenological periods and combines them with a temporal threshold (e.g., NDVI consistently below 0.2) to identify abandoned land. This method performs well in continuously cultivated plains, but is limited by the mixed pixel effects of low- and medium-resolution imagery, resulting in a high error rate in identifying fragmented plots.

[0005] 2. Random Forest Classification Model. By integrating multispectral bands, vegetation indices (EVI, LSWI), and texture features, it significantly improves the classification accuracy of complex land cover. However, its reliance on the continuity of optical imagery limits its applicability in cloudy and rainy areas.

[0006] 3. Multi-source data fusion technology: Combining optical remote sensing data with radar data, the advantage of radar's cloud penetration compensates for the lack of optical data. However, the low spatial resolution of radar data limits the precise identification of small-scale abandoned land.

[0007] 4. Time series correction algorithm. This algorithm uses a sliding window to remove isolated pixels with sudden changes over multiple years and optimizes the results by combining a priori rules (e.g., if the land has been abandoned for ≥ 2 years). However, static correction rules are difficult to adapt to the annual dynamic changes in phenological characteristics, leading to long-term error accumulation.

[0008] The shortcomings of the above-mentioned prior art are:

[0009] 1. Resolution and mixed pixel issues. Using medium- and low-resolution data such as MODIS or Landsat makes it difficult to accurately identify the small-scale, fragmented, abandoned land parcels that are prevalent in southern China's hilly regions, leading to significant underestimation of the area in the classification results.

[0010] 2. Data loss due to interference. Traditional optical imagery lacks effective observation frequency in cloudy and rainy areas, resulting in time series interruptions. This in turn causes abnormal fluctuations in vegetation indices such as NDVI and EVI, leading to the misidentification of periods of data loss as abandoned land.

[0011] 3. Cumulative effect. Existing time series correction algorithms rely solely on static rules and fail to consider the annual dynamic changes in phenological characteristics, resulting in classification errors being gradually amplified during long-term monitoring.

[0012] 4. Insufficient generalization capability. Classification models based on a single data source, such as optical remote sensing or radar data, struggle to cope with complex land cover scenarios (e.g., transition from abandoned land to shrubland), and their classification accuracy drops significantly in vegetation transition zones.

[0013] The above technical bottlenecks seriously restrict the accuracy, timeliness and large-scale promotion potential of abandoned farmland monitoring. Summary of the Invention

[0014] In view of this, the present invention addresses the deficiencies in the existing technology, and its main purpose is to provide a method for identifying abandoned farmland based on multi-source remote sensing data and time series correction, which aims to achieve efficient and accurate monitoring of the phenomenon of abandoned farmland through systematic data processing and analysis technology.

[0015] To achieve the above object, the present invention adopts the following technical solutions:

[0016] A method for identifying abandoned farmland based on multi-source remote sensing data and time series correction includes the following steps:

[0017] Step 1: Data acquisition and preprocessing. The specific operations are as follows:

[0018] 1.1. Remote sensing data acquisition: Select multi-source remote sensing image data covering the growing season data of the target area;

[0019] 1.2. Cloud Mask and Atmospheric Correction: When using optical remote sensing data, cloud masking algorithms are first used to remove cloud interference to ensure image clarity. Next, atmospheric correction is performed to eliminate the effects of atmospheric scattering and absorption on the image, ensuring spectral consistency of the data.

[0020] 1.3. Geometric Registration and Accuracy Correction: All images are geometrically registered to ensure spatial alignment between different images and temporal and spatial consistency. In addition, the images are radiometrically corrected to eliminate sensor differences and atmospheric effects to ensure data accuracy.

[0021] 1.4. Radar image processing: Radar data is denoised using Refined Lee filtering, and further radiometric calibration and terrain correction are performed. The processed radar images are used to extract backscatter coefficients, which facilitates analysis of different ground object types.

[0022] 1.5. Data integration and spatiotemporal consistency: By fusing optical and radar images, we ensure the spatiotemporal consistency of data from different sources, providing reliable input data for subsequent analysis.

[0023] Step 2: Extraction of key phenological periods: Based on the NDVI and NDSI time series, two sensitive periods are extracted through filtering, smoothing, and extreme value detection algorithms: the minimum nutrient period and the peak nutrient period. The phenological period positioning error is controlled within 5 days to ensure the accuracy of the time series analysis.

[0024] Step 3: Feature library construction: Integrate multi-source remote sensing data and spectral indices to construct a multidimensional feature set. Through principal component analysis and feature importance screening, optimize feature dimensions and improve the discrimination efficiency of the classification model.

[0025] Step 4: Random Forest Model Training and Prediction: Using the random forest algorithm, we input a multi-source feature set and optimize model parameters through cross-validation. We also apply oversampling and class weight adjustment to address sample imbalance. The model outputs a probability map of abandoned land, which is then segmented using a threshold to generate a binary classification result, enabling regional-scale mapping of the spatial distribution of abandoned land.

[0026] Step 5: Post-processing and accuracy assessment: Spatial filtering is performed on the classification results to remove noise and generate a refined abandoned land distribution map; the confusion matrix is ​​calculated based on an independent test set to evaluate the overall accuracy, Kappa coefficient and category-specific indicators, analyze the main sources of misjudgment and propose the optimization results of temporal continuity rules; the final results meet the needs of high-resolution monitoring and support decision-making applications.

[0027] Preferably, in step 1.1, Landsat satellite images are used to obtain remote sensing data, and Landsat-5, Landsat-7 and Landsat-8 multi-temporal optical remote sensing image data and Sentinel-2 optical image data and Sentinel-1 radar image data are obtained from Google Earth Engine.

[0028] Preferably, the specific operations of step 2 are as follows:

[0029] 2.1. NDVI and NDSI time series analysis: Using NDVI and NDSI time series data, we extracted key phenological periods using an extreme value detection algorithm. NDVI reflects vegetation growth, while NDSI is highly sensitive to bare soil.

[0030] 2.2. Sensitive Period Extraction: By analyzing changes in NDVI and NDSI, the minimum nutrient period and the peak nutrient period are extracted. LVP is usually the stage with the least vegetation cover and significant bare soil, while PVP is the stage with saturated vegetation cover and can reflect the maximum vegetation growth state of cultivated land.

[0031] 2.3. Error control: When extracting phenological periods, smoothing filtering and extreme value detection algorithms are used to ensure that the phenological period positioning error is controlled within 5 days, thus ensuring the accuracy of time series analysis;

[0032] 2.4. Phenological period correction: In order to accurately calibrate LVP and PVP, corrections are performed based on the growth cycles of different crops to ensure that the extracted phenological periods adapt to changes in different regions and climatic conditions.

[0033] Preferably, the specific operations of step 3 are as follows:

[0034] 3.1. Fusion of spectral index and remote sensing data: Combining multiple spectral indices such as NDVI, EVI, and BSI with different types of remote sensing images to construct a multidimensional feature set;

[0035] 3.2. Feature dimension optimization: Principal component analysis and feature importance screening algorithms are used to remove redundant features and retain only the features that are most effective in distinguishing different landform types;

[0036] 3.3. Feature Importance Screening: The algorithm calculates the impact of different features on the classification results and selects the most representative features. This process can significantly improve the discrimination ability of the classification model and reduce the impact of noise data on the classification results.

[0037] 3.4. Feature library expansion and update: As remote sensing data is continuously updated, the feature library will also be continuously expanded to ensure that it matches the latest remote sensing image data.

[0038] Preferably, the specific operations of step 4 are as follows:

[0039] 4.1. Model Training and Cross-Validation: We used a random forest algorithm for training, inputting a multi-source feature set, and optimized the number and depth of decision trees through cross-validation. We also adjusted the parameters of the decision trees to ensure that the model could accurately distinguish between different land cover types.

[0040] 4.2. Addressing sample imbalance: Use oversampling and class weight adjustment techniques to ensure that different categories of cultivated land and non-cultivated land are adequately represented in model training;

[0041] 4.3. Model Output and Classification Results: The trained RF model outputs a probability map of abandoned land, which is converted into a binary classification result by setting a threshold. Based on this classification result, an accurate spatial distribution map of abandoned land can be generated.

[0042] 4.4. Model Evaluation and Optimization: Evaluate the classification accuracy of the model by comparing it with an independent test set; accuracy indicators include overall accuracy, Kappa coefficient, and category accuracy. Based on these indicators, further optimize the model parameters to improve classification accuracy.

[0043] Preferably, the specific operations of step 5 are as follows:

[0044] 5.1. Spatial filtering and noise removal: Spatial filtering is performed on the classification results to remove noise and generate a refined abandoned land distribution map;

[0045] 5.2. Accuracy Assessment and Error Analysis: Use the confusion matrix method to evaluate the accuracy of classification results, calculate the overall accuracy, Kappa coefficient, and user and producer accuracy for each category; analyze different sources of misclassification and propose improvement suggestions and correction plans;

[0046] 5.3. Accuracy Improvement and Model Optimization: Based on the accuracy assessment results, the model will be optimized, especially the distinction between abandoned land and non-cultivated land will be further adjusted. Misclassifications will be corrected to further improve the robustness of the model.

[0047] 5.4. Final results generation: Based on the revised classification results, a wasteland monitoring map is generated, and decision-making support data is provided to help local governments and relevant departments formulate reasonable land management and ecological protection strategies.

[0048] Compared with the prior art, the present invention has obvious advantages and beneficial effects. Specifically, it can be seen from the above technical solution that:

[0049] The present invention first integrates multi-source remote sensing image data and performs strict preprocessing on it to ensure the temporal and spatial consistency of the data. Through in-depth analysis of NDVI and NDSI time series, key phenological period information is accurately extracted. In the feature construction stage, multiple spectral indices are integrated, and principal component analysis and feature importance screening are used to optimize feature dimensions and improve model performance. The random forest model is used for training and prediction, combined with oversampling and category weight adjustment techniques to solve the sample imbalance problem and ultimately generate a high-precision spatial distribution map of abandoned land. Post-processing and precision evaluation are used to ensure the reliability and accuracy of the classification results. This method effectively overcomes the shortcomings of traditional monitoring technology in terms of scientificity, efficiency and precision, and provides strong technical support for the protection and management of cultivated land. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 It is a processing flow chart of the present invention. DETAILED DESCRIPTION

[0051] The present invention discloses a method for identifying abandoned farmland based on multi-source remote sensing data and time series correction, comprising the following steps:

[0052] Step 1: Data acquisition and preprocessing. Data acquisition and preprocessing are the foundation of this invention and involve the integration, cleaning, and correction of remote sensing image data. In order to accurately monitor abandoned land, it is necessary to select appropriate remote sensing data and perform rigorous preprocessing on it. The specific operations are as follows:

[0053] 1.1. Remote Sensing Data Acquisition: Multi-source remote sensing image data covering the growing season of the target area was selected. Landsat satellite imagery was used for remote sensing data acquisition, which provides data with a spatial resolution of 30 meters, can effectively capture surface features, and is suitable for abandoned land monitoring. Multi-temporal optical remote sensing image data such as Landsat-5, Landsat-7, and Landsat-8, as well as Sentinel-2 optical imagery and Sentinel-1 radar imagery data were acquired from Google Earth Engine (GEE), covering a long time scale (1989-2021).

[0054] 1.2. Cloud Mask and Atmospheric Correction: When using optical remote sensing data, cloud masking algorithms are first used to remove cloud interference to ensure image clarity. Next, atmospheric correction is performed to eliminate the effects of atmospheric scattering and absorption on the image and ensure the spectral consistency of the data.

[0055] 1.3. Geometric Registration and Accuracy Correction: All images are geometrically registered to ensure spatial alignment between different images and temporal and spatial consistency. In addition, the images are radiometrically corrected to eliminate sensor differences and atmospheric effects to ensure data accuracy.

[0056] 1.4. Radar image processing: Radar data is denoised using Refined Lee filtering, and further radiometric calibration and terrain correction are performed. The processed radar images are used to extract backscatter coefficients, which facilitates the analysis of different land feature types.

[0057] 1.5. Data integration and spatiotemporal consistency: By fusing optical images with radar images, the spatiotemporal consistency of data from different sources is ensured, providing reliable input data for subsequent analysis.

[0058] Step 2: Extract key phenological periods: Based on the NDVI and NDSI time series, two sensitive periods are extracted through filtering, smoothing, and extreme value detection algorithms: the minimum nutrient period (LVP, where the lowest NDVI value reflects the characteristics of bare soil) and the peak nutrient period (PVP, where the peak NDVI value indicates saturated vegetation cover). The phenological period positioning error is controlled within 5 days to ensure the accuracy of the time series analysis. The specific steps are as follows:

[0059] 2.1. NDVI and NDSI time series analysis: NDVI (normalized difference vegetation index) and NDSI (bare soil index) time series data were used to extract key phenological periods using an extreme value detection algorithm. NDVI reflects vegetation growth, while NDSI is highly sensitive to bare soil.

[0060] 2.2. Sensitive period extraction: By analyzing the changes in NDVI and NDSI, the minimum vegetative period (LVP) and the peak vegetative period (PVP) are extracted. The LVP is usually the stage with the least vegetation cover and significant bare soil, while the PVP is the stage with saturated vegetation cover and can reflect the maximum vegetation growth state of cultivated land.

[0061] 2.3. Error control: When extracting phenological periods, smoothing filtering and extreme value detection algorithms are used to ensure that the phenological period positioning error is controlled within 5 days, thereby ensuring the accuracy of time series analysis.

[0062] 2.4. Phenological period correction: In order to accurately calibrate LVP and PVP, corrections are performed based on the growth cycles of different crops to ensure that the extracted phenological periods adapt to changes in different regions and climatic conditions.

[0063] Step 3: Feature library construction: Integrate multi-source remote sensing data and spectral indices such as NDVI, EVI, and BSI to construct a multidimensional feature set. Through principal component analysis (PCA) and feature importance screening, optimize feature dimensions and improve the discrimination efficiency of the classification model. The specific operations are as follows:

[0064] 3.1. Fusion of Spectral Indices and Remote Sensing Data: A multidimensional feature set is constructed by combining multiple spectral indices such as NDVI, EVI, and BSI, as well as different types of remote sensing images (optical and radar data). These features can describe the different change patterns of the surface in detail, helping to improve classification accuracy.

[0065] 3.2. Feature dimension optimization: Principal component analysis and feature importance screening algorithms are used to remove redundant features and retain only the features that can best distinguish different landform types to improve the efficiency of the classification model.

[0066] 3.3. Feature importance screening: The algorithm calculates the impact of different features on the classification results and selects the most representative features. This process can significantly improve the discrimination ability of the classification model and reduce the impact of noise data on the classification results.

[0067] 3.4. Feature library expansion and update: As remote sensing data is continuously updated, the feature library will also be continuously expanded to ensure that it matches the latest remote sensing image data, thereby improving the adaptability of the model and the reliability of long-term monitoring.

[0068] Step 4: Random Forest Model Training and Prediction: Using the random forest algorithm, we input a multi-source feature set and optimize model parameters (number and depth of decision trees) through cross-validation. We also apply oversampling and class weight adjustment to address sample imbalance. The model outputs a probability map of abandoned land, which is then segmented using a threshold to generate a binary classification result, enabling regional-scale mapping of the spatial distribution of abandoned land. The specific steps are as follows:

[0069] 4.1. Model training and cross-validation: A random forest algorithm is used for training, with multi-source feature sets as input. The number and depth of decision trees are optimized through cross-validation. The parameters of the decision trees are adjusted to ensure that the model can accurately distinguish between different land cover types.

[0070] 4.2. Addressing sample imbalance: Use oversampling and class weight adjustment techniques to ensure that different categories of cultivated land and non-cultivated land are adequately represented in model training.

[0071] 4.3. Model Output and Classification Results: The trained RF model outputs a probability map of abandoned land, which is converted into a binary classification result by setting a threshold. Based on this classification result, an accurate spatial distribution map of abandoned land can be generated.

[0072] 4.4. Model Evaluation and Optimization: Evaluate the classification accuracy of the model by comparing it with an independent test set; accuracy indicators include overall accuracy, Kappa coefficient, and category accuracy. Based on these indicators, further optimize the model parameters to improve classification accuracy.

[0073] Step 5: Post-processing and accuracy assessment: Spatial filtering is performed on the classification results to remove noise and generate a refined abandoned land distribution map. The confusion matrix is ​​calculated based on an independent test set to evaluate the overall accuracy, Kappa coefficient, and category-specific indicators. The main sources of misjudgment are analyzed and the optimization results of temporal continuity rules are proposed. The final results meet the needs of high-resolution monitoring and support decision-making applications. The specific operations are as follows:

[0074] 5.1. Spatial Filtering and Noise Removal: Spatial filtering is performed on the classification results to remove noise and generate a refined map of abandoned land distribution. Spatial filtering can effectively smooth the classification results, reduce the impact of misclassification, and ensure the clarity and accuracy of the final map.

[0075] 5.2. Accuracy Assessment and Error Analysis: Use the confusion matrix method to evaluate the accuracy of the classification results, calculate the overall accuracy, Kappa coefficient, and the user accuracy and producer accuracy of each category; and propose improvement suggestions and correction plans by analyzing different sources of misjudgment.

[0076] 5.3. Accuracy Improvement and Model Optimization: Based on the accuracy assessment results, the model is optimized, especially the distinction between abandoned land and non-cultivated land is further adjusted. Misclassifications (such as grassland or shrubs being mistakenly identified as abandoned land) are corrected to further improve the robustness of the model.

[0077] 5.4. Final results generation: Based on the revised classification results, a wasteland monitoring map is generated, and decision-making support data is provided to help local governments and relevant departments formulate reasonable land management and ecological protection strategies. Example

[0078] 1. Data acquisition: Obtain Landsat-7 / 8, Sentinel-2 optical imagery, and Sentinel-1 radar imagery data for Sites A and B from 2019 to 2024. Perform cloud detection and declouding on the optical imagery to ensure data integrity during key phenological periods.

[0079] 2. Extraction of key phenological periods: Based on the time series analysis of NDVI and NDSI, data during LVP and PVP were extracted.

[0080] 3. Feature Library Construction: Combining multi-band data from optical and radar images, we constructed a spectral feature library for abandoned and non-abandoned land. We incorporated multiple spectral indices such as NDVI, EVI, GCVI, BSI, and LSWI as features.

[0081] 4. Model Training and Prediction: The model is trained using a random forest algorithm, with input features including multi-band optical imagery data, backscatter coefficients from radar imagery, and multiple spectral indices. The trained model is used to predict the amount of abandoned farmland in 2024 and generate a map of abandoned land distribution.

[0082] 5. Accuracy Assessment: The classification accuracy of the model was evaluated using the confusion matrix and Kappa coefficient. The results showed that the overall accuracy of site A was 91.06%, and the Kappa coefficient was 82.14%; the overall accuracy of site B was 93.96%, and the Kappa coefficient was 81.75%.

[0083] The method of the present invention has the following advantages:

[0084] 1. Significantly improved classification accuracy: By combining the phenological characteristics of optical images with the anti-interference capabilities of radar data, it effectively distinguishes abandoned and fallow land, reduces the risk of misjudgment of mixed pixels and vegetation transition zones, and achieves high-confidence spatial recognition.

[0085] 2. Enhanced resistance to environmental interference: Utilizing the cloud-penetrating properties of radar data, the system compensates for the data loss of optical images in cloudy and rainy areas, ensuring monitoring continuity under complex climatic conditions. At the same time, the red-edge band enhances the sensitivity to the physiological state of vegetation.

[0086] 3. Computing efficiency and scalability optimization: Based on a cloud-based parallel computing framework, it enables automated and rapid processing of large areas, significantly shortens analysis cycles, and supports high-frequency dynamic monitoring needs.

[0087] 4. Improve spatial resolution: Using high-resolution images can more accurately capture small-scale, fragmented cases of abandoned farmland.

[0088] The technical principles of the present invention have been described above with reference to specific embodiments. These descriptions are intended solely to illustrate the principles of the present invention and are not to be construed in any way as limiting the scope of protection of the present invention. Based on the explanations herein, those skilled in the art will readily conceive of other specific embodiments of the present invention without inventive effort, and such embodiments will fall within the scope of protection of the present invention.

Claims

1. A method for identifying abandoned farmland based on multi-source remote sensing data and time series correction, characterized by comprising the following steps: Step 1: Data acquisition and preprocessing. The specific operations are as follows: 1.

1. Remote sensing data acquisition: Select multi-source remote sensing image data covering the growing season data of the target area; 1.

2. Cloud Mask and Atmospheric Correction: When using optical remote sensing data, first remove cloud interference through a cloud mask algorithm; then perform atmospheric correction; 1.

3. Geometric registration and accuracy correction: All images are geometrically registered; in addition, the images are also radiometrically corrected; 1.

4. Radar image processing: Radar data is denoised using Refined Lee filtering, and further radiometric calibration and terrain correction are performed. The processed radar images are used to extract backscatter coefficients, which facilitates analysis of different ground object types. 1.

5. Data integration and spatiotemporal consistency: By fusing optical images with radar images; Step 2: Extract key phenological periods: Based on the NDVI and NDSI time series, two sensitive periods are extracted through filtering, smoothing, and extreme value detection algorithms: the minimum nutrient period and the peak nutrient period. The phenological period positioning error is controlled within 5 days to ensure the accuracy of the time series analysis. The specific steps are as follows: 2.

1. NDVI and NDSI time series analysis: Using NDVI and NDSI time series data, we extracted key phenological periods using an extreme value detection algorithm. NDVI reflects vegetation growth, while NDSI is highly sensitive to bare soil. 2.

2. Sensitive Period Extraction: By analyzing changes in NDVI and NDSI, the minimum nutrient phase (LVP) and the peak nutrient phase (PVP) are extracted. LVP is usually the stage with the least vegetation cover and significant bare soil, while PVP is the stage with saturated vegetation cover and can reflect the maximum vegetation growth state of cultivated land. 2.

3. Error control: When extracting phenological periods, smoothing filtering and extreme value detection algorithms are used to ensure that the phenological period positioning error is controlled within 5 days; 2.

4. Phenological period correction: In order to accurately calibrate LVP and PVP, the phenological period is corrected according to the growth cycle of different crops to ensure that the extracted phenological period adapts to changes in different regions and climatic conditions; Step 3: Feature library construction: Integrate multi-source remote sensing data and spectral index to build a multidimensional feature set, and optimize the feature dimensions through principal component analysis and feature importance screening; Step 4: Random Forest Model Training and Prediction: Using the random forest algorithm, we input a multi-source feature set and optimize model parameters through cross-validation. We also apply oversampling and class weight adjustment to address sample imbalance. The model outputs a probability map of abandoned land, which is then segmented using a threshold to generate a binary classification result, enabling regional-scale mapping of the spatial distribution of abandoned land. Step 5: Post-processing and accuracy assessment: Perform spatial filtering on the classification results to remove noise and generate a refined abandoned land distribution map; calculate the confusion matrix based on the independent test set, evaluate the overall accuracy, Kappa coefficient and category-specific indicators, analyze the main sources of misjudgment and propose the optimization results of the temporal continuity rules.

2. The method for identifying abandoned farmland based on multi-source remote sensing data and time series correction according to claim 1, wherein: In step 1.1, Landsat satellite images are used to acquire remote sensing data, and Landsat-5, Landsat-7, and Landsat-8 multi-temporal optical remote sensing image data, as well as Sentinel-2 optical image data and Sentinel-1 radar image data are acquired from Google Earth Engine.

3. The method for identifying abandoned farmland based on multi-source remote sensing data and time series correction according to claim 1, wherein: The specific operations of step 3 are as follows: 3.

1. Fusion of spectral index and remote sensing data: Combining multiple spectral indices such as NDVI, EVI, and BSI with different types of remote sensing images to construct a multidimensional feature set; 3.

2. Feature dimension optimization: Principal component analysis and feature importance screening algorithms are used to remove redundant features and retain only the features that are most effective in distinguishing different landform types; 3.

3. Feature importance screening: Use algorithms to calculate the impact of different features on the classification results and select the most representative features; 3.

4. Feature library expansion and update: As remote sensing data is continuously updated, the feature library will also be continuously expanded to ensure that it matches the latest remote sensing image data.

4. The method for identifying abandoned farmland based on multi-source remote sensing data and time series correction according to claim 1, wherein: The specific operations of step 4 are as follows: 4.

1. Model Training and Cross-Validation: We used a random forest algorithm for training, inputting a multi-source feature set, and optimized the number and depth of decision trees through cross-validation. We also adjusted the parameters of the decision trees to ensure that the model could accurately distinguish between different land cover types. 4.

2. Addressing sample imbalance: Use oversampling and class weight adjustment techniques to ensure that different categories of cultivated land and non-cultivated land are adequately represented in model training; 4.

3. Model Output and Classification Results: The trained RF model outputs a probability map of abandoned land, which is converted into a binary classification result by setting a threshold. Based on this classification result, an accurate spatial distribution map of abandoned land can be generated. 4.

4. Model Evaluation and Optimization: Evaluate the classification accuracy of the model by comparing it with an independent test set; accuracy indicators include overall accuracy, Kappa coefficient, and category accuracy. Based on these indicators, further optimize the model parameters to improve classification accuracy.

5. The method for identifying abandoned farmland based on multi-source remote sensing data and time series correction according to claim 1, wherein: The specific operations of step 5 are as follows: 5.

1. Spatial filtering and noise removal: Spatial filtering is performed on the classification results to remove noise and generate a refined abandoned land distribution map; 5.

2. Accuracy Assessment and Error Analysis: Use the confusion matrix method to evaluate the accuracy of classification results, calculate the overall accuracy, Kappa coefficient, and user and producer accuracy for each category; analyze different sources of misclassification and propose improvement suggestions and correction plans; 5.

3. Accuracy Improvement and Model Optimization: Based on the accuracy assessment results, the model will be optimized, especially the distinction between abandoned land and non-cultivated land will be further adjusted. Misclassifications will be corrected to further improve the robustness of the model. 5.

4. Final results generation: Based on the revised classification results, a wasteland monitoring map is generated, and decision-making support data is provided to help formulate reasonable land management and ecological protection strategies.

Images