Land change dynamic monitoring system and method based on multi-modal remote sensing and incremental learning

Abstract

The invention relates to a land change dynamic monitoring system and method based on multi-mode remote sensing and incremental learning, and belongs to the field of remote sensing monitoring, change detection and machine learning. The system comprises a comparative learning driven feature decoupling module, a progressive knowledge migration module, a change sample self-calibration module, a multi-granularity early warning trigger module and a difference sample backtracking and memory enhancement module, the feature decoupling module constructs an orthogonal semantic and change space and standardizes feature mapping, and the knowledge migration module realizes model updating through teacher and student networks. The self-calibration module evaluates the quality of unmarked samples and automatically incorporates the unmarked samples into a training set, the early warning trigger module realizes graded early warning and sensitive land class weighting, and the backtracking and memory enhancement module stores difficult samples and performs periodic equilibrium training. Under the condition of a small number of new samples, the change recognition accuracy of the system reaches 85%, the false alarm rate is reduced by 62%, the key change detection rate is kept above 95%, and the marking cost is greatly saved.

Description

Technical Field

[0001] The present invention relates to a land change dynamic monitoring system and method based on multi-modal remote sensing and incremental learning, belonging to the fields of remote sensing monitoring, change detection, and machine learning, and particularly to the technical field of intelligent analysis of multi-source remote sensing data using incremental learning technology to achieve land change dynamic monitoring. Background Art

[0002] Currently, land change monitoring technology mainly relies on a combination of manual field surveys and traditional remote sensing analysis. For example, the patent document with the publication number CN107516282A discloses a "land change investigation information management system integrating three inspections", which includes a server-side data overlay analysis system, a front-end information management query system, and a mobile information collection system, realizing the integration of three functional modules of inspection, investigation, and verification to improve work efficiency and data security.

[0003] However, the existing technology has the following deficiencies:

[0004] Firstly, the data acquisition method is single, mainly relying on manual field surveys to obtain data, and fails to fully utilize automated means such as remote sensing, resulting in low data acquisition efficiency and limited coverage. Secondly, the analysis ability is weak, and most systems only adopt traditional GIS overlay analysis methods, unable to achieve automatic recognition and classification of complex surface cover changes. Moreover, the monitoring mode is passive and unable to actively detect changes. It is necessary to arrange inspections and investigations manually to discover changes, with poor timeliness. In addition, the knowledge accumulation and iteration ability are insufficient. The system does not have the ability to learn from historical data, and each investigation is independent, lacking a knowledge precipitation and iteration optimization mechanism. Finally, the terminal function is limited. The mobile terminal is only used as a data collection tool, lacking edge computing ability and local intelligent analysis function, and having poor adaptability to newly emerging land type change types. Summary of the Invention

[0005] The purpose of the present invention is to provide a land change dynamic monitoring system and method based on multi-modal remote sensing and incremental learning, aiming to solve the problems of passive data acquisition, weak analysis ability, lack of knowledge accumulation, and poor adaptability existing in the prior art.

[0006] The present invention proposes a land change dynamic monitoring system based on multi-modal remote sensing and incremental learning, including:

[0007] A contrastive learning-driven feature decoupling module for constructing two orthogonal feature spaces, namely a semantic space and a change space, and mapping the original features to a normalized hypersphere through a non-linear projection head network;

[0008] Progressive knowledge transfer module, connected to the contrast learning-driven feature decoupling module, for model update based on the two-way knowledge exchange architecture of the teacher network and the student network;

[0009] Changed sample self-calibration module, connected to the progressive knowledge transfer module, for scoring unlabeled changed samples based on the pseudo-label quality assessment network and automatically incorporating high-quality samples into the training set according to the dynamic threshold;

[0010] Multi-granularity warning trigger module, connected to the changed sample self-calibration module, for hierarchical warning based on the three-level warning architecture and the sensitive land class weighted matrix;

[0011] Differential sample backtracking and memory reinforcement module, connected to the multi-granularity warning trigger module and the contrast learning-driven feature decoupling module, for storing difficult samples and periodic balanced training based on the experience replay buffer of the hierarchical storage architecture.

[0012] Preferably, the contrast learning-driven feature decoupling module includes:

[0013] Dual-space feature representation unit, for constructing two orthogonal feature spaces of the semantic space and the change space;

[0014] Nonlinear projection head network unit, for mapping the original features to the normalized hypersphere, the nonlinear projection head network unit includes a two-layer MLP structure, the input dimension is 2048, the hidden layer dimension is 512, and the output dimension is 128;

[0015] Contrast loss calculation unit, for calculating the contrast loss of the feature representation based on the InfoNCE loss function, and the temperature parameter τ of the InfoNCE loss function is set to 0.07.

[0016] Preferably, the progressive knowledge transfer module includes:

[0017] Two-way knowledge distillation network unit, for realizing knowledge exchange between the teacher network and the student network, where the number of parameters of the student network is 40% of that of the teacher network;

[0018] Soft label temperature adaptation unit, for dynamically adjusting the temperature parameter T, the initial temperature T is set to 2.0, and the adjustment range is [1.0, 4.0];

[0019] Feature alignment loss unit, for calculating the multi-level feature alignment loss L feat , and weights of 0.2, 0.3, and 0.5 are assigned to the shallow, middle, and deep features respectively.

[0020] Preferably, the changed sample self-calibration module includes:

[0021] Pseudo-label quality assessment network unit, which is used to score samples based on uncertainty assessment and consistency verification, and output sample reliability scores in the range of 0-1;

[0022] Dynamic threshold adjustment unit, which is used to dynamically adjust the threshold θ according to the characteristics of historical sample distribution. The initial threshold θ is set to 0.75, and the adjustment range is [0.65, 0.85];

[0023] Reinforcement learning loop unit, which is used to automatically incorporate high-quality pseudo-label samples into the training set and transfer low-quality samples to the manual review queue.

[0024] Preferably, the multi-granularity early warning trigger module includes:

[0025] Three-level early warning architecture unit, which is used to build three-level early warning mechanisms of observation level, attention level and alarm level based on the change probability;

[0026] Sensitive land class weighted matrix unit, which is used to build a 10×10 land class conversion sensitivity matrix W and assign higher weights to key supervised land class conversion types;

[0027] Temporal anomaly detection unit, which is used to analyze the NDVI change sequence based on the LSTM model, output the anomaly score Z, and trigger an additional verification process when Z is greater than 3σ.

[0028] Preferably, the differential sample backtracking and memory reinforcement module includes:

[0029] Experience replay buffer unit, which is used to store 10,000 samples based on a hierarchical storage architecture, divided into 5 libraries according to the change type and difficulty, and each library stores 2,000 samples;

[0030] Hard example mining unit, which is used to mark samples with the top 20% loss values and confidence levels lower than 0.6 as hard examples and assign 3 times the sampling weight;

[0031] Periodic equilibrium training unit, which is used to sample 64 historical samples from the experience replay buffer every 5 batches and mix them with the current batch for training.

[0032] Preferably, the contrast learning-driven feature decoupling module, the progressive knowledge transfer module, the changed sample self-calibration module, the multi-granularity early warning trigger module and the differential sample backtracking and memory reinforcement module are implemented through a distributed architecture, including:

[0033] Cloud training cluster, configured with 16-core GPU servers and deploying the PyTorch framework;

[0034] Edge computing unit, based on domestic Feiteng CPUs and Ascend AI chips, running TensorRT-optimized models;

[0035] The data center adopts a distributed storage system and supports PB-level remote sensing data management.

[0036] Preferably, the system provides a standardized API interface, including:

[0037] A data access interface that supports the import of remote sensing data in formats such as GeoTIFF and HDF5;

[0038] A model inference interface designed with RESTful API, supporting batch inference and real-time analysis;

[0039] An early warning push interface that supports WebSocket real-time notifications and standardized JSON alarm information.

[0040] Preferably, the bidirectional knowledge distillation network unit calculates the total loss L through the following loss function:

[0041] L = αL ce + βL kd + γL feat ,

[0042] where α = 0.5, β = 0.3, γ = 0.2, L ce is the cross-entropy loss, L kd is the knowledge distillation loss, L feat is the feature alignment loss.

[0043] A method for dynamic monitoring of land change in multi-modal remote sensing and incremental learning, including:

[0044] Construct two orthogonal feature spaces, a semantic space and a change space, and map the original features to a normalized hypersphere through a non-linear projection head network;

[0045] Realize model update based on the bidirectional knowledge exchange architecture of the teacher network and the student network;

[0046] Score the unlabeled change samples based on the pseudo-label quality assessment network and automatically include high-quality samples in the training set according to the dynamic threshold;

[0047] Realize hierarchical early warning based on the three-level early warning architecture and the sensitive land type weighting matrix;

[0048] Realize the storage of difficult samples and periodic balanced training based on the experience replay buffer with a hierarchical storage architecture.

[0049] The beneficial effects of the present invention are as follows:

[0050] Under the condition of only 20 groups of new land type samples, the change recognition accuracy rate reaches 85%, which is 25 percentage points higher than the traditional method; the false alarm rate is reduced by 62%, while maintaining a key change detection rate of over 95%; only 25 out of every 100 new change samples need to be manually reviewed, saving 75% of the annotation cost; the system response time is improved from the hour level of the traditional method to the minute level; after the model is continuously updated 10 times, the recognition accuracy rate for the initial change categories remains above 92%. The present invention realizes the paradigm shift of land change monitoring from "passive" to "active early warning", with significant technological innovation and practical value. Brief Description of the Drawings

[0051] Figure 1 It is the overall architecture diagram of the multi-modal remote sensing and incremental learning land change dynamic monitoring system in the embodiment of the present invention;

[0052] Figure 2 It is the structural schematic diagram of the contrastive learning-driven feature decoupling module in the embodiment of the present invention;

[0053] Figure 3 It is the structural schematic diagram of the progressive knowledge transfer module in the embodiment of the present invention;

[0054] Figure 4 It is the structural schematic diagram of the change sample self-calibration module in the embodiment of the present invention;

[0055] Figure 5 It is the structural schematic diagram of the multi-granularity early warning trigger module in the embodiment of the present invention;

[0056] Figure 6 It is the structural schematic diagram of the differential sample backtracking and memory reinforcement module in the embodiment of the present invention;

[0057] Figure 7 It is the performance retention curve diagram based on incremental learning in the embodiment of the present invention;

[0058] Figure 8 It is the comparison diagram of the actual application effects in the embodiment of the present invention. Detailed Embodiment

[0059] Please refer to the attached Figure 1-8 , and the present invention will be further described in detail below with reference to the attached drawings and specific embodiments.

[0060] Embodiment 1

[0061] As Figure 1 shown, the multi-modal remote sensing and incremental learning land change dynamic monitoring system includes a contrastive learning-driven feature decoupling module 1, a progressive knowledge transfer module 2, a change sample self-calibration module 3, a multi-granularity early warning trigger module 4, and a differential sample backtracking and memory reinforcement module 5.

[0062] The contrastive learning-driven feature decoupling module 1 is used to construct two orthogonal feature spaces, namely the semantic space and the variation space, and map the original features to the normalized hypersphere through the non-linear projection head network. This module effectively decouples the land change information in the feature space, realizes the rapid identification of new change patterns, and improves the adaptability of the system to new changes.

[0063] The progressive knowledge transfer module 2 is connected to the contrastive learning-driven feature decoupling module 1 and is used to update the model based on the two-way knowledge exchange architecture of the teacher network and the student network. This module solves the catastrophic forgetting problem in incremental learning, enabling the system to maintain a high recognition ability for the learned change types when incorporating new types of change samples.

[0064] The change sample self-calibration module 3 is connected to the progressive knowledge transfer module 2 and is used to score the unlabeled change samples based on the pseudo-label quality evaluation network and automatically incorporate high-quality samples into the training set according to the dynamic threshold. This module greatly reduces the dependence on manual annotation and realizes the self-optimization of the model in the real application environment.

[0065] The multi-granularity warning trigger module 4 is connected to the change sample self-calibration module 3 and is used to realize hierarchical warning based on the three-level warning architecture and the sensitive land use type weighted matrix. This module realizes the precise hierarchical supervision of land changes, ensures that important changes are responded to in a timely manner, and avoids waste of system resources.

[0066] The differential sample backtracking and memory reinforcement module 5 is connected to the multi-granularity warning trigger module 4 and the contrastive learning-driven feature decoupling module 1 and is used to store difficult samples and perform periodic balanced training based on the experience replay buffer of the hierarchical storage architecture. This module fundamentally solves the problems of class imbalance and knowledge forgetting in incremental learning, enabling the system to maintain a high recognition ability for various changes in the long term.

[0067] These five modules form a complete positive enhancement loop, jointly constructing a highly intelligent land change monitoring technology framework. The contrastive learning-driven feature decoupling module 1 provides high-quality semantic and variation feature representations for the progressive knowledge transfer module 2; the progressive knowledge transfer module 2 provides more reliable model inference capabilities, improving the quality of the pseudo-labels in the change sample self-calibration module 3; the change sample self-calibration module 3 continuously optimizes the model performance, enabling the multi-granularity warning trigger module 4 to more accurately identify changes; the high-value samples captured by the multi-granularity warning trigger module 4 are preferentially incorporated into the experience replay buffer of the differential sample backtracking and memory reinforcement module 5; and the diverse historical samples stored in the differential sample backtracking and memory reinforcement module 5 provide rich positive and negative sample pairs for the contrastive learning-driven feature decoupling module 1, further enhancing the discrimination ability of the feature representation.

[0068] Example 2

[0069] As Figure 2 shown, the contrast learning-driven feature decoupling module 1 includes a dual-space feature representation unit 11, a non-linear projection head network unit 12, and a contrast loss calculation unit 13.

[0070] The dual-space feature representation unit 11 is used to construct two orthogonal feature spaces, namely the semantic space and the change space, which are respectively used to represent the category attributes of ground objects and the temporal change characteristics of ground objects. The semantic space focuses on capturing the static features of land cover types, such as category information of buildings, farmlands, water bodies, etc.; while the change space focuses on capturing the dynamic features of ground objects changing over time, such as change patterns of farmland converted into construction land, forest land converted into farmland, etc. By decoupling these two attributes in the feature space, the system can more effectively identify and classify different types of land changes.

[0071] The non-linear projection head network unit 12 is used to map the original features onto a normalized hypersphere. This unit adopts a two-layer MLP structure, with an input dimension of 2048, a hidden layer dimension of 512, and an output dimension of 128. The main function of this unit is to map the high-dimensional features extracted from multi-modal data to a low-dimensional feature space through non-linear transformation, making the features more compact and discriminative. In actual implementation, this unit first performs a linear transformation on the input features to reduce the dimension to 512, then applies the ReLU activation function to introduce non-linearity, and finally maps the features to a 128-dimensional space through a linear layer and performs L2 normalization, so that all feature vectors are located on the unit hypersphere.

[0072] The contrast loss calculation unit 13 is used to calculate the contrast loss of the feature representation based on the InfoNCE loss function, and the temperature parameter τ of the InfoNCE loss function is set to 0.07. This unit implements the core mechanism of contrast learning. By maximizing the mutual information of positive sample pairs and minimizing the mutual information of negative sample pairs, the feature representation becomes more discriminative. The InfoNCE loss function can be expressed as:

[0073]

[0074] where s i and s j are the feature representations of positive sample pairs, s k is the feature representation of all samples, τ is the temperature parameter, set to 0.07, which is used to control the smoothness of the feature distribution, and N is the batch size.

[0075] In practical applications, the contrastive learning-driven feature decoupling module 1 effectively represents land change information through the collaborative work of the above three units. Experiments show that this module improves the change recognition accuracy under small sample conditions from 65% of the baseline to 78%. Especially for newly emerging change types, the recognition ability is significantly improved.

[0076] Embodiment 3

[0077] As Figure 3 shown, the progressive knowledge transfer module 2 includes a bidirectional knowledge distillation network unit 21, a soft label temperature adaptation unit 22, and a feature alignment loss unit 23.

[0078] The bidirectional knowledge distillation network unit 21 is used to achieve knowledge communication between the teacher network and the student network, where the number of parameters of the student network is 40% of that of the teacher network. This unit constructs an innovative bidirectional knowledge communication architecture, which is different from the way that the teacher network unidirectionally guides the student network in traditional knowledge distillation. In this unit, the teacher network represents the complete trained model, containing all the knowledge of historical data; while the student network is an incrementally updated model, focusing on learning newly added change types. Through bidirectional knowledge communication, the student network can not only obtain historical knowledge from the teacher network, but also feedback the knowledge of newly added types to the teacher network, realizing the bidirectional flow of knowledge and the co-evolution of the model.

[0079] The soft label temperature adaptation unit 22 is used to dynamically adjust the temperature parameter T. The initial temperature T is set to 2.0, and the adjustment range is [1.0, 4.0]. The core function of this unit is to automatically adjust the smoothness of the soft label according to the complexity of the sample distribution to optimize the knowledge distillation process. When facing complex or ambiguous samples, the system will automatically increase the temperature parameter T value (up to 4.0) to generate a smoother probability distribution, thus retaining more inter-class relationship information; while for simple and clear samples, the T value will be decreased (down to 1.0) to make the probability distribution more concentrated, highlighting the main class information. This dynamic adjustment mechanism enables the knowledge distillation process to adapt to change samples of different complexities, improving the accuracy and robustness of knowledge transfer.

[0080] The feature alignment loss unit 23 is used to calculate the multi-level feature alignment loss L_feat, and weights 0.2, 0.3, and 0.5 are assigned to the shallow, middle, and deep features respectively. This unit ensures the effective transfer of knowledge at different levels by aligning the feature representations of the teacher network and the student network at different abstraction levels. Shallow features mainly contain low-level visual information such as edges and textures, and are assigned a relatively low weight of 0.2; middle features contain intermediate semantic information such as local structures of ground objects, and are assigned a medium weight of 0.3; deep features contain high-level semantic information such as ground object categories and change patterns, and are assigned the highest weight of 0.5. Through this multi-level feature alignment mechanism, the student network can comprehensively inherit the knowledge structure of the teacher network, rather than just the final classification result.

[0081] The above three units work together to form a complete progressive knowledge transfer framework. The total loss function L can be expressed as:

[0082] L = αL ce + βL kd + γL feat ,

[0083] where α = 0.5, β = 0.3, γ = 0.2, L ce is the cross-entropy loss, L kd is the knowledge distillation loss, and L f eat is the feature alignment loss.

[0084] In practical applications, this module successfully solves the catastrophic forgetting problem in incremental learning, enabling the system to maintain high performance when incorporating new types of change samples. Experiments have shown that this framework increases the overall performance retention rate of the model from 74% to 91% after adding 20% new categories, significantly improving the long-term stability and adaptability of the system.

[0085] Example 4

[0086] As Figure 4 shown, the change sample self-calibration module 3 includes a pseudo-label quality assessment network unit 31, a dynamic threshold adjustment unit 32, and an enhanced learning loop unit 33.

[0087] The pseudo-label quality assessment network unit 31 is used to score samples based on uncertainty assessment and consistency verification, and output a sample reliability score within the range of 0-1. This unit adopts a quality assessment model designed with unsupervised metrics and can evaluate the quality of pseudo-labels generated by the model for unlabeled samples. The uncertainty assessment module quantifies the confidence of the model in the prediction results by calculating the entropy of the predicted probability distribution. The lower the entropy value, the higher the confidence of the model. The consistency verification module evaluates the stability of the prediction by the prediction consistency after applying different data augmentations and model parameter perturbations to the same sample. After comprehensive evaluation of the two metrics, a reliability score within the range of 0-1 is generated for each pseudo-label sample. The higher the score, the better the quality of the pseudo-label.

[0088] The dynamic threshold adjustment unit 32 is used to dynamically adjust the threshold θ according to the characteristics of the historical sample distribution. The initial threshold θ is set to 0.75, and the adjustment range is [0.65, 0.85]. This unit uses an exponential moving average algorithm to dynamically adjust the threshold for judging the quality of pseudo-labels. The system will automatically adjust the value of the threshold θ according to the performance metrics on the recent validation set and the overall quality distribution of the pseudo-label samples. In the case of good model performance and high overall quality of pseudo-labels, the system will appropriately increase the threshold (up to 0.85 at most) to ensure that only the highest-quality pseudo-labels are automatically included in the training set. While in the area with poor performance or greater processing difficulty, the system will appropriately lower the threshold (down to 0.65 at least) to ensure that there are enough samples for model optimization. This dynamic adjustment mechanism enables the system to maintain the best self-optimization effect under different application environments and data distributions.

[0089] The reinforcement learning loop unit 33 is used to automatically include high-quality pseudo-label samples in the training set and transfer low-quality samples to the manual review queue. This unit constructs a self-supervised reinforcement loop, automatically including samples with a score higher than the dynamic threshold θ by the pseudo-label quality assessment network unit 31 in the training set for model update, while samples with a score lower than the threshold enter the manual review queue for review by professionals. Samples reviewed manually will also be included in the training set and used to further optimize the pseudo-label quality assessment network. This closed-loop mechanism continuously improves the model performance and pseudo-label quality, forming a virtuous cycle and gradually reducing the dependence on manual review.

[0090] These three units together constitute the changed sample self-calibration module 3, which realizes the self-optimization of the model in the actual application environment through pseudo-label quality assessment, dynamic threshold adjustment, and reinforcement learning loop. The test results show that this module enables the monitoring system to only require 25 manual reviews for every 100 new changed samples added, saving 75% of the annotation cost, while maintaining high model performance, proving its significant advantage in reducing manual workload while maintaining high system performance.

[0091] Example 5

[0092] As Figure 5 shown, the multi - granularity warning trigger module 4 includes a three - level warning architecture unit 41, a sensitive land - type weighted matrix unit 42, and a time - series anomaly detection unit 43.

[0093] The three - level warning architecture unit 41 is used to construct a three - level warning mechanism at the observation level, attention level, and alarm level based on the change probability. This unit designs a refined three - level warning mechanism: when the change probability is greater than 0.5 but less than 0.7, the system generates an observation - level warning, adds the changed area to the observation list, and conducts regular reviews; when the change probability is greater than 0.7 but less than 0.85, the system generates an attention - level warning and arranges a low - frequency inspection plan; when the change probability is greater than 0.85, the system generates an alarm - level warning and arranges immediate verification and high - frequency monitoring. Each warning level corresponds to a different intervention response strategy, from low - frequency observation to immediately dispatching field investigators, ensuring the efficient allocation of system resources while ensuring that important changes are not overlooked.

[0094] The sensitive land - type weighted matrix unit 42 is used to construct a 10×10 land - type conversion sensitivity matrix W, and assigns higher weights to key supervised land - type conversion types. This unit constructs a 10×10 land - type conversion sensitivity matrix W according to national land management policies and supervision priorities, covering all possible conversion combinations between 10 common land - use types (such as cultivated land, forest land, grassland, construction land, etc.). Each element W ij in the matrix represents the sensitivity of the conversion from type i to type j. The larger the value, the more attention is required for this type of conversion. For example, the sensitivity weight for the conversion of cultivated land to construction land is set to 2.5 (the highest), while the sensitivity weight for the conversion of wasteland to forest land is only 0.8 (lower). The system multiplies the original change probability P by the corresponding sensitivity weight W to obtain the weighted change probability P' = P×W, further improving the detection priority of key supervised change types.

[0095] The time - series anomaly detection unit 43 is used to analyze the NDVI change sequence based on the LSTM model, output an anomaly score Z, and trigger an additional verification process when Z is greater than 3σ. This unit introduces a time - series anomaly detection algorithm based on the long short - term memory network (LSTM), which is specifically used to identify abnormal changes that do not conform to historical change patterns. The system maintains a 12 - month time - series record of NDVI (Normalized Difference Vegetation Index) for each monitoring area, and learns the normal seasonal change rules through the LSTM model. When it is detected that the NDVI change in a certain area deviates significantly from the expected pattern (anomaly score Z > 3σ, that is, exceeding 3 standard deviations), even if its change probability is not high, the system will trigger an additional verification process to prevent missed detection of possible abnormal changes. This time - series - pattern - based anomaly detection mechanism is particularly suitable for detecting those gradual changes or land changes that are confused with seasonal changes.

[0096] These three units work together to form a complete multi-granularity early warning trigger system. Compared with the traditional single-threshold trigger mechanism, this module reduces the false alarm rate by 62%, while maintaining a detection rate of over 95% for critical changes, achieving efficient allocation of regulatory resources and accurate identification of critical changes, and greatly improving the practicality and reliability of the system.

[0097] Example 6

[0098] As Figure 6 shown, the differential sample backtracking and memory enhancement module 5 includes an experience replay buffer unit 51, a hard example mining unit 52, and a periodic balance training unit 53.

[0099] The experience replay buffer unit 51 is used to store 10,000 samples based on a hierarchical storage architecture, divided into 5 libraries according to change types and difficulty levels, each storing 2,000 samples. This unit designs a buffer with a hierarchical storage architecture, with a total capacity of 10,000 historical samples, divided into 5 sub-libraries according to change types and difficulty levels, and each sub-library stores 2,000 samples. The 5 sub-libraries correspond to different types of change patterns: one for storing basic change types (such as the mutual conversion of agricultural land and construction land), one for storing seasonal change types (such as seasonal changes in water areas), one for storing progressive change types (such as vegetation degradation), one for storing rare change types (such as sudden changes caused by disasters), and one for storing hard example samples (boundary cases where the model is prone to misjudgment). This hierarchical storage structure ensures that the system can evenly retain different types of historical knowledge and avoid the over-forgetting of certain types of knowledge caused by data distribution drift.

[0100] The hard example mining unit 52 is used to mark samples with the top 20% loss values and confidence levels lower than 0.6 as hard examples and assign them a 3-fold sampling weight. This unit implements a hard example recognition algorithm based on loss values and confidence levels, which can automatically discover samples that the model has difficulty classifying correctly. In each training batch, the system records the loss value and prediction confidence level of each sample, marks samples with the top 20% loss values and confidence levels lower than 0.6 as hard examples, preferentially stores them in the hard example library of the experience replay buffer, and assigns a 3-fold sampling weight in subsequent training. This hard example mining mechanism enables the model to pay more attention to those challenging boundary cases and rare patterns, effectively improving the model's adaptability to complex change scenarios.

[0101] The periodic equilibrium training unit 53 is used to sample 64 historical samples from the experience replay buffer every 5 batches and mix them with the current batch for training. This unit designs a periodic training scheduling mechanism. During the incremental update process of the model, every time k = 5 new data batches are processed, the system will sample m = 64 historical samples from the experience replay buffer and mix them with the new samples of the current batch for training. The sampling process adopts a priority-based random sampling strategy, where difficult examples and rare class samples have a higher sampling probability, ensuring that the model can continuously review those important but easily forgotten knowledge points. This periodic equilibrium training mechanism effectively prevents the catastrophic forgetting problem that occurs in the long-term incremental learning process of the model, enabling the system to maintain the memory of historical knowledge while learning new knowledge.

[0102] These three units together constitute the differential sample backtracking and memory reinforcement module 5, which fundamentally solves the problems of class imbalance and knowledge forgetting in incremental learning. Experiments show that after introducing this module, the recognition accuracy of the system for the initial changed categories still remains above 92% after 10 consecutive updates, which is 18 percentage points higher than the baseline method, demonstrating the significant effect of this module in maintaining the long-term stability of the system.

[0103] Embodiment 7

[0104] The contrastive learning-driven feature decoupling module 1, the progressive knowledge transfer module 2, the changed sample self-calibration module 3, the multi-granularity warning trigger module 4, and the differential sample backtracking and memory reinforcement module 5 are implemented through a distributed architecture, including a cloud training cluster, edge computing units, and a data center.

[0105] The cloud training cluster is configured with 16-core GPU servers and deploys the PyTorch framework. This cluster is mainly responsible for the deep learning model training and large-scale data processing tasks of the system. Each server is equipped with 16-core high-performance GPUs, supporting large-scale parallel computing, and runs the PyTorch framework (v1.10+) for model training. The cloud cluster implements the core algorithms of the contrastive learning-driven feature decoupling module 1, the progressive knowledge transfer module 2, and the differential sample backtracking and memory reinforcement module 5, processes computationally intensive tasks, and regularly updates and distributes the optimized model parameters to the edge computing units.

[0106] The edge computing units are based on domestic Feiteng CPUs and Ascend AI chips and run TensorRT-optimized models. These edge units are deployed on monitoring points and field survey devices and are responsible for real-time data processing and preliminary analysis. Each edge unit is loaded with a lightweight deep learning model optimized by TensorRT. The model size is controlled within 200MB, which is adapted to the architectures of domestic Feiteng CPUs and Ascend AI chips and can operate efficiently in a low-power environment. The edge units mainly implement some functions of the change sample self-calibration module 3 and the multi-granularity warning trigger module 4, can process multi-source remote sensing data in real time on site, generate preliminary change analysis results, and only upload key data to the cloud, greatly reducing the data transmission volume and system response time.

[0107] The data center adopts a distributed storage system, supporting PB-level remote sensing data management. The data center is responsible for storing and managing a large amount of historical remote sensing data and change records. The system adopts a distributed storage architecture, supports PB-level data storage, and implements an efficient spatio-temporal indexing mechanism, enabling the system to quickly retrieve and access historical data in a specific area and specific time period. The data center also implements an automated data preprocessing process, including orthorectification, atmospheric correction, mosaicking, etc. of remote sensing images, to ensure the data quality of the input model.

[0108] The three are interconnected through a high-speed network to form a complete collaborative working system. The cloud training cluster is responsible for complex model training and overall optimization, the edge computing units are responsible for on-site real-time analysis and preliminary decision-making, and the data center is responsible for storing and managing a large amount of data. Through this hierarchical architecture design, the system can not only make full use of the powerful computing resources of the cloud for complex model training, but also achieve low-latency real-time response through edge computing. At the same time, relying on centralized data storage to ensure data integrity and traceability, greatly improving the practicality and efficiency of the system.

[0109] Embodiment 8

[0110] This system provides standardized API interfaces, including data access interfaces, model inference interfaces, and warning push interfaces.

[0111] The data access interface supports the import of remote sensing data in formats such as GeoTIFF and HDF5. This interface provides a unified data import specification and supports multiple common geospatial data formats, including GeoTIFF, HDF5, NetCDF, Shapefile, etc. The interface adopts the RESTful design style, provides batch upload and incremental synchronization functions, and supports resume from breakpoint and data integrity verification. The data access interface also includes a data quality assessment module, which can automatically detect and mark low-quality data (such as excessive cloud cover, sensor failures, etc.) to ensure the data quality of the input system. In addition, the interface also supports the spatio-temporal alignment of multi-source data, automatically handling the spatial resolution differences and time series inconsistencies between different remote sensing data sources.

[0112] The model inference interface adopts the RESTful API design and supports batch inference and real-time analysis. This interface provides a standardized model calling method. The client only needs to provide the geographical range and time range of the target area, and the system will automatically obtain the relevant data and perform change detection analysis, returning the change probability map and the prediction results of the change type. The interface supports two working modes: batch inference mode (processing large-scale historical data) and real-time analysis mode (processing newly acquired real-time data). To improve the system response performance, the interface implements a request queue and task priority management mechanism to ensure that important and urgent analysis tasks can be processed preferentially. The interface also provides a model interpretation function, which can generate a heat map to show the areas concerned by the model, helping users understand the change detection results.

[0113] The early warning push interface supports WebSocket real-time notifications and standardized JSON alarm information. This interface is responsible for pushing change early warning information to the user system, supports the WebSocket real-time notification mechanism to ensure that critical early warnings can be delivered to the user terminal in a timely manner. The early warning information adopts the standardized JSON format, including key information such as the early warning level, change type, geographical location, occurrence time, change probability, and relevant evidence. The interface supports multiple early warning subscription modes. Users can set subscription conditions in multiple dimensions such as geographical area, change type, and early warning level, and only receive the early warning information they care about. The interface also implements an early warning aggregation function, which can merge and push multiple similar early warnings in the same area within a short period of time to avoid information redundancy. In addition, the interface also supports an early warning confirmation and feedback mechanism, allowing users to confirm the received early warnings, mark false alarms, or provide supplementary information, and these feedbacks will be fed back to the system as important data sources to further optimize the early warning mechanism.

[0114] These three types of standardized API interfaces constitute the external interaction layer of the system, ensuring that the system can be seamlessly integrated with external systems such as existing geographic information systems and land resource management systems. The interface design follows open standards and best practices, making the system have good scalability and compatibility and being able to adapt to the needs of different application scenarios.

[0115] Example 9

[0116] In the progressive knowledge transfer module 2, the bidirectional knowledge distillation network unit 21 calculates the total loss L through the following loss function:

[0117] L = αL ce + βL kd + γL feat ,

[0118] where α = 0.5, β = 0.3, γ = 0.2, L ce is the cross-entropy loss, L kd is the knowledge distillation loss, L feat is the feature alignment loss.

[0119] The cross-entropy loss L ce is used to measure the gap between the model prediction and the true label, and is the most commonly used classification loss function in supervised learning, which can be expressed as:

[0120]

[0121] where N is the number of samples, C is the number of classes, y ij is the true label (0 or 1) of sample i belonging to class j, and p ij is the probability that the model predicts sample i belonging to class j.

[0122] The knowledge distillation loss L kd is used to measure the gap between the output of the student network and the soft label of the teacher network, and prompts the student network to learn the knowledge of the teacher network, which can be expressed as:

[0123]

[0124] where z t and z s are the output logits of the teacher network and the student network respectively, T is the temperature parameter, which is dynamically adjusted by the soft label temperature adaptation unit 22, and KL represents the Kullback-Leibler divergence, which is used to measure the difference between two probability distributions.

[0125] The feature alignment loss L feat is used to measure the difference between the intermediate layer features of the student network and the teacher network, and ensure the effective transfer of knowledge at different abstraction levels, which can be expressed as:

[0126]

[0127] where and are the feature maps of the first layer of the teacher network and the student network respectively, wl is the weight of the corresponding layer, L is the selected number of layers, and MSE is the mean square error function. In this embodiment, the weight of the shallow features is 0.2, the weight of the middle features is 0.3, and the weight of the deep features is 0.5.

[0128] Through the weighted combination of the above three loss functions, the bidirectional knowledge distillation network unit 21 can comprehensively capture and transfer the knowledge in the model, ensuring that the student network can not only learn new change types but also maintain the recognition ability of the learned change types. Specifically, when implemented, the system first uses the cross-entropy loss L ce to train the student network on new samples, and then uses the knowledge distillation loss L kd and the feature alignment loss L feat to guide the student network to learn from the teacher network. Finally, the teacher network is updated according to the performance of the student network on new samples, realizing bidirectional knowledge flow. This design enables the system to stably accumulate knowledge during the continuous learning process, avoiding the common performance oscillation and catastrophic forgetting problems in traditional incremental learning.

[0129] Embodiment 10

[0130] A method for dynamic monitoring of land change based on multi-modal remote sensing and incremental learning, comprising the following steps:

[0131] Step 1, construct two orthogonal feature spaces, namely the semantic space and the change space, and map the original features to the normalized hypersphere through a non-linear projection head network.

[0132] This step first extracts the original features from multi-source remote sensing data, including spectral features, texture features, temporal features, etc. Then, the system constructs two orthogonal feature spaces through a deep neural network: the semantic space focuses on capturing the static features of land cover types, and the change space focuses on capturing the dynamic features of ground objects changing over time. Next, the system uses a non-linear projection head network to map these original features to the normalized hypersphere to enhance the discriminability of the features. Finally, the system calculates the contrast loss through the InfoNCE loss function to optimize the feature representation, making the features of the same change type closer and the features of different change types more separated.

[0133] Step 2, implement model update based on the bidirectional knowledge exchange architecture of the teacher network and the student network.

[0134] This step first initializes the teacher network (fully trained model) and the student network (incrementally updated model). When there are new change type samples, the system first trains the student network on the new samples using cross-entropy loss so that it can recognize the new change types. Then, through the knowledge distillation and feature alignment mechanisms, the system guides the student network to learn historical knowledge from the teacher network, avoiding forgetting the change types that have been learned. Next, the system evaluates the performance of the student network on the new samples and appropriately updates the teacher network to achieve two-way knowledge flow. Finally, the system optimizes the two networks using the total loss function L (the weighted sum of cross-entropy loss, knowledge distillation loss, and feature alignment loss) to complete the model update.

[0135] Step 3: Score the unlabeled change samples based on the pseudo-label quality assessment network and automatically include the high-quality samples in the training set according to the dynamic threshold.

[0136] This step first generates pseudo-labels for the unlabeled change samples using the current model. Then, the system scores these pseudo-labels through the pseudo-label quality assessment network, and the scoring is based on two aspects: prediction uncertainty and consistency. Next, the system uses a dynamic threshold adjustment strategy to determine the current quality threshold θ, with an initial value of 0.75, which is adjusted within the range of [0.65, 0.85] according to the distribution characteristics of historical samples. Subsequently, the system automatically includes the samples with scores higher than the threshold θ in the training set, and the samples with scores lower than the threshold are transferred to the manual review queue. Finally, the system updates the model using the expanded training set and adjusts the quality assessment network according to the performance of the updated model to form a self-optimizing closed loop.

[0137] Step 4: Implement hierarchical early warning based on the three-level early warning architecture and the sensitive land cover weighted matrix.

[0138] This step first performs change detection on the monitoring area to calculate the change probability P. Then, according to the original land cover and target land cover of the changed area, the system looks up the corresponding weights from the sensitive land cover weighted matrix W and calculates the weighted change probability P' = P × W. Next, the system performs anomaly detection on the NDVI time series data to calculate the anomaly score Z. Then, based on the weighted change probability P' and the anomaly score Z, the system determines the early warning level: when P' is greater than 0.5 but less than 0.7, it is the observation level; when P' is greater than 0.7 but less than 0.85, it is the attention level; when P' is greater than 0.85 or Z > 3σ, it is the alarm level. Finally, the system triggers corresponding response measures according to different early warning levels, including regular rechecks, arranging inspections, or immediate verification, etc.

[0139] Step 5: Implement difficult sample storage and periodic balanced training based on the experience replay buffer of the hierarchical storage architecture.

[0140] This step first constructs an experience replay buffer with hierarchical storage. The total capacity is 10,000 samples, which are divided into 5 sub-banks, with 2,000 samples in each sub-bank. Then, during the training process, the system identifies difficult samples, that is, samples with a loss value ranking in the top 20% and a confidence level lower than 0.6, and preferentially stores them in the experience replay buffer. Next, every time the system processes 5 new data batches, it samples 64 historical samples from the experience replay buffer and mixes them with the current batch for training. Subsequently, the system assigns different sampling weights to different samples according to the sample difficulty and importance to ensure that difficult samples and rare category samples have a higher replay frequency. Finally, through this periodic balanced training mechanism, the system maintains the memory of historical knowledge and prevents catastrophic forgetting.

[0141] Through the collaborative implementation of the above five steps, this method realizes the intelligent monitoring and dynamic early warning of land changes, significantly improves the comprehensiveness, accuracy, and timeliness of change detection, reduces the labor cost, and provides strong technical support for the refined management of land resources.

[0142] Example 11: Application Case

[0143] This example demonstrates an actual application case to verify the effect of the present invention in an actual environment.

[0144] Application Background: A certain area needs to monitor the land use changes within a range of 20 square kilometers, with a focus on the conversion of cultivated land to construction land. The area obtains 4 periods of multi-spectral satellite images and 2 periods of UAV aerial photography data every year.

[0145] Application Process: First, the system learns 10 common land change types from historical data, including the conversion patterns between different types such as cultivated land, forest land, grassland, water body, and construction land. Then, during the monitoring process, the system identifies 3 suspected changes from cultivated land to construction land and generates 2 attention-level early warnings and 1 alarm-level early warning. For the alarm-level early warning, the system automatically triggers a UAV aerial photography task to obtain high-resolution images for verification.

[0146] Through the change sample self-calibration mechanism, the system automatically incorporates 2 confirmed change samples into the training set and triggers model update. The updated model successfully identifies 4 similar small-scale changes during subsequent monitoring, which are easily overlooked by traditional methods. At the same time, through the progressive knowledge transfer mechanism, after learning new change patterns, the system's recognition ability for the learned change types does not decline, maintaining stable performance.

[0147] Through one year of continuous monitoring, the system has detected a total of 32 land changes, including 12 cases of unauthorized conversion of cultivated land to construction land. Compared with the traditional manual survey-based method, this system discovers changes 15 days earlier on average, improving the timeliness of law enforcement. At the same time, the detection rate has increased by 27%, and the false alarm rate has decreased by 62%, greatly improving the monitoring efficiency and accuracy.

[0148] This case fully verifies the superior performance of the present invention in the actual application environment, especially the innovative advantages in small-sample learning, incremental learning, and precise early warning, demonstrating the practical value and promotion potential of this technical solution.

[0149] The present invention is implemented using a distributed architecture, including a cloud training cluster, edge computing units, and a data center, and can adapt to land change monitoring requirements of various scales. The system provides standardized API interfaces, supporting seamless integration with existing geographic information systems and land and resources management systems.

[0150] Compared with the prior art, the present invention has significant technical advantages: under the condition of only 20 groups of new land type samples, the change recognition accuracy rate reaches 85%, 25 percentage points higher than the traditional method; the false alarm rate is reduced by 62%, while maintaining a key change detection rate of more than 95%; only 25 out of every 100 new change samples need to be manually reviewed, saving 75% of the annotation cost; the system response time is improved from the hour level of the traditional method to the minute level; after the model is continuously updated 10 times, the recognition accuracy rate for the initial change categories still remains above 92%.

[0151] By realizing the paradigm shift of land change monitoring from "passive" to "active early warning", the present invention provides an efficient and precise technical means for land and resources supervision departments, with significant social and economic benefits.

[0152] The above are only the preferred embodiments of the present invention and are not used to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A land change dynamic monitoring system for multimodal remote sensing and incremental learning, characterized in that, Including: A contrastive learning-driven feature decoupling module for constructing two orthogonal feature spaces, namely a semantic space and a variation space, and mapping the original features onto a normalized hypersphere through a non-linear projection head network; A progressive knowledge transfer module connected to the contrastive learning-driven feature decoupling module for implementing model updates based on a bidirectional knowledge exchange architecture between a teacher network and a student network; A changed sample self-calibration module connected to the progressive knowledge transfer module for scoring unlabeled changed samples based on a pseudo-label quality assessment network and automatically incorporating high-quality samples into the training set according to a dynamic threshold; A multi-granularity warning trigger module connected to the changed sample self-calibration module for implementing hierarchical warnings based on a three-level warning architecture and a sensitive land class weighted matrix; A differential sample backtracking and memory reinforcement module connected to the multi-granularity warning trigger module and the contrastive learning-driven feature decoupling module for storing difficult samples and performing periodic balanced training based on an experience replay buffer with a hierarchical storage architecture.

2. The system according to claim 1, wherein The contrastive learning-driven feature decoupling module includes: A dual-space feature representation unit for constructing two orthogonal feature spaces, namely a semantic space and a variation space; A non-linear projection head network unit for mapping the original features onto a normalized hypersphere. The non-linear projection head network unit includes a two-layer MLP structure with an input dimension of 2048, a hidden layer dimension of 512, and an output dimension of 128; A contrastive loss calculation unit for calculating the contrastive loss of the feature representation based on the InfoNCE loss function, where the temperature parameter τ of the InfoNCE loss function is set to 0.

07.

3. The system according to claim 1, characterized in that, The progressive knowledge transfer module includes: A bidirectional knowledge distillation network unit for realizing knowledge exchange between a teacher network and a student network, where the number of parameters of the student network is 40% of that of the teacher network; A soft label temperature adaptive unit for dynamically adjusting the temperature parameter T, with an initial temperature T set to 2.0 and an adjustment range of [1.0, 4.0]; Feature alignment loss unit, used to calculate the multi-level feature alignment loss L feat , and weights of 0.2, 0.3, and 0.5 are assigned to the shallow, middle, and deep features respectively.

4. The system according to claim 1, wherein The changed sample self-calibration module includes: A pseudo-label quality assessment network unit for scoring samples based on uncertainty assessment and consistency verification and outputting a sample reliability score within the range of 0 - 1; A dynamic threshold adjustment unit for dynamically adjusting the threshold θ according to the historical sample distribution characteristics, with an initial threshold θ set to 0.75 and an adjustment range of [0.65, 0.85]; An enhanced learning loop unit for automatically incorporating high-quality pseudo-label samples into the training set and transferring low-quality samples to the manual review queue.

5. The system according to claim 1, characterized in that The multi-granularity warning trigger module includes: A three-level warning architecture unit for constructing a three-level warning mechanism of an observation level, a concern level, and an alarm level based on the change probability; A sensitive land class weighted matrix unit for constructing a 10×10 land class conversion sensitivity matrix W and assigning higher weights to key supervised land class conversion types; A time series anomaly detection unit for analyzing the NDVI change sequence based on an LSTM model, outputting an anomaly score Z, and triggering an additional verification process when Z is greater than 3σ.

6. The system according to claim 1, characterized in that, The differential sample backtracking and memory reinforcement module includes: An experience replay buffer unit for storing 10,000 samples based on a hierarchical storage architecture, divided into 5 libraries according to change types and difficulty levels, each storing 2,000 samples; A hard example mining unit for marking samples with the top 20% of loss values and a confidence level lower than 0.6 as hard examples and assigning a 3-fold sampling weight; A periodic balanced training unit for sampling 64 historical samples from the experience replay buffer every 5 batches and mixing them for training with the current batch.

7. The system according to claim 1, wherein The contrastive learning-driven feature decoupling module, the progressive knowledge transfer module, the changed sample self-calibration module, the multi-granularity warning trigger module, and the differential sample backtracking and memory reinforcement module are implemented through a distributed architecture, including: A cloud training cluster configured with 16-core GPU servers and deploying the PyTorch framework; An edge computing unit based on domestic Feiteng CPUs and Ascend AI chips, running a TensorRT-optimized model; A data center adopting a distributed storage system to support PB-level remote sensing data management.

8. The system according to claim 1, wherein The system provides standardized API interfaces, including: A data access interface supporting the import of remote sensing data in formats such as GeoTIFF and HDF5; A model inference interface designed with RESTful API, supporting batch inference and real-time analysis; A warning push interface supporting WebSocket real-time notifications and standardized JSON alarm information.

9. The system according to claim 3, characterized in that, The two-way knowledge distillation network unit calculates the total loss L through the following loss function: L = αL ce + βL kd + γL feat , Among them, α = 0.5, β = 0.3, γ = 0.2, L ce is the cross-entropy loss, L kd is the knowledge distillation loss, L feat is the feature alignment loss.

10. A method for dynamically monitoring land change by multi-modal remote sensing and incremental learning, using the system according to any one of claims 1-9, characterized in that, Including: Constructing two orthogonal feature spaces, a semantic space and a change space, and mapping the original features to a normalized hypersphere through a non-linear projection head network; Implementing model updates based on a two-way knowledge exchange architecture between the teacher network and the student network; Scoring unlabeled changed samples based on a pseudo-label quality assessment network and automatically incorporating high-quality samples into the training set according to a dynamic threshold; Implementing hierarchical warnings based on a three-level warning architecture and a sensitive land cover weighted matrix; Implementing the storage of difficult samples and periodic balanced training based on an experience replay buffer with a hierarchical storage architecture.

Images