Remote sensing image change detection method and device fusing lightweight-deep features and frequency domain information

Abstract

The invention discloses a remote sensing image change detection method and device fusing lightweight-deep features and frequency domain information, and the method comprises the steps: obtaining a dual-time-phase remote sensing image of a to-be-detected cultivated land, and carrying out the preprocessing of the dual-time-phase remote sensing image; inputting the preprocessed dual-time-phase remote sensing image into a cultivated land non-agrochemical detection model for processing to obtain a cultivated land non-agrochemical detection result; wherein the cultivated land non-agrochemical detection model is obtained after training based on a lightweight segmentation and depth feature collaborative model training method. The method not only shows excellent performance in farmland non-agrochemical detection tasks, but also has the comprehensive advantages of simple structure, strong generalization ability, flexible deployment and the like, and is suitable for various practical application scenes of remote sensing image change detection.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and in particular to a method and apparatus for detecting changes in remote sensing images by fusing lightweight-deep features and frequency domain information. Background Technology

[0002] Change detection (CD) in remote sensing images is a core task in intelligent remote sensing interpretation, widely used in fields such as farmland use monitoring, urban expansion analysis, and disaster damage assessment. Its goal is to accurately identify areas of change in ground features within remote sensing images from different time phases. Current mainstream methods primarily employ convolutional neural networks (CNNs) or visual transformer architectures for feature extraction and classification.

[0003] CNN architecture excels at modeling local spatial context information, effectively capturing edge and texture features in images and performing well in high-resolution remote sensing images. However, limited by its finite receptive field and layer-by-layer stacking characteristics, traditional CNNs struggle to model long-range dependencies, resulting in insufficient ability to model cross-scale semantic relationships. In contrast, the Transformer architecture, through its self-attention mechanism, can model global context and is suitable for handling remote sensing images with complex semantic relationships. However, Transformers typically require large-scale labeled data, their generalization ability is highly dependent on downstream tasks, and they are prone to overfitting in spatially imbalanced scenarios of remote sensing images.

[0004] In recent years, Vision Foundation Models (VFMs) such as Segment Anything Model (SAM) have demonstrated strong generalization capabilities in natural image segmentation tasks. Their superior performance across multiple downstream tasks is attributed to the general visual feature representations obtained through large-scale pre-training. However, directly transferring VFMs to remote sensing image change detection scenarios still faces numerous challenges. On the one hand, existing methods heavily rely on large-scale, precisely labeled pixel-level change tags, limiting their application in large-scale, low-label environments. On the other hand, remote sensing images commonly contain non-semantic interference factors such as changes in illumination conditions, seasonal alternation, and land cover evolution. Existing models lack effective semantic-change decoupling mechanisms, easily misclassifying pseudo-changes as real changes, leading to high false detection rates. Furthermore, since VFMs are mostly trained on natural images, the features they learn are difficult to fully adapt to the complex geometric structures and radiometric characteristics of remote sensing images, resulting in distorted feature representations, blurred edges, and semantic structural breaks. Meanwhile, ground feature changes exhibit significant scale diversity in remote sensing images. A single-branch network structure struggles to integrate local details with global semantic information, resulting in limited model recognition capabilities for small targets and complex background changes. More importantly, dual-temporal remote sensing images are often affected by factors such as imaging angle, time span, and differences in sensor parameters during acquisition, leading to significant style inconsistencies that interfere with accurate modeling and identification of changed areas. In summary, the direct application of VFMs in remote sensing change detection tasks is still constrained by multiple factors, necessitating structural and mechanistic optimization and adaptation tailored to the characteristics of remote sensing data.

[0005] Based on the above background analysis, current remote sensing image change detection methods still have significant shortcomings in several key technical aspects. First, many existing methods rely primarily on a single backbone network structure (such as U-Net, ResNet, or Siamese architecture) for feature extraction, making it difficult to simultaneously capture spatial details and deep semantic information. This is particularly problematic in scenarios with small object changes or complex backgrounds, easily leading to missed detections and false detections. Second, while introducing multi-scale features can enhance the model's spatial perception capability to some extent, most methods employ relatively simple fusion strategies, lacking effective cross-scale collaborative mechanisms and failing to fully integrate structural information and semantic representations from features at different levels. Furthermore, most current models are still limited to feature processing in the spatial domain, neglecting the structural stability and saliency of changes inherent in the frequency domain of remote sensing images. This makes style transfer difficult to achieve, further affecting the accurate modeling of changed regions. Simultaneously, in terms of salient region modeling, traditional attention mechanisms tend to focus on single-dimensional modeling of channels or space, lacking multi-dimensional joint modeling and dynamic adjustment capabilities, limiting the model's response to key regions. Finally, due to the slight displacement of ground features or geometric deviations that may exist between two temporal phases in remote sensing images, existing methods struggle to achieve accurate feature alignment, especially in areas with blurred edges or small-scale changes. This often results in response misalignment or boundary breaks, severely impacting the accuracy and robustness of change detection. In conclusion, designing a remote sensing change detection model with multi-dimensional collaborative modeling, frequency domain supplementary representation, and high-precision alignment capabilities remains a key challenge in current research. Summary of the Invention

[0006] To address the technical problems existing in the prior art, this invention proposes a remote sensing image change detection method and apparatus that integrates lightweight-deep features and frequency domain information. This overcomes the shortcomings of the prior art and effectively solves the problems of inaccurate extraction of semantic information from farmland in complex scenarios, leading to missed detections and false detections. Furthermore, considering that existing supervised learning methods heavily rely on large amounts of manually labeled data, making it difficult to meet the needs of global-scale farmland change monitoring, the method proposed in this invention reduces the labeling burden while improving the robustness and generalization ability of the detection.

[0007] To achieve the above objectives, this invention provides a remote sensing image change detection method that integrates lightweight-deep features and frequency domain information, comprising:

[0008] Acquire dual-temporal remote sensing images of the cultivated land to be detected and perform preprocessing;

[0009] The preprocessed dual-temporal remote sensing images are input into the farmland non-agriculturalization detection model for processing to obtain the farmland non-agriculturalization detection results;

[0010] Among them, the farmland non-agriculturalization detection model was obtained after training based on a model training method that combines lightweight segmentation and deep feature collaboration.

[0011] Preferably, the process of training the farmland non-agriculturalization detection model includes:

[0012] A historical dual-temporal remote sensing image sample set is acquired, and the historical dual-temporal remote sensing image sample set is uniformly resized and enhanced. The sample set includes several samples, each of which includes a historical dual-temporal remote sensing image of cultivated land and the corresponding identification information of cultivated land non-agriculturalization detection results.

[0013] The processed samples are divided into training set, validation set and test set according to a preset ratio. Each sample contains an image pair and its corresponding change label.

[0014] The training set is used to train a preset deep learning model to obtain a trained deep learning model.

[0015] The trained deep learning model is validated and tested using the validation set and the test set, and the deep learning model is optimized to obtain the farmland non-agriculturalization detection model.

[0016] Preferably, the preset deep learning model includes:

[0017] Dual-branch feature extraction module: This module is used to input the acquired dual-temporal remote sensing images into the lightweight segmentation backbone Fast-SAM branch and the deep convolutional neural network ResNet50 branch with strong semantic representation capabilities, respectively, to extract multi-scale feature maps.

[0018] Channel dimensionality reduction module: used to unify the number of channels for the features extracted by the Fast-SAM branch;

[0019] Multi-scale attention feature aggregation module: used to fuse channel attention and spatial attention mechanisms, extract local context information from feature maps through multi-scale deep convolutional structures and concatenate them, and then fuse the concatenated results to obtain feature maps of different scales;

[0020] Gated multi-scale context aggregation module: It is used to concatenate multi-scale features from Fast-SAM branch and ResNet50 branch in the channel dimension, and model the context semantic distribution through global average pooling path of multiple scales to generate semantic attention map. It also introduces learnable gating unit to perform channel-level and spatial-level weighted modulation on the semantic attention map.

[0021] Frequency domain channel-selective feature exchange module: used to exchange the center spectrum of selected channels in the frequency domain;

[0022] Decision-level fusion classification module: used to classify and fuse feature maps of different scales for output.

[0023] Preferably, the final output feature of the multi-scale attention feature aggregation module is obtained by adding the weighted sum of the enhanced features to the original input:

[0024]

[0025] In the formula, Out represents the final output feature of the multi-scale attention feature aggregation module, and W represents the output feature of the multi-scale attention feature aggregation module. s These are learnable weight parameters. and All are multi-scale feature representations enhanced through attention mechanisms. and These are the original input features for the corresponding branches.

[0026] Preferably, the processing procedure of the frequency domain channel selective feature exchange module includes:

[0027] Perform a two-dimensional fast Fourier transform on the selected operation channel and then perform spectrum centering.

[0028] Exchange the spectral information of the dual-temporal images in the central region of the frequency domain;

[0029] The image is restored by inverse Fourier transform, where only the selected channels are frequency domain swapped, while the other channels remain unchanged;

[0030] Wherein, the frequency domain center region Ω is defined as:

[0031] Ω={(u,v)|u0-l≤u <u0+l,v0-l≤v<v0+l};

[0032]

[0033] In the formula, l represents the controllable center radius, u0 and v0 are the frequency domain center coordinates after spectrum centering, u and v are the horizontal and vertical coordinates of the pixel in the frequency domain, and H and W are the height and width of the input feature map, respectively.

[0034] Preferably, the decision-level fusion classification module adopts a weighted voting mechanism to fuse classification results at different scales, and balances the loss weights of each scale output through an adaptive loss function;

[0035] The method for balancing the loss weights of the output at each scale using an adaptive loss function is as follows:

[0036]

[0037] loss = cross_entropy();

[0038] L=a*loss(result1)+b*loss(result2)+c*loss(result4)+d*loss(result);

[0039] Among them, c ′ y is the number of categories; i p represents the true label of class i; i is the probability of the i-th class predicted by the model; a, b, c, and d are adaptive parameters; cross_entropy() is the cross-entropy function; result1, result2, result4, and result are the transformed results at the first, second, and third scales and after final fusion, respectively; L is the final adaptive weighted total loss; loss(·) represents the cross-entropy loss value calculated for the output at the specified scale. entropy This represents the cross-entropy loss function.

[0040] The present invention also provides a model training device for the remote sensing image change detection method that integrates lightweight-deep features and frequency domain information, comprising:

[0041] The sample acquisition unit is used to acquire a sample set and divide the sample set into a training set, a validation set and a test set according to a preset ratio. The sample set includes several samples, and each sample includes a historical dual-temporal remote sensing image of cultivated land and the corresponding identification information of the non-agriculturalization detection result of cultivated land.

[0042] The model training unit is used to train a preset deep learning model using the training set to obtain a trained deep learning model. The preset deep learning model includes a dual-branch feature extraction module, a channel dimensionality reduction module, a multi-scale attention feature fusion module, a gated multi-scale context aggregation module, a frequency domain channel selective feature exchange module, and a decision-level fusion classification module.

[0043] The model optimization unit is used to validate and test the trained deep learning model using the validation set and the test set, optimize the deep learning model, and obtain a farmland non-agriculturalization detection model.

[0044] The model training unit includes:

[0045] The multi-scale feature extraction and fusion module is used to input the acquired dual-temporal remote sensing images into the lightweight segmentation backbone Fast-SAM branch and the deep convolutional neural network ResNet50 branch with strong semantic representation capabilities to extract multi-scale feature maps.

[0046] The number of channels for the features extracted by the Fast-SAM branch is unified using a channel dimensionality reduction module.

[0047] By utilizing a multi-scale attention feature fusion module, channel attention and spatial attention mechanisms are fused. Local contextual information in the feature maps is extracted and concatenated through a multi-scale deep convolutional structure. The concatenated results are then fused to obtain feature maps at different scales.

[0048] Using the decision-level fusion classification module, the non-agriculturalization of cultivated land is detected on the feature maps at each scale after fusion, and the detection results of each non-agriculturalization of cultivated land are fused to obtain the final non-agriculturalization detection result of cultivated land.

[0049] By optimizing and balancing the loss between the farmland non-agriculturalization detection results at various scales and the fused farmland non-agriculturalization detection results in the training process using an adaptive loss function, the training weights that meet the training conditions are output, resulting in the trained deep learning model.

[0050] The present invention also provides a remote sensing image change detection device for the remote sensing image change detection method that integrates lightweight-deep features and frequency domain information, comprising:

[0051] The image acquisition module is used to acquire dual-temporal remote sensing images of the cultivated land to be detected and to perform preprocessing.

[0052] The detection module is used to input the preprocessed dual-temporal remote sensing images into the farmland non-agriculturalization detection model for processing, and obtain the farmland non-agriculturalization detection results.

[0053] The farmland non-agriculturalization detection model is obtained through a model training device.

[0054] An electronic device includes a processor and a memory, the memory storing a computer program that is loaded and executed by the processor to implement the steps in the method.

[0055] A storage medium storing a computer program that can be read by a computer, the computer program being configured to execute the steps of the method at runtime.

[0056] Compared with the prior art, the present invention has the following advantages and technical effects:

[0057] This invention constructs a dual-branch, multi-scale feature extraction architecture that integrates the lightweight segmentation model FastSAM and the deep semantic network ResNet50. By combining modules such as multi-scale attention feature aggregation, gated multi-scale context fusion, frequency domain channel-selective interaction, and decision-level classification fusion, it achieves accurate detection of farmland non-agriculturalization change areas in complex remote sensing scenarios. This structure balances shallow spatial details with deep semantic information, effectively enhancing the model's sensitivity to changes in areas with blurred boundaries, small targets, and complex backgrounds. Simultaneously, the introduction of a multi-scale attention mechanism strengthens the response capability to key change areas and suppresses interference from non-target areas. Furthermore, this invention utilizes a frequency domain channel-selective interaction mechanism to locally exchange spectral features from remote sensing images of different time phases, achieving style consistency while maintaining spatial structure stability and effectively mitigating spurious change problems caused by imaging conditions and sensor differences. The gated multi-scale context aggregation module can dynamically adjust feature weights at different scales and between branches, thereby improving the representation effect of salient areas and enhancing global context modeling capabilities. Through a dual-axis cross-attention mechanism, the model further enhances its ability to perceive spatial structure changes, helping to refine change boundaries and enhance the discrimination ability of geometrically heterogeneous regions. By combining decision-level fusion classification with an adaptive loss function optimization strategy, the model can integrate information at different scales and achieve stable convergence, improving overall detection accuracy and robustness. This invention not only demonstrates excellent performance in detecting farmland conversion to non-agricultural uses, but also possesses comprehensive advantages such as simple structure, strong generalization ability, and flexible deployment, making it suitable for various practical application scenarios in remote sensing image change detection. Attached Figure Description

[0058] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0059] Figure 1 This is a schematic diagram of the implementation environment of an embodiment of the present invention;

[0060] Figure 2 This is a schematic diagram illustrating the process of training a farmland non-agriculturalization detection model according to an embodiment of the present invention.

[0061] Figure 3 This is a schematic diagram of a model training method based on lightweight segmentation and deep feature collaboration according to an embodiment of the present invention;

[0062] Figure 4 This is a schematic diagram of the dual-branch feature extraction module according to an embodiment of the present invention;

[0063] Figure 5 This is a schematic diagram of the channel dimensionality reduction module according to an embodiment of the present invention;

[0064] Figure 6This is a schematic diagram of the multi-scale attention feature aggregation module in an embodiment of the present invention;

[0065] Figure 7 This is a schematic diagram of the gated multi-scale context aggregation module according to an embodiment of the present invention;

[0066] Figure 8 This is a schematic diagram of the frequency domain channel selective feature exchange module according to an embodiment of the present invention;

[0067] Figure 9 This is a flowchart of the remote sensing image change detection method that integrates lightweight-deep features and frequency domain information according to an embodiment of the present invention.

[0068] Figure 10 This is a schematic diagram of a remote sensing image according to an embodiment of the present invention. Detailed Implementation

[0069] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0070] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0071] This embodiment provides a remote sensing image change detection method, a remote sensing image change detection device, a model training device, an electronic device, and a storage medium that integrate lightweight-deep features and frequency domain information. A preset deep learning model is trained using a training set to obtain a trained deep learning model. The trained deep learning model is then validated and tested using a validation set and a test set to optimize the deep learning model and obtain a farmland non-agriculturalization detection model. This farmland non-agriculturalization detection model is then used to detect dual-temporal remote sensing images of the farmland to be detected, yielding farmland non-agriculturalization detection results.

[0072] The method provided in this embodiment may involve artificial intelligence (AI) technology and can be implemented based on artificial intelligence technology, such as using deep learning to train a corresponding model using samples.

[0073] Machine Learning (ML) is a multidisciplinary field involving probability theory, statistics, approximation theory, convex analysis, and algorithm complexity theory. It specifically studies how computers can simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to endow computers with intelligence; its applications span all areas of artificial intelligence.

[0074] Deep learning (DL) specifically refers to machine learning based on deep neural network models and methods. It has developed based on statistical machine learning, artificial neural network, and other algorithmic models, combined with the development of modern big data and high computing power. The most important technical feature of deep learning is its ability to automatically extract features.

[0075] The aforementioned machine learning and deep learning typically include techniques such as neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and instructional learning.

[0076] like Figure 1 This is a schematic diagram of the implementation environment. The implementation environment may include training equipment and usage equipment.

[0077] Both the training equipment and the equipment used are computer devices; optionally, the computer device is a terminal device, such as a mobile phone, tablet computer, PC (Personal Computer) or other electronic devices; or, the computer device is a server, which can be a single server, a server cluster composed of multiple servers, or a cloud computing service center. This embodiment of the invention does not limit this.

[0078] Training equipment refers to computer equipment capable of training and learning deep learning models. Optionally, the training equipment has the ability to acquire deep learning models and train and learn them according to application requirements. For example, the training equipment acquires deep learning models from other devices via a network and then trains them using training samples according to application requirements, so that the deep learning model has the ability to obtain a model for detecting non-agriculturalization of arable land. Optionally, the training equipment has the ability to build neural networks. It can build deep learning models on its own according to application requirements and then train and learn them. For example, in order to obtain the detection results of non-agriculturalization of arable land from dual-temporal remote sensing images, the training equipment builds a deep learning model on its own and then trains and learns it using samples according to application requirements.

[0079] The device used refers to a computer device that has the requirement to use a deep learning model. Optionally, the device uses the data to obtain a farmland non-agriculturalization detection model from other devices via the network according to the application requirements. For example, if the device has the requirement to detect farmland non-agriculturalization, it can obtain a farmland non-agriculturalization detection model that has been trained and learned from other devices via the network, and use the farmland non-agriculturalization detection model to perform farmland non-agriculturalization detection.

[0080] Based on this, the technical solution of the present invention will be described and explained below with reference to several examples.

[0081] Example 1

[0082] This embodiment proposes a remote sensing image change detection method that integrates lightweight-deep features and frequency domain information, such as... Figure 9 ,include:

[0083] Acquire dual-temporal remote sensing images of the cultivated land to be detected and perform preprocessing;

[0084] The preprocessed dual-temporal remote sensing images are input into the farmland non-agriculturalization detection model for processing to obtain the farmland non-agriculturalization detection results;

[0085] Among them, the farmland non-agriculturalization detection model was obtained after training based on a model training method that combines lightweight segmentation and deep feature collaboration.

[0086] Furthermore, the process of training the model for detecting the non-agricultural use of arable land includes:

[0087] A historical dual-temporal remote sensing image sample set is acquired, and the historical dual-temporal remote sensing image sample set is uniformly resized and enhanced. The sample set includes several samples, each of which includes a historical dual-temporal remote sensing image of cultivated land and the corresponding identification information of cultivated land non-agriculturalization detection results.

[0088] The processed samples are divided into training set, validation set and test set according to a preset ratio. Each sample contains an image pair and its corresponding change label.

[0089] The training set is used to train a preset deep learning model to obtain a trained deep learning model.

[0090] The trained deep learning model is validated and tested using the validation set and the test set, and the deep learning model is optimized to obtain the farmland non-agriculturalization detection model.

[0091] Specifically, the process of training the model for detecting the non-agricultural use of arable land includes, for example: Figure 2 :

[0092] Step S110: Obtain a sample set and divide it into a training set, a validation set and a test set according to the proportion. The sample set includes several samples, each of which includes historical dual-temporal remote sensing images of cultivated land and corresponding identification information of cultivated land non-agriculturalization detection results.

[0093] Step S120: Train the preset deep learning model using the training set to obtain the trained deep learning model.

[0094] Step S130: Validate and test the trained deep learning model using the validation set and test set, optimize the deep learning model, and obtain the farmland non-agriculturalization detection model.

[0095] In this embodiment, each sample in step S110 includes a historical dual-temporal remote sensing image of cultivated land and corresponding identification information of cultivated land non-agriculturalization detection results. The identification information of cultivated land non-agriculturalization detection results is an image that identifies the size and location of cultivated land non-agriculturalization in the historical dual-temporal remote sensing image. This image can be a feature map or it can be obtained by converting a feature map.

[0096] Furthermore, the preset deep learning model includes:

[0097] Dual-branch feature extraction module: This module is used to input the acquired dual-temporal remote sensing images into the lightweight segmentation backbone Fast-SAM branch and the deep convolutional neural network ResNet50 branch with strong semantic representation capabilities, respectively, to extract multi-scale feature maps.

[0098] Channel dimensionality reduction module: used to unify the number of channels for the features extracted by the Fast-SAM branch;

[0099] Multi-scale attention feature aggregation module: used to fuse channel attention and spatial attention mechanisms, extract local context information from feature maps through multi-scale deep convolutional structures and concatenate them, and then fuse the concatenated results to obtain feature maps of different scales;

[0100] Gated multi-scale context aggregation module: It is used to concatenate multi-scale features from Fast-SAM branch and ResNet50 branch in the channel dimension, and model the context semantic distribution through global average pooling path of multiple scales to generate semantic attention map. It also introduces learnable gating unit to perform channel-level and spatial-level weighted modulation on the semantic attention map.

[0101] Frequency domain channel-selective feature exchange module: used to exchange the center spectrum of selected channels in the frequency domain;

[0102] Decision-level fusion classification module: used to classify and fuse feature maps of different scales for output.

[0103] Specifically, in this embodiment, the deep learning model preset in step S120 includes a dual-branch feature extraction module, a channel dimensionality reduction module, a multi-scale attention feature aggregation module, a gated multi-scale context aggregation module, a frequency domain channel selective feature exchange module, and a decision-level fusion classification module.

[0104] The dual-branch feature extraction module can use the pre-trained dual branches (Fast-SAM and ResNet50) to extract multi-scale features from the samples in the training set. In order to balance local detail sensitivity and global semantic consistency, the network introduces an adapter module to effectively integrate the two types of features, thereby enhancing the feature expression ability and the discriminability of change areas. This is beneficial for locating farmland semantic information in complex scenes, thereby identifying farmland changes in small areas and complex scenes.

[0105] The multi-scale attention feature aggregation module can compensate for the loss of details in the multi-scale feature extraction process through cross-scale fusion;

[0106] The gated multi-scale context aggregation module effectively integrates the shallow spatial detail features extracted by the FastSAM branch and the deep semantic features extracted by the ResNet50 branch. This module aims to integrate feature representations from two heterogeneous backbone networks across branches and scales. By introducing a multi-scale context modeling mechanism, it fully captures spatial semantic information under different receptive fields. At the same time, combined with a gated modulation mechanism, it achieves dynamic enhancement of salient regions and effective suppression of redundant background information.

[0107] In the frequency domain channel-selective feature exchange module, to further enhance the model's ability to model fine-grained structural variations across time phases, flexible spectral migration range constraints are achieved through a channel-controllable strategy while maintaining the stability of the image spatial structure. Local frequency domain transformation enhances the modeling ability for mesoscale structural differences. At the same time, since the operation is limited to a reversible process at the center of the frequency domain, the introduction of edge artifacts can be effectively avoided, ensuring the consistency and discriminability of the original image structure and providing stable support for subsequent feature representation.

[0108] Furthermore, step S130 specifically includes:

[0109] The trained deep learning model is validated using a validation set, and the model parameters are optimized and updated by combining the validation results with the optimizer. The optimizer here can be the Adam optimizer, which can optimize and update the model parameters to accelerate the convergence process and improve optimization efficiency. A cosine annealing learning rate scheduler is used to dynamically adjust the learning rate during training to avoid skipping the optimal solution or getting stuck in local optima when approaching the optimal solution.

[0110] The validated deep learning model is tested using a test set, and a qualified farmland change detection model is output.

[0111] A model training method based on lightweight segmentation and deep feature collaboration is proposed. This method combines a bi-branch feature extraction module, a channel dimensionality reduction module, a multi-scale attention feature aggregation module, a gated multi-scale context aggregation module, a frequency domain channel selective feature exchange module, and a decision-level fusion classification module. It extracts multi-scale features from dual-temporal remote sensing images and performs interactive fusion, global information mining, highlighting change-related features, and suppressing change-irrelevant features on the features at each scale. This enables the model to not only integrate features related to cultivated land at different scales, but also accurately locate the semantic information of cultivated land in complex scenarios and accurately identify changes in cultivated land for non-agricultural purposes in complex scenarios. This solves the problem that existing cultivated land change detection methods are not accurate enough in detecting the semantic information of cultivated land in complex scenarios, which easily leads to missed detections and false detections.

[0112] Specifically, such as Figure 3 ,include:

[0113] Step S210: Acquire several samples and adjust the size of all samples according to the set pixels. Each sample includes a historical dual-temporal remote sensing image of cultivated land and the corresponding identification information of cultivated land non-agriculturalization detection results. In this embodiment, the set pixels can be set as needed, and can be, but is not limited to, 256*256 pixels or 512*512 pixels.

[0114] Step S220: Enhance the samples to obtain a sample set, wherein the enhancement methods include random rotation, horizontal flipping, and vertical flipping.

[0115] This step involves data augmentation using a random number p, including: random rotation to increase the diversity of the dataset and improve the model's robustness to image rotation changes; horizontal flipping to increase the diversity of the dataset and improve the model's robustness to horizontal changes in the image; and vertical flipping to increase the diversity of the dataset and improve the model's robustness to vertical changes in the image.

[0116] Step S230: Divide the sample set into a training set, a validation set, and a test set according to the specified proportions. The proportions used for dividing the sample set in this step can be set as needed.

[0117] Furthermore, the pre-defined deep learning model is trained using the training set to obtain the trained deep learning model, including:

[0118] Step S310: The dual-branch feature extraction module is used to extract and fuse features at multiple scales from the samples in the training set, resulting in feature maps at different scales, such as... Figure 4 .

[0119] To address the challenge of co-modeling spatial detail information and high-level semantic features in farmland change detection, a novel dual-branch co-coding network architecture is proposed. This network consists of a lightweight segmentation backbone (Fast-SAM) branch and a deep convolutional neural network (ResNet50) branch with strong semantic representation capabilities, constructing a two-stream feature extraction framework that balances local detail sensitivity with global semantic consistency. Specifically, the input dual-temporal remote sensing images... The data are input into both the Fast-SAM and ResNet50 branches to extract features at different scales. To achieve cross-branch and cross-layer feature fusion, an adapter module is introduced to effectively integrate the two types of features, thereby enhancing feature representation and the discriminative power of changing regions. This design effectively improves the model's detection accuracy and robustness in complex changing regions, especially demonstrating excellent performance in real-world scenarios such as blurred plot boundaries and heterogeneous background interference.

[0120] Step S320 involves channel dimensionality reduction of the multi-scale features extracted by the Fast-SAM branch using the channel dimensionality reduction module, such as... Figure 5 ;

[0121] A multi-scale channel dimensionality reduction module (Adapter) is used to perform channel dimensionality reduction on the multi-scale features extracted by the Fast-SAM branch in order to effectively integrate the two types of features, thereby enhancing the feature representation ability and the discriminative power of the change region;

[0122] Adapter16=nn.Conv2d(640,chan_num,kernel_size=1,stride=1,padding=0,bias=False);

[0123] self.Adapter8=nn.Conv2d(320,chan_num,kernel_size=1,stride=1,padding=0,bias=False);

[0124] self.Adapter4=nn.Conv2d(160, chan_num, kernel_size=1, stride=1, padding=0, bias=False).

[0125] Step S330 utilizes a multi-scale attention feature aggregation module, which integrates channel attention and spatial attention mechanisms, and extracts local contextual information through a multi-scale deep convolutional structure. The concatenated results are then used for feature fusion to obtain feature maps at different scales.

[0126] To enhance the expressive power of dual-temporal features and the ability to discriminate changing regions, this embodiment designs a multi-scale attention feature aggregation module (Multi-scale Effective Feature Aggregation, MAFA). For example... Figure 6 This module integrates channel attention and spatial attention mechanisms, and extracts local contextual information through multi-scale deep convolutional structures, thereby achieving significant enhancement of key change regions and effective suppression of background noise. The proposed Multi-Scale Effective Feature Aggregation (MEFA) module aims to achieve robust feature representation through a dual-domain attention mechanism and adaptive multi-scale fusion. This module consists of three key components: dual-branch feature decoupling, dynamic feature calibration, and residual feature optimization.

[0127] Specifically, the input dual-temporal feature map and First, channel importance and spatial saliency are modeled using concatenated channel attention (CA) and spatial attention (SA) to generate weighted features:

[0128]

[0129] Subsequently, the two enhanced feature maps are concatenated along the channel dimension and then compressed in dimension using a 1×1 convolution to obtain the fused feature F:

[0130]

[0131] The fused feature F is equally divided into four sub-features. Three of these sub-features are input to depthwise separable convolutional channels with different kernel sizes (3×3, 5×5, and 7×7) to extract local structural information under different receptive fields. The remaining sub-feature serves as a skip connection to preserve the original information. Finally, the four sub-features are concatenated along the channel dimension and fused using a 1×1 convolution to obtain the fused feature F. ms :

[0132] F ms =Conv 1×1 ([F3,F5,F7,F s ]);

[0133] To guide the subsequent feature fusion process, the MAFA module uses F... ms Applying the Sigmoid activation function generates weights W s Used to dynamically adjust the fusion weights of bi-temporal features:

[0134] W s =σ(F ms );

[0135] The final output feature is obtained by adding the weighted sum of the enhanced features to the original input, in the following form:

[0136]

[0137] In the formula, Out represents the final output feature of the multi-scale attention feature aggregation module, and W represents the output feature of the multi-scale attention feature aggregation module. s These are learnable weight parameters used to adjust the fusion ratio of enhanced features. and All are multi-scale feature representations enhanced through attention mechanisms. and These are the original input features for the corresponding branches, used to preserve low-level information and enhance gradient flow through skip connections.

[0138] The Multi-Scale Attention Feature Aggregation (MAFA) module, by introducing multi-scale convolutional branches and attention mechanisms, not only enhances the richness of feature representation but also significantly improves the model's ability to perceive small regions and fine-grained changes, making it suitable for change detection tasks in high-resolution remote sensing images.

[0139] Step S340: The shallow spatial detail features extracted by the FastSAM branch and the deep semantic features extracted by the ResNet50 branch are effectively fused using the gated multi-scale context aggregation module.

[0140] The gated multi-scale context aggregation module first concatenates multi-scale features from FastSAM and ResNet50 along the channel dimension, and then models the contextual semantic distribution through global average pooling paths at multiple scales to generate a semantic attention map. Subsequently, a learnable gating unit is introduced to perform channel-level and spatial-level weighted modulation of the fused features, guiding the model to focus on change-sensitive regions and suppressing responses to static or interfering regions. Furthermore, the module employs a residual connection structure to preserve key information from the original features, which helps stabilize training and improve semantic alignment performance.

[0141] The specific steps are as follows:

[0142] To fully integrate the deep semantic information extracted from the backbone network with the shallow details in skip connections, this embodiment proposes a Gated Multi-scale Context Aggregation Module (GMCA), such as... Figure 7 .

[0143] This module enhances the responsiveness to salient regions and the suppression of irrelevant backgrounds during feature fusion by combining multi-scale context modeling with gating mechanisms.

[0144] Given shallow features and deep features First, the data is stitched together along the channel dimension to obtain fused features. To effectively capture contextual information under different receptive fields, AMSA-FFM designed three adaptive average pooling paths of different scales (1×1, 2×2, and 4×4), denoted as follows:

[0145]

[0146] Each attention path has its output size unified through an upsampling operation, and the results are weighted and fused to form the overall attention map.

[0147] A = A1 + A2 + A3;

[0148] To further enhance the model's response to significantly changing regions, a gating mechanism is introduced. This mechanism modulates feature channels or spatial locations through learning-based gating, thereby dynamically adjusting the weight distribution of fused features and enhancing the model's ability to perceive key regions.

[0149] G = Sigmoid(Conv) 1×1 (X m ));

[0150] A gate =A·G;

[0151] X m ′=X m ·A gate ;

[0152] Considering that the fused feature dimension is 2C, the module uses 1×1 convolution to compress the channels to the original dimension C, while introducing residual paths to retain key information of the main branches:

[0153]

[0154] X f =Conv1×1 (X m ′)+X r ;

[0155] To further refine the fusion results, an additional attention conditioning term is introduced:

[0156]

[0157] The GMCA module constructs an efficient and robust feature fusion framework by integrating multi-scale contextual modeling, gating mechanisms, and residual connections. This module employs multi-scale average pooling to extract rich global contextual information and introduces a gating mechanism to dynamically adjust attention intensity, enhancing the expressive power of salient regions. Simultaneously, residual structures mitigate gradient vanishing and preserve the original feature distribution. Furthermore, a dual-branch attention alignment mechanism further improves the matching between shallow and deep features. Overall, GMCA effectively promotes the collaborative fusion of low-level details and high-level semantics, significantly enhancing the model's ability to perceive edge contours and fine-grained changes.

[0158] To further enhance the model's ability to model fine-grained structural variations across time phases, this embodiment designs a frequency-domain channel-selective feature interchange (FCFI) module, such as... Figure 8 This module, while maintaining the stability of the image spatial structure, performs local region information exchange in the frequency domain to achieve low-cost, high-fidelity feature style reconstruction.

[0159] Specifically, let the source image and the target image be respectively First, select a set of operation channel indices based on the controllable parameters channel_option and step:

[0160]

[0161] Subsequently, in the selected channel Perform a two-dimensional Fast Fourier Transform (FFT) and then perform spectral centering:

[0162]

[0163] To achieve structurally consistent style transfer, only the complex spectral information of the central region Ω is exchanged in the frequency domain, where Ω is defined as:

[0164] Ω={(u,v)|u0-l≤u <u0+l,v0-l≤v<v0+l};

[0165]

[0166] Where l represents the controllable center radius, u0 and v0 are the frequency domain center coordinates after spectrum centering, u and v are the x and y coordinates of the pixel in the frequency domain, and H and W are the height and width of the input feature map, respectively.

[0167] Next, the following frequency domain swap will be performed:

[0168]

[0169] After frequency domain swapping, the image is reconstructed using inverse FFT:

[0170]

[0171] Final output Maintain the same shape and data type as the input, only for channels. The region underwent frequency domain replacement:

[0172]

[0173] This module has significant advantages: it achieves flexible spectral migration range constraints through a channel-controllable strategy; local frequency domain transformation enhances the ability to model mesoscale structural differences; and since the operation is limited to a reversible process at the center of the frequency domain, it can effectively avoid the introduction of edge artifacts, ensure the consistency and discriminability of the original image structure, and provide stable support for subsequent feature representation.

[0174] Step S350: The decision-level fusion classification module is used to detect the non-agriculturalization of cultivated land in the feature maps of each scale after fusion, and the detection results of each non-agriculturalization of cultivated land are fused to obtain the final detection result of non-agriculturalization of cultivated land.

[0175] In this embodiment, the decision-level fusion classification module adopts a weighted voting method to dynamically adjust the decision weights of the output results at each scale, ensuring that the final classification result takes into account both shallow and deep semantic features. The specific steps are as follows:

[0176] (1) First, input the feature maps at each scale into the corresponding classifiers to obtain the detection results of farmland non-agriculturalization at three different scales. The specific steps are as follows:

[0177]

[0178] x1=interpolate(x1,size=imgsize);

[0179] x2=interpolate(x2,size=imgsize);

[0180] x4=interpolate(x4,size=imgsize);

[0181] result1 = Classifier1(x1);

[0182] result2 = Classifier2(x2);

[0183] result4 = Classifier4(x4);

[0184] The specific process of the classifier is as follows:

[0185] result=Conv3(Relu(BN(Conv3(x1))));

[0186] In the formula, interpolate is the upsampling function, Conv3 is the 3*3 convolution operation, BN is the batch normalization, ReLU is the activation function, and Classifier1, Classifier2, and Classifier4 are three classifiers, one for each scale.

[0187] (2) The three different scales of farmland non-agriculturalization detection results are fused to obtain the final farmland non-agriculturalization detection results. The specific steps are as follows:

[0188] result=torch.cat([result1,result2,result4],dim=1);

[0189] result = Conv1(result);

[0190] Wherein, Conv1: 1*1 convolution.

[0191] Step S360 involves optimizing and balancing the loss between the farmland non-agriculturalization detection results at various scales during training and the fused farmland non-agriculturalization detection results using an adaptive loss function. This ensures the accuracy and consistency of the detection results. The specific steps are as follows:

[0192] The proportions of the four losses are adjusted using adaptive parameters, and the calculation process is as follows: The loss function can be replaced with other loss functions.

[0193]

[0194] loss = cross_entropy();

[0195] L=a*loss(result1)+b*loss(result2)+c*loss(result4)+d*loss(result);

[0196] Among them, c ′ y is the number of categories; i p represents the true label of class i; i `i` represents the probability of the i-th class predicted by the model; `a`, `b`, `c`, and `d` are adaptive parameters; `cross_entropy()` is the cross-entropy function; `result1`, `result2`, `result4`, and `result` are the transformed results at the first, second, and third scales, respectively, and the final fused result; `L` is the final adaptive weighted total loss, used to optimize model parameters during training; `loss(·)` represents the cross-entropy loss value calculated for the output at the specified scale. entropy This represents the cross-entropy loss function.

[0197] In the steps of this embodiment, a new loss function is used for results at different scales, and the loss weights between different scales and different categories are adjusted by adaptive parameters. This makes full use of features at different scales and focuses on semantic features of interest in complex backgrounds, thus alleviating the problem of false detection and false negative detection caused by complex backgrounds.

[0198] Step S370: Repeat the above steps to output the training weights that meet the training conditions, thus obtaining the trained deep learning model. The training conditions include the number of iterations and the stability of the loss value.

[0199] Specifically, such as Figure 10 Six sets of historical dual-temporal remote sensing images were acquired. The first row contains remote sensing images of temporal phase one, the second row contains remote sensing images of temporal phase two, the third row contains the identification information of the farmland non-agriculturalization detection results for each set of historical dual-temporal remote sensing images, and the fourth row contains the farmland non-agriculturalization detection results detected using the farmland non-agriculturalization detection model trained in this embodiment. Black represents unchanged areas, white represents changed areas, green represents false detection areas, and red represents missed detection areas. Comparison shows that the missed detection and false reduction areas are within the allowable range. The model trained in this embodiment is suitable for farmland non-agriculturalization detection.

[0200] This embodiment combines a dual-branch feature extraction module, a multi-scale attention feature aggregation module, and a gated multi-scale context aggregation module to effectively address the problems of inaccurate extraction of semantic information from farmland in complex scenarios, leading to missed detections and false detections. Furthermore, considering that existing supervised learning methods heavily rely on large amounts of manually labeled data, making it difficult to meet the needs of global-scale farmland change monitoring, the method proposed in this embodiment reduces the labeling burden while improving the robustness and generalization ability of the detection. Specifically:

[0201] A dual-stream feature extraction framework is adopted, which uses pre-trained Fast-SAM and ResNet50 to extract multi-scale features from dual-temporal remote sensing images. The two branches are independent of each other while interacting with each other, making full use of the advantages of different branches and different scale features, so as to locate the semantic features of cultivated land in complex backgrounds.

[0202] A channel dimensionality reduction module (Adapter) is used to effectively integrate the two types of features, thereby enhancing the feature representation ability and the discriminative power of changing regions;

[0203] A multi-scale attention feature aggregation module is employed to enhance the expressive power of dual-temporal features and the ability to discriminate changing regions. This module integrates channel attention and spatial attention mechanisms and extracts local contextual information through multi-scale deep convolutional structures, thereby achieving significant enhancement of key changing regions and effective suppression of background noise. The proposed multi-scale effective feature aggregation (MEFA) module aims to achieve robust feature representation through dual-domain attention mechanisms and adaptive multi-scale fusion.

[0204] A gated multi-scale context aggregation module is employed to transform spatial domain features into the frequency domain. Then, a method similar to spatial attention and channel attention is used to interact with the features from both temporal phases. This strategy minimizes style differences between the two temporal remote sensing images, achieving style unification and suppressing non-semantic variations and missed detections caused by style inconsistencies. To effectively fuse shallow spatial detail features extracted by the FastSAM branch with deep semantic features extracted by the ResNet50 branch, this module aims to integrate feature representations from two heterogeneous backbone networks across branches and scales. By introducing a multi-scale context modeling mechanism, it fully captures spatial semantic information under different receptive fields. Simultaneously, combined with a gated modulation mechanism, it achieves dynamic enhancement of salient regions and effective suppression of redundant background information. Specifically, the GMCA module first concatenates multi-scale features from FastSAM and ResNet50 along the channel dimension and models the contextual semantic distribution through global average pooling paths at multiple scales to generate a semantic attention map. Subsequently, a learnable gating unit is introduced to perform channel-level and spatial-level weighted modulation of the fused features, guiding the model to focus on change-sensitive regions and suppressing responses to static or interfering regions. Furthermore, the module employs a residual connection structure to preserve key information from the original features, which helps stabilize training and improve semantic alignment. The GMCA module balances shallow structural details with high-level semantic abstraction during feature fusion, significantly improving the model's discriminative ability and robustness in complex scenes, and providing richer and more consistent feature representations for subsequent change detection.

[0205] A frequency-domain channel-selective feature exchange module is employed to further enhance the model's ability to model fine-grained structural variations across time phases. While maintaining the stability of the image's spatial structure, the module performs local region information exchange in the frequency domain, achieving low-cost, high-fidelity feature style reconstruction. A channel-controllable strategy enables flexible spectral migration range constraints; local frequency-domain transformation enhances the modeling ability for mesoscale structural differences; and because the operation is confined to a reversible process at the center of the frequency domain, it effectively avoids the introduction of edge artifacts, ensuring the consistency and discriminability of the original image structure and providing stable support for subsequent feature representation.

[0206] Example 2

[0207] This embodiment discloses a model training device for a remote sensing image change detection method that integrates lightweight-deep features and frequency domain information, comprising:

[0208] The sample acquisition unit is used to acquire a sample set and divide the sample set into a training set, a validation set and a test set according to a preset ratio. The sample set includes several samples, and each sample includes a historical dual-temporal remote sensing image of cultivated land and the corresponding identification information of the non-agriculturalization detection result of cultivated land.

[0209] The model training unit is used to train a preset deep learning model using the training set to obtain a trained deep learning model. The preset deep learning model includes a dual-branch feature extraction module, a channel dimensionality reduction module, a multi-scale attention feature fusion module, a gated multi-scale context aggregation module, a frequency domain channel selective feature exchange module, and a decision-level fusion classification module.

[0210] The model optimization unit is used to validate and test the trained deep learning model using the validation set and the test set, optimize the deep learning model, and obtain a farmland non-agriculturalization detection model.

[0211] The model training unit includes:

[0212] The multi-scale feature extraction and fusion module is used to input the acquired dual-temporal remote sensing images into the lightweight segmentation backbone Fast-SAM branch and the deep convolutional neural network ResNet50 branch with strong semantic representation capabilities to extract multi-scale feature maps.

[0213] The number of channels for the features extracted by the Fast-SAM branch is unified using a channel dimensionality reduction module.

[0214] By utilizing a multi-scale attention feature fusion module, channel attention and spatial attention mechanisms are fused. Local contextual information in the feature maps is extracted and concatenated through a multi-scale deep convolutional structure. The concatenated results are then fused to obtain feature maps at different scales.

[0215] Using the decision-level fusion classification module, the non-agriculturalization of cultivated land is detected on the feature maps at each scale after fusion, and the detection results of each non-agriculturalization of cultivated land are fused to obtain the final non-agriculturalization detection result of cultivated land.

[0216] By optimizing and balancing the loss between the farmland non-agriculturalization detection results at various scales and the fused farmland non-agriculturalization detection results in the training process using an adaptive loss function, the training weights that meet the training conditions are output, resulting in the trained deep learning model.

[0217] Example 3

[0218] This embodiment also provides a remote sensing image change detection device applied to the aforementioned remote sensing image change detection method that fuses lightweight-deep features and frequency domain information, comprising:

[0219] The image acquisition module is used to acquire dual-temporal remote sensing images of the cultivated land to be detected and to perform preprocessing.

[0220] The detection module is used to input the preprocessed dual-temporal remote sensing images into the farmland non-agriculturalization detection model for processing, and obtain the farmland non-agriculturalization detection results.

[0221] The farmland non-agriculturalization detection model is obtained through a model training device.

[0222] Example 4

[0223] This embodiment also provides an electronic device, including a processor and a memory, wherein the memory stores a computer program, which is loaded and executed by the processor to implement the steps in the remote sensing image change detection method that fuses lightweight-deep features and frequency domain information.

[0224] Specifically, the processor can be a central processing unit (CPU), a general-purpose processor, a digital signal processor (DSP), an ASIC, an FPGA, or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It can implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of these embodiments. It can also be a combination that implements computational functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, etc. The memory can include, but is not limited to, various media capable of storing computer programs, such as a USB flash drive, a read-only memory, a portable hard drive, a magnetic disk, or an optical disk.

[0225] Example 5

[0226] A storage medium storing a computer program readable by a computer, the computer program being configured to execute steps in a remote sensing image change detection method that fuses lightweight-deep features and frequency domain information during runtime.

[0227] Specifically, the aforementioned storage media may include, but are not limited to, USB flash drives, read-only memory, portable hard drives, magnetic disks, optical disks, and other media capable of storing computer programs.

[0228] Those skilled in the art will understand that this embodiment can be provided as a method, system, or computer program product. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of this application can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0229] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A remote sensing image change detection method that integrates lightweight-deep features and frequency domain information, characterized in that, include: Acquire dual-temporal remote sensing images of the cultivated land to be detected and perform preprocessing; The preprocessed dual-temporal remote sensing images are input into the farmland non-agriculturalization detection model for processing to obtain the farmland non-agriculturalization detection results; Among them, the farmland non-agriculturalization detection model was obtained after training based on a model training method that combines lightweight segmentation and deep feature collaboration.

2. The remote sensing image change detection method according to claim 1, characterized in that, The process of training the farmland non-agriculturalization detection model includes: A historical dual-temporal remote sensing image sample set is acquired, and the historical dual-temporal remote sensing image sample set is uniformly resized and enhanced. The sample set includes several samples, each of which includes a historical dual-temporal remote sensing image of cultivated land and the corresponding identification information of cultivated land non-agriculturalization detection results. The processed samples are divided into training set, validation set and test set according to a preset ratio. Each sample contains an image pair and its corresponding change label. The training set is used to train a preset deep learning model to obtain a trained deep learning model. The trained deep learning model is validated and tested using the validation set and the test set, and the deep learning model is optimized to obtain the farmland non-agriculturalization detection model.

3. The remote sensing image change detection method according to claim 2, characterized in that, The preset deep learning model includes: Dual-branch feature extraction module: This module is used to input the acquired dual-temporal remote sensing images into the lightweight segmentation backbone Fast-SAM branch and the deep convolutional neural network ResNet50 branch with strong semantic representation capabilities, respectively, to extract multi-scale feature maps. Channel dimensionality reduction module: used to unify the number of channels for the features extracted by the Fast-SAM branch; Multi-scale attention feature aggregation module: used to fuse channel attention and spatial attention mechanisms, extract local context information from feature maps through multi-scale deep convolutional structures and concatenate them, and then fuse the concatenated results to obtain feature maps of different scales; Gated multi-scale context aggregation module: It is used to concatenate multi-scale features from Fast-SAM branch and ResNet50 branch in the channel dimension, and model the context semantic distribution through global average pooling path of multiple scales to generate semantic attention map. It also introduces learnable gating unit to perform channel-level and spatial-level weighted modulation on the semantic attention map. Frequency domain channel-selective feature exchange module: used to exchange the center spectrum of selected channels in the frequency domain; Decision-level fusion classification module: used to classify and fuse feature maps of different scales for output.

4. The remote sensing image change detection method according to claim 3, characterized in that, The final output feature of the multi-scale attention feature aggregation module is obtained by adding the weighted sum of the enhanced features to the original input: In the formula, Out represents the final output feature of the multi-scale attention feature aggregation module, and W represents the output feature of the multi-scale attention feature aggregation module. s These are learnable weight parameters. and All are multi-scale feature representations enhanced through attention mechanisms. and These are the original input features for the corresponding branches.

5. The remote sensing image change detection method according to claim 3, characterized in that, The processing procedure of the frequency domain channel selective feature exchange module includes: Perform a two-dimensional fast Fourier transform on the selected operation channel and then perform spectrum centering. Exchange the spectral information of the dual-temporal images in the central region of the frequency domain; The image is restored by inverse Fourier transform, where only the selected channels are frequency domain swapped, while the other channels remain unchanged; Wherein, the frequency domain center region Ω is defined as: Ω={(u,v)|u0-l≤u <u0+l,v0-l≤v<v0+l}; In the formula, l represents the controllable center radius, u0 and v0 are the frequency domain center coordinates after spectrum centering, u and v are the horizontal and vertical coordinates of the pixel in the frequency domain, and H and W are the height and width of the input feature map, respectively.

6. The remote sensing image change detection method according to claim 3, characterized in that, The decision-level fusion classification module adopts a weighted voting mechanism to fuse classification results at different scales and balances the loss weights of each scale output through an adaptive loss function. The method for balancing the loss weights of the output at each scale using an adaptive loss function is as follows: loss = cross_entropy(); L=a*loss(result1)+b*loss(result2)+c*loss(result4)+d*loss(result); Among them, c ′ y is the number of categories; i p represents the true label of class i; i is the probability of the i-th class predicted by the model; a, b, c, and d are adaptive parameters; cross_entropy() is the cross-entropy function; result1, result2, result4, and result are the transformed results at the first, second, and third scales and after final fusion, respectively; L is the final adaptive weighted total loss; loss(·) represents the cross-entropy loss value calculated for the output at the specified scale. entropy This represents the cross-entropy loss function.

7. A model training device for the remote sensing image change detection method fusing lightweight-deep features and frequency domain information as described in any one of claims 1-6, characterized in that, include: The sample acquisition unit is used to acquire a sample set and divide the sample set into a training set, a validation set and a test set according to a preset ratio. The sample set includes several samples, and each sample includes a historical dual-temporal remote sensing image of cultivated land and the corresponding identification information of the non-agriculturalization detection result of cultivated land. The model training unit is used to train a preset deep learning model using the training set to obtain a trained deep learning model. The preset deep learning model includes a dual-branch feature extraction module, a channel dimensionality reduction module, a multi-scale attention feature fusion module, a gated multi-scale context aggregation module, a frequency domain channel selective feature exchange module, and a decision-level fusion classification module. The model optimization unit is used to validate and test the trained deep learning model using the validation set and the test set, optimize the deep learning model, and obtain a farmland non-agriculturalization detection model. The model training unit includes: The multi-scale feature extraction and fusion module is used to input the acquired dual-temporal remote sensing images into the lightweight segmentation backbone Fast-SAM branch and the deep convolutional neural network ResNet50 branch with strong semantic representation capabilities to extract multi-scale feature maps. The number of channels for the features extracted by the Fast-SAM branch is unified using a channel dimensionality reduction module. By utilizing a multi-scale attention feature fusion module, channel attention and spatial attention mechanisms are fused. Local contextual information in the feature maps is extracted and concatenated through a multi-scale deep convolutional structure. The concatenated results are then fused to obtain feature maps at different scales. Using the decision-level fusion classification module, the non-agriculturalization of cultivated land is detected on the feature maps at each scale after fusion, and the detection results of each non-agriculturalization of cultivated land are fused to obtain the final non-agriculturalization detection result of cultivated land. By optimizing and balancing the loss between the farmland non-agriculturalization detection results at various scales and the fused farmland non-agriculturalization detection results in the training process using an adaptive loss function, the training weights that meet the training conditions are output, resulting in the trained deep learning model.

8. A remote sensing image change detection device applied to the remote sensing image change detection method fusing lightweight-deep features and frequency domain information as described in any one of claims 1-6, characterized in that, include: The image acquisition module is used to acquire dual-temporal remote sensing images of the cultivated land to be detected and to perform preprocessing. The detection module is used to input the preprocessed dual-temporal remote sensing images into the farmland non-agriculturalization detection model for processing, and obtain the farmland non-agriculturalization detection results. The farmland non-agriculturalization detection model is obtained by training the model training device as described in claim 7.

9. An electronic device, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program, which is loaded and executed by the processor to implement the steps of the method as described in any one of claims 1-6.

10. A storage medium, characterized in that, The storage medium stores a computer program that can be read by a computer, the computer program being configured to execute the steps of the method as described in any one of claims 1-6 when it is run.

Images