Large-scale forest felling drawing method based on computer vision model and Sentinel-1 time sequence SAR (Synthetic Aperture Radar) data

Abstract

The invention discloses a large-scale forest felling drawing method based on a computer vision model and Sentinel-1 time sequence SAR data, relates to the technical field of forest felling detection and forest felling drawing, and solves the problems that an existing forest felling drawing model based on the Sentinel-1 time sequence SAR data is low in precision and lacks testing of cross-regional transferability. According to the method, acquired image data are trained by establishing a forest cutting cartographic model SAMSR, and a final model outputs a forest cutting probability graph. According to the SAMSR model, the accuracy of forest felling mapping is improved, and ecological system management is supported by providing detailed opinions to forest long-term dynamics.

Description

Technical Field

[0001] The present invention relates to the technical field of forest felling detection and forest felling mapping, and in particular to a large-scale forest felling mapping method based on a computer vision model and Sentinel-1 time series SAR data. Background Art

[0002] Deep learning models, such as U-Net, can automatically learn complex features from Sentinel-1 data. It is able to capture complete and solid logging patches of different shapes and orientations, and outperforms traditional methods in logging mapping (such as existing literature: Monthly Forest Deforestation Mapping Using Dense Time Series Sentinel-1 SAR Images and Deep Learning, 2022, Environmental Remote Sensing). However, the U-Net model still faces challenges when applied to Sentinel-1 data for forest deforestation mapping. Sentinel-1 SAR images are prone to speckle noise, which makes it difficult to extract clear and reliable features required for accurate analysis, and brings difficulties to U-Net when dealing with the specific multi-scale complexity of forest environments. In addition, in terms of context perception, the model's reliance on local domains may limit its ability to capture broader contextual information in the vast, heterogeneous forest landscapes of different locations.

[0003] With the advancement of computer vision technology, base image segmentation models customized for remote sensing data have emerged, opening up new avenues for improving large-scale deforestation mapping. Recently, Meta AI introduced the Segment Anything Model (SAM), a powerful computer vision base model trained on natural images that can accurately segment complex scenes. The emergence of SAM provides new insights for further improving deforestation mapping based on remote sensing data (such as existing literature: Segment Everything, 2023, Proceedings of the IEEE / CVF International Conference on Computer Vision). However, the original SAM model was trained on natural images, and the model structure lacks a module for handling SAR noise. As a result, SAM is difficult to generalize from natural images to SAR data, especially in identifying complex spatial patterns of deforestation, limiting its effectiveness in large-scale mapping. Therefore, there is an urgent need to develop specialized model architectures to effectively exploit the advantages of Sentinel-1 SAR imagery while addressing its inherent challenges. Recent studies have shown that adding adapters to the basic SAM structure and increasing CNN branches can suppress noise of small objects in ultrasound images and enhance the extraction of fuzzy boundaries. However, the role of this approach in deforestation monitoring and mapping based on Sentinel-1 data still needs further evaluation.

[0004] Deep learning models require a large amount of high-quality annotated data for training, which is costly and time-consuming to obtain. Second, the heterogeneity of terrain, species composition, and harvesting patterns makes it difficult to transfer models across regions. Ideally, the combination of transfer and active learning techniques with adaptive SAM can save computing resources while maximizing the use of limited labeled data, providing a promising solution to address the efficiency and cost challenges of large-scale logging mapping. On the one hand, transfer learning allows models to apply pre-trained knowledge to new tasks, reducing the need for labeled data and enhancing cross-regional adaptability. Fine-tuning on local data is one of the most commonly used transfer learning techniques, which can reduce the need for large amounts of labeled data and enhance the portability of models across different geographic regions and forest types. On the other hand, active learning addresses these issues by optimizing the training process. Specifically, it can select the most representative samples, avoid redundancy, and minimize costs while optimizing model performance.

[0005] Therefore, in this paper, a new framework, namely the framework based on Sentinel-1SAR data and SAM (SAMSR), is proposed to map large-scale deforestation in four pan-tropical hotspots using time-series Sentinel-1 data. Our specific objectives include: (1) Evaluate the performance of SAMSR using the original SAM model and the U-Net model as benchmarks. (2) Determine the optimal transfer strategy of the framework based on transfer learning methods in four large test sites with different environmental conditions (Rondonia, Brazil, Guangxi Zhuang Autonomous Region, China, California, USA, and Hainan, China). (3) Evaluate the effectiveness of active learning methods in reducing the number of samples required for model training and improving the efficiency of SAMSR. Summary of the invention

[0006] In order to solve the problems of low accuracy of the existing forest felling mapping model based on Sentinel-1 time series SAR data and lack of test on cross-regional transferability, the present invention provides a large-scale forest felling mapping method based on computer vision model and Sentinel-1 time series SAR data.

[0007] A large-scale forest clearing mapping method based on computer vision model and Sentinel-1 time series SAR data consists of the following steps:

[0008] Step 1: Collect high-resolution Planet image data and Sentinel-1SAR image data of felled samples;

[0009] Step 2: preprocess the image data collected in step 1 and divide it into a training set, a validation set and a test set;

[0010] Step 3: Establish a forest felling mapping model SAMSR, use the training set to train the SAMSR model, and obtain a trained SAMSR model;

[0011] The SAMSR model includes a ViT branch layer, a CNN branch layer, a CBA layer, a hint encoder and a mask decoder;

[0012] The training set is input into the ViT branch layer and the CNN branch layer at the same time; the ViT branch layer is used to extract the global features of the image, and the CNN layer is used to extract the local features of the image, and the global features and the local features are input into the CBA layer for feature weighted fusion to extract the feature map;

[0013] Inputting the feature map into a mask decoder to generate an image segmentation mask;

[0014] The real image segmentation map of the prompt encoder is input into the mask decoder and compared with the generated image segmentation mask to finally generate a forest felling probability map;

[0015] Step 4: Use the test set to test the trained SAMSR model, and use the SAMSR model and Sentinel-1 data to perform monthly forest felling mapping.

[0016] Beneficial effects of the present invention:

[0017] 1. In the mapping method described in the present invention, the dynamic change characteristics of forest felling are extracted by combining Sentinel-1SAR data with the SAM model, which solves the problems of insufficient monitoring accuracy and poor timeliness in existing methods. Using Sentinel-1SAR data, not only can the forest felling situation be monitored in a timely manner, but also stable remote sensing data support can be provided under various regional conditions. This method can greatly improve the logging monitoring capabilities in large-scale forest areas, effectively address the problem of illegal logging, and provide technical support for global environmental governance. In addition, this method can also overcome the obstacles of traditional remote sensing data such as complex terrain and clouds, and further expand the application boundaries of remote sensing technology. The leading nature of this method has brought disruptive changes to forest resource management and promoted the innovative development of the industry.

[0018] 2. The forest felling mapping method based on time series SAR data of the present invention can monitor large-scale forest felling under various environmental conditions, solves the pain point of weather influence of traditional optical remote sensing, fills the market gap, and meets the urgent need of global forest resource monitoring. By providing accurate forest felling monitoring data, it can help governments and international organizations reduce environmental and economic losses caused by illegal felling and improve resource management efficiency. Environmental protection organizations, carbon emission management agencies, forestry companies, etc. will optimize their resource management and protection programs by adopting this technology, thereby saving costs and improving benefits. In addition, since the data of the Sentinel-1 satellite is free and open to the public, this greatly reduces the cost of data acquisition and increases the market promotion potential of this technology.

[0019] 3. In the mapping method described in the present invention, a SAMSR model for mapping large-scale deforestation using Sentinel-1 imagery and an adaptive SAM model was proposed and evaluated. The SAMSR model performed well on four pan-tropical study areas, and its results consistently outperformed the baseline original SAM and U-Net models, with IoU improvements of 28.30% and 6.25%, respectively. The optimal migration strategy across regions was further determined, and the potential of active learning to improve model efficiency was explored. The SAMSR model found that the multi-region combination model had the best migration performance, and active learning further improved the efficiency, achieving saturation accuracy with approximately 20%-70% fewer training samples.

[0020] 4. The SAMSR model described in the present invention improves the accuracy of forest harvest mapping and supports ecosystem management by providing detailed insights into the long-term dynamics of forests. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 A flowchart of the large-scale forest felling mapping method based on computer vision model and Sentinel-1 time series SAR data according to the present invention;

[0022] Figure 2 It is the principle diagram of the SAMSR model in the method of the present invention;

[0023] Figure 3 The figure shows the spatial representation of the reference and prediction results of the SAMSR model, SAM and U-Net model described in the present invention. The first row is the high-resolution Planet image, the second and third rows are the VV and VH of Sentinel-1. The fourth row is the manually drawn labels, and the fifth, sixth and seventh rows are the prediction results of the U-Net, SAM and SAMSR models respectively.

[0024] Figure 4The IoU values ​​of the random method and the active method based on the multi-region comprehensive model under different samples of an embodiment of the present invention are shown in Figure 2. The study areas are (a) Rondonia (Ron), Brazil, (b) Guangxi Zhuang Autonomous Region (GX), China, (c) California (CA), the United States, and (d) Hainan (HN), China.

[0025] Figure 5 Renderings of long-term monthly deforestation mapping for four study regions; the three blue circled areas are shown in magnified view. Adjacent to these magnified images are Planet images before and after deforestation, and the bar charts below are monthly deforestation statistics for the four study regions: (a) Rondônia (Ron), Brazil, (b) Guangxi Zhuang Autonomous Region (GX), China, (c) California (CA), United States, and (d) Hainan (HN), China. DETAILED DESCRIPTION

[0026] Specific implementation method 1. Combination Figures 1 to 5 This embodiment describes a large-scale forest felling mapping method based on a computer vision model and Sentinel-1 time series SAR data. The method can establish a forest felling mapping model SAMSR based on Sentinel-1 data of the forest and combined with an improved SAM model, that is, a felling mapping model based on the SAM model and SAR (SAM-based and SAR, SAMSR), which can quickly identify and map felling in the forest. The method is implemented by the following steps:

[0027] Step 1: Collect high-resolution Planet image data and Sentinel-1SAR image data for felled samples in different study areas;

[0028] The Sentinel-1SAR image data includes vertical-vertical polarization (VV) and vertical-horizontal polarization (VH) acquisition data; the Sentinel-1SAR image data can be time series images of the study area at different time points to ensure the integrity of seasonal change information in the forest felling area.

[0029] Step 2: preprocess the collected high-resolution Planet image data and Sentinel-1SAR image data, namely: preprocess the image data into a consistent size to ensure consistency, and use the high-resolution Planet image data as annotation data (mask), i.e., forest felling label; remove noise, radiation calibration, terrain correction and data cropping from the Sentinel-1SAR image data; and reasonably divide the ratios of the training set, validation set and test set.

[0030] Step 3: Based on the improved SAM model, a forest felling mapping model SAMSR is constructed. This model inherits the overall architecture of the SAM model, retains its prompt encoder and mask decoder, and is optimized for the special needs of remote sensing image segmentation.

[0031] like Figure 2 As shown, the SAMSR model includes a ViT branch layer (Vision Transformer), a CNN branch layer, a cross-branch attention layer (CBA layer), a prompt encoder and a mask decoder; the ViT branch layer includes an adapter, a window transformer and a global transformer;

[0032] The adapter is used to dynamically adjust the global position embedding to adapt to the environmental characteristics of different study areas and enhance the model's migration ability. By tuning the input Sentinel-1 features to maximize alignment with the pre-trained visual transformer encoder, the model's feature extraction ability is improved. The SAMSR model adopts a cross-branch attention mechanism and integrates the advantages of CNN and ViT branches. It can better capture local and global features in the detection of forest logging areas and improve the accuracy of logging area recognition. Specifically:

[0033] The input image first enters the ViT branch layer and the CNN branch layer at the same time. The image is embedded and resized through the adapter in the ViT branch layer. Then, the feature map is divided into multiple local windows by the window transformer. Each window is processed independently, and finally enters the multi-head self-attention mechanism of the global transformer for global feature extraction.

[0034] The input image is passed through multiple convolutional pooling blocks in the CNN branch layer to extract fine-grained local information. Then the features of the ViT branch layer and the CNN branch layer enter the CBA layer to extract the feature map by weighted fusion of the two branch features. The feature map is then passed to the mask decoder to generate the final image segmentation mask. Then the encoder is prompted that the input real image segmentation map can be compared with the generated image segmentation mask to finally generate a forest felling probability map; each pixel value in the felling probability map is the felling probability, and pixels with a felling probability greater than 0.5 are felled. The SAMSR model is trained for 200 epochs. After each epoch, the SAMSR model calculates the intersection over union (IoU) on the validation set. If the current IoU is better than the previous best IoU, the current SAMSR model will be saved as the best model. If the performance on the validation set no longer improves after multiple epochs, the early stopping strategy can be used to terminate the training.

[0035] In this implementation, the image data input to the ViT branch layer and the CNN branch layer of the SAMSR model are both two-band Sentinel-1SAR images (VV and VH) and corresponding annotation data (mask), and the size of each input image is 128×128 pixels; the ViT branch layer is responsible for global feature extraction. In order to make the ViT branch layer better adapt to the low contrast and blurred boundaries in remote sensing images, the model introduces an adapter. The adapter aligns the input features and adjusts the position embedding through maximum pooling and convolution to adapt to the resolution change of the input image.

[0036] The SAMSR model introduces a parallel CNN branch layer to extract local features. The CNN branch layer extracts fine-grained local information through multiple convolutional pooling blocks. This information is crucial to capturing complex-shaped targets such as logging patches. Each convolutional pooling block consists of a convolutional layer, batch normalization, and activation function ReLU. Through this branch, the model can better process the details in the SAR image, especially those complex boundaries and internal structures of logging patches.

[0037] The CBA layer integrates local features from the CNN branch layer in each patch of the ViT branch layer, enhancing the model's ability to capture detail information. Specifically, the CBA layer uses the global features of the ViT branch layer as a query and establishes a global dependency relationship with the local features of the CNN branch layer, thereby achieving effective fusion of local and global information in the model. After multi-layer feature extraction and fusion, the model generates a 128×128×1 forest felling probability map. The value of each pixel represents the probability that the pixel is judged as forest felling. Pixels exceeding 0.5 are marked as felled areas, and the remaining pixels are non-falling areas. During the training process, the SAMSR model training process uses the dice loss function and the binary cross entropy loss function for supervision, the optimizer is Adam, the initial learning rate is 0.0001, the batch size is 8, and the training is 200 epochs in total.

[0038] Step 4: After the SAMSR model training is completed, transfer learning and active learning are used to evaluate the transferability of the model and realize monthly forest felling monitoring; specifically: migrate the trained SAMSR model to other target areas, and retrain the model based on samples from the target area to determine the best migration strategy.

[0039] Determination of the model migration strategy includes retraining samples from target areas and evaluating the adaptability of the model in these areas.

[0040] Step 5. In the model transfer training, active learning strategy is added to further reduce the required local and target domain samples and improve the model efficiency. Finally, monthly forest felling mapping of the study area is realized based on the SAMSR model and Sentinel-1SAR image data.

[0041] The active learning strategy reduces the number of required samples by selecting the most representative samples in the study area for retraining while ensuring the high performance of the model.

[0042] The felling probability maps output by the SAMSR model are post-processed to generate the final monthly felling maps, which can be used to evaluate the felling activities each month. The SAMSR model is compared with the original SAM model and the U-Net model. The performance evaluation of the proposed SAMSR model and other comparative models includes the overall accuracy of the felling area, user accuracy, producer accuracy, F1-score, and IoU to comprehensively evaluate the performance of the model in different study areas and time. The monthly forest felling mapping results generated based on the SAMSR model can timely reflect the dynamic changes of forest resources and provide accurate forest management decision support for relevant departments.

[0043] Specific implementation method 2: Combination Figures 1 to 5 This embodiment is described as an example of a large-scale forest felling mapping method based on computer vision model SAM and Sentinel-1 time series SAR data described in Specific Embodiment 1:

[0044] Taking deforestation as an example, a deforestation mapping model is constructed using Sentinel-1 data. Based on SAM, a SAMSR model is proposed. The SAMSR model retains the prompt encoder and mask decoder structure, while adding a parallel CNN branch to process the input Sentinel-1 image in parallel with the visual transformer (ViT) branch. The CNN and ViT branch layers work together, and the cross-branch attention module establishes global dependencies between them, thereby producing the final image feature embedding, constructing a deforestation model, and then inputting the Sentinel-1 data of the study area into the optimal model to realize deforestation mapping. The specific process is:

[0045] 1. Collection of felling samples;

[0046] High-resolution reference image data acquisition;

[0047] High-resolution satellite images can clearly identify harvested patches and collect reference data for subsequent SAMSR model training, parameterization, and testing; as shown in Table 1. Table 1 shows the total number of samples in the study area and the number of samples used for training, parameterization, and testing. Among them, PAR is the perimeter area ratio, Q1 is the lower quartile, M is the median, and Q3 is the upper quartile.

[0048] Table 1

[0049]

[0050]

[0051] 3-meter resolution images obtained from the Planet website between January 2016 and January 2023 were used. Specific point locations were generated in different study areas based on the principle of uniform spatiotemporal distribution to ensure sufficient true positive and true negative logging samples. Each 1280m×1280m square plot or its surrounding 128×128 pixel area was used as the basic spatial unit for ground reference markers and dataset segmentation.

[0052] Sentinel-1 data collection;

[0053] Since the end of 2014, the European Space Agency's Sentinel-1 data have been used for deforestation mapping, and the L1-level GRD products are processed to achieve a spatial resolution of 10 meters. This process includes the application of orbit files, the removal of GRD boundaries and thermal noise, followed by radiometric calibration and terrain correction. Given the sensitivity of Sentinel-1 data to terrain changes, the standardized L1-level GRD data are further refined using 30-meter SRTM DEM data to ensure accurate representation of terrain relief.

[0054] 2. Construction of SAMSR model;

[0055] Based on the overall architecture of the SAM model, SAMSR was developed to generate deforestation maps from Sentinel-1 SAR images. Figure 2As shown in Figure 2, the SAMSR model retains the hint encoder and mask decoder structure of SAM while adding parallel CNN branches to extract richer local features. This enhancement is crucial for accurately segmenting complex shapes such as logging patches as it improves boundary and internal structure detection. The CNN branch consists of sequential convolutional pooling blocks that process the input in parallel with the visual transformer (ViT) branch, ensuring that both have the same spatial resolution. The lightweight CNN design prevents overfitting during training. SAMSR’s cross-branch attention (CBA) establishes global dependencies between them, improving segmentation accuracy. The ViT branch layer includes an adapter that fine-tunes the global position embedding and aligns the input features. Initially, the position embedding is downsampled by max pooling with a stride of 2 and a kernel size of 2 to maintain the same resolution as the embedding sequence. Subsequently, a convolution operation with a kernel size of 3×3 is applied to further fine-tune the position embedding, helping the ViT branch to process smaller inputs more effectively. The CNN branch layer and the ViT branch layer work together, and the CBA layer establishes global dependencies between them, resulting in the final image feature embedding. The SAMSR model uses weights trained on the SA-1B dataset, which consists of 1 billion annotated images, to initialize the parameters of SAM. The training process is supervised by a combined loss function, including dice loss and binary cross entropy loss. Adam is used as the optimizer with an initial learning rate of 0.0001 and a batch size of 8, and the model is trained for 200 epochs. The input size is 128×128 pixels with 2 bands (VV and VH), and the output size is 128×128×1, where each pixel value represents the probability of felling, and pixels with a probability greater than 0.5 are classified as felling.

[0056] In this implementation, the construction of a baseline model U-Net is also included;

[0057] A lightweight U-Net model architecture is adopted as a baseline for mapping forest deforestation and compared with the SAMSR model. In the “downsampling” part of the U-Net, pooling reduces the spatial dimension, and we set four convolutional layers where the channels are expanded from 8 to 128 to capture complex features. The “upsampling” part then upgrades these features by merging them with the corresponding downsampled features through skip connections to preserve spatial details. ReLU activation is used in all convolutional layers except the last layer, where the sigmoid function outputs values ​​between 0 and 1. The U-Net is trained using a combination of binary cross entropy loss and dice loss, and the Adam optimizer is set to an initial learning rate of 0.0001.

[0058] 4. Comparison of model accuracy evaluation;

[0059] In this implementation, the performance of the models (SAMSR, U-Net, and SAM) was evaluated using an independent test dataset. To evaluate the accuracy of the models, several metrics were calculated, including overall accuracy, user accuracy, producer accuracy, F1 score, and IoU as defined in equations (1)-(5).

[0060]

[0061] In the formula, TP stands for true positive; TN stands for true negative; FP stands for false positive; and FN stands for false negative. Table 2 shows the accuracy comparison of the three models in each study area.

[0062] Table 2

[0063]

[0064]

[0065] The results show that based on an independent test set, the SAMSR model outperforms the U-Net model and the SAM model in all scenarios, with an overall accuracy ranging from 86% to 94% (the comprehensive model for all study areas reaches 89%). The user and producer accuracies are 77%-83% and 71%-93% respectively (for the comprehensive model for all study areas, the user and producer accuracies are 82% and 80% respectively). The F1 value and IoU value of SAMSR are 0.74-0.88 and 0.58-0.78 respectively, outperforming the SAM and U-Net models by 28.30% and 6.25% in IoU respectively (the comprehensive model for all study areas has an F1 value of 0.81 and an IoU value of 0.68). The combined model based on U-Net for the four study areas has an overall accuracy of 88%, and the user and producer accuracies are 82% and 75% respectively. The F1 is 0.78 and the IoU is 0.64. The comprehensive model based on the SAM model has the lowest accuracy, with an IoU of 0.53.

[0066] like Figure 3 As shown, comparing the spatial mapping results with the reference and baseline prediction results, SAMSR has the greatest similarity with the reference labels in terms of spatial location and boundary consistency. In contrast, the U-Net model produced more false negative results than the SAMSR model, while the SAM model had the worst prediction results.

[0067] 5. Transferability assessment of the SAMSR model;

[0068] To evaluate the transferability of the SAMSR model across the four study sites, the following experiments were conducted: (1) A SAMSR model was trained for each study area using local samples, starting with 10 input objects and then increasing the training data until the performance stabilized (measured by IoU). (2) Each saturated local model was then applied to the remaining target study areas, gradually adding samples from the target area until saturation. (3) Finally, the trained multi-region comprehensive model was applied to another target area, gradually increasing the samples. The appropriate sample size for stabilizing the SAMSR model was evaluated by comparing the performance of the locally trained models and the transferred models, and their transferability across different study sites was assessed. Table 3 summarizes the maximum F1 and IoU values ​​of the transfers of the individual models for each study area.

[0069] Table 3

[0070]

[0071]

[0072] The results show that the multi-region integrated model generally produces better accuracy (IoU 0.60-0.77, F1 0.75-0.87) compared with the single-region model transfer.

[0073] In this implementation, an active learning strategy is adopted;

[0074] When applying transfer learning models, it is crucial to use new local samples to expand the application scenarios. Traditionally, this involves random sample selection, which can have a negative impact on model performance if not done properly. The improved greedy sampling (iGS) method is used to select representative samples, which enhances the model's generalization ability to real-world scenarios. iGS outperforms traditional active learning algorithms and achieves comparable performance with fewer samples while saving computing resources. The iGS process includes: (1) GS is input (GSx), and samples are selected based on their proximity to the centroid. (2) Output GS (GSy), samples are selected based on the diversity of output values. (3) iGS, samples are selected based on the diversity of input and output values. Iteration continues until the maximum number of samples (K) is reached. In order to compare iGS with random sampling, 9 sets of sample gradient experiments were conducted.

[0075] like Figure 4 As shown, Figure 4 Comparison of IoU values ​​of random and active methods based on multi-region comprehensive models under different samples. The results show that active learning methods continuously improve accuracy, especially when there are fewer samples, outperforming random sampling. Active learning reaches or exceeds the saturation accuracy of traditional random selection methods, reducing samples by about 20%–70%.

[0076] 6. If Figure 5 As shown, monthly forest felling maps were drawn;

[0077] Monthly deforestation maps from 2021 to 2023 were produced using a previously determined optimal mapping model (SAMSR).

[0078] The central regions of GX and HN were selected for mapping and analysis. Figure 5 Monthly deforestation maps for each study area are plotted over a long time series, with the three blue circled areas shown in magnified view. Adjacent to these magnified images are Planet images before and after deforestation, and in the lower right corner is a chart of monthly deforestation statistics. The monthly deforestation plotted in the figure is highly consistent with the high-resolution Planet imagery.

[0079] In addition, according to the mapping results, corresponding protection measures can be taken for the logging areas in forest change monitoring, and forests can be protected and managed in a focused manner, which is conducive to precise forest management. Moreover, it also helps biodiversity conservation and carbon flux research because it provides valuable insights into long-term forest dynamics, especially in large ecosystems.

[0080] The technical features of the above-described embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0081] The above-mentioned embodiments only express several implementation methods of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. A large-scale forest felling mapping method based on computer vision models and Sentinel-1 time series SAR data, characterized by: The method consists of the following steps: Step 1: Collect high-resolution Planet image data and Sentinel-1SAR image data of felled samples; Step 2: pre-process the image data collected in step 1 and divide it into a training set, a validation set and a test set; Step 3: Establish a forest felling mapping model SAMSR, use the training set to train the SAMSR model, and obtain a trained SAMSR model; The SAMSR model includes a ViT branch layer, a CNN branch layer, a CBA layer, a prompt encoder and a mask decoder; The training set is input into the ViT branch layer and the CNN branch layer at the same time; the ViT branch layer is used to extract the global features of the image, and the CNN layer is used to extract the local features of the image, and the global features and the local features are input into the CBA layer for feature weighted fusion to extract the feature map; Inputting the feature map into a mask decoder to generate an image segmentation mask; The real image segmentation map of the prompt encoder is input into the mask decoder and compared with the generated image segmentation mask to finally generate a forest felling probability map; Step 4: Use the test set to test the trained SAMSR model, and use the SAMSR model and Sentinel-1 data to perform monthly forest felling mapping.

2. The large-scale forest felling mapping method based on computer vision model and Sentinel-1 time series SAR data according to claim 1, characterized in that: In step 2, data preprocessing includes preprocessing the image data into a consistent size; using the high-resolution Planet image data as annotation data; removing noise, radiometric calibration, terrain correction and data cropping of the Sentinel-1SAR image data.

3. The large-scale forest felling mapping method based on computer vision model and Sentinel-1 time series SAR data according to claim 1, characterized in that: Step three also includes setting the number of epochs for training the SAMSR model. During the SAMSR model training process, the dice loss function and the binary cross entropy loss function are used for supervision, the optimizer is Adam, the initial learning rate is 0.0001, and the batch size is 8. After each epoch, the SAMSR model calculates the IoU on the validation set. If the current IoU is better than the previous best IoU, the current SAMSR model is the optimal model until the end of training.

4. The large-scale forest felling mapping method based on computer vision model and Sentinel-1 time series SAR data according to claim 1, characterized in that: The training set uses Sentinel-1SAR images in the VV and VH polarization bands and label images with a resolution of 128×128 pixels.

5. The large-scale forest felling mapping method based on computer vision model and Sentinel-1 time series SAR data according to claim 1, characterized in that: Each pixel value in the felling probability map is the felling probability, and pixels with a felling probability greater than 0.5 are felled.

6. The large-scale forest felling mapping method based on computer vision model and Sentinel-1 time series SAR data according to claim 1, characterized in that: The ViT branch layer includes an adapter, a window transformer and a global transformer; The adapter is used to embed and resize the input image, and then divide the image with embedded feature map into multiple local windows through window transformer, each window is processed independently, and finally the global transformer with multi-head attention mechanism is used to extract global features and input them into CBA layer.

7. The large-scale forest felling mapping method based on computer vision model and Sentinel-1 time series SAR data according to claim 1, characterized in that: The CNN branch layer extracts local features of the image through multiple convolutional pooling blocks and inputs the local features into the CBA layer so that the features of the two branches are weightedly fused to extract the feature map.

8. The large-scale forest felling mapping method based on computer vision model and Sentinel-1 time series SAR data according to claim 1, characterized in that: Step 4 also includes evaluating the performance of the SAMSR model; specifically: The trained SAMSR model was tested for its transferability, and the optimal migration strategy of the SAMSR model was determined based on retraining of samples in the target area. An active learning strategy was also added. Finally, monthly forest felling mapping of the study area was achieved based on the SAMSR model and Sentinel-1 data.

9. The large-scale forest felling mapping method based on computer vision model and Sentinel-1 time series SAR data according to claim 8, characterized in that: The migration strategy includes retraining samples in the target area to evaluate the adaptability of the SAMSR model in the target area; the active learning strategy is to retrain by selecting representative samples in the study area.

Images