Ecological environment risk monitoring and analysis method and system based on image recognition

Abstract

The invention relates to the technical field of ecological environment risk monitoring, and discloses an ecological environment risk monitoring and analysis method and system based on image recognition, and the method comprises the steps: obtaining the aerial photography monitoring information of a target frozen soil region, constructing a two-dimensional reference feature library of an aerial photography image in each category, carrying out the preprocessing and enhancement processing of the aerial photography image, and obtaining a two-dimensional feature library of the aerial photography image; and performing confusion suppression processing on the enhanced feature map by using background reference features and spectral difference attention weights, performing feature splicing in combination with extracted shallow features, middle features and edge features, outputting a semantic segmentation mask, and generating an ecological environment risk analysis result. Therefore, on the basis of the aerial image of the unmanned aerial vehicle, the aerial correlation parameters and the surface temperature, remote sensing adaptive radiation correction is applied to low-altitude monitoring optimization through two-dimensional scene division and reference feature library construction, so that accurate identification and dynamic quantification of a risk target are realized; and a reliable technical support is provided for evaluating the ecological environment risk of the plateau permafrost region.

Description

Technical Field

[0001] This invention relates to the field of ecological and environmental risk monitoring technology, and in particular to an ecological and environmental risk monitoring and analysis method and system based on image recognition. Background Technology

[0002] As an ecologically fragile zone, the dynamic changes in the surface micro-topography (permafrost cracks 2-5 cm wide and vegetation degradation patches 5-10 m² in area) of the plateau permafrost region are the core basis for ecological risk assessment.

[0003] Current mainstream monitoring relies on drone aerial photography and image processing technology, but this scenario has highly specific pain points: First, the strong ultraviolet radiation at high altitudes easily leads to spectral distortion in RGB images, with high spectral overlap between permafrost, degraded vegetation patches, and non-degraded vegetation; second, the variability in drone shooting conditions exacerbates distortion. Variations in shooting angles (pitch ±15°, roll ±10°) caused by different time periods (40%-60% difference in ultraviolet intensity between morning, noon, and evening) and different flight paths alter the spectral reflection path, while different shooting angles (front lighting / backlighting / side lighting) reduce the spectral difference between the target and the background; third, seasonal freeze-thaw cycles (freezing / thawing / stabilization periods) cause dynamic changes in the morphology and spectrum of risk targets, with the same target exhibiting 30%-40% grayscale differences at different stages. Traditional semantic segmentation models struggle to establish a correlation between shooting conditions, freeze-thaw stages, and the true characteristics of the target, resulting in target features being overwhelmed by noise. This ultimately leads to high rates of missed and false detections of risk targets, large area / width quantification errors, and an inability to accurately support dynamic assessment of ecological risks in high-altitude permafrost regions.

[0004] Therefore, improving the accuracy and precision of ecological risk assessment for monitoring dynamic changes in surface micro-topography in plateau permafrost regions is a technical problem that urgently needs to be solved. Summary of the Invention

[0005] This invention provides a method and system for monitoring and analyzing ecological and environmental risks based on image recognition, aiming to solve at least one of the above-mentioned technical problems.

[0006] To achieve the above objectives, this invention provides a method for monitoring and analyzing ecological and environmental risks based on image recognition, the method comprising the following steps: S1: Obtain aerial monitoring information of the target permafrost area; wherein, the aerial monitoring information includes aerial images, aerial correlation parameters, and surface temperature collected by the UAV during patrol monitoring; S2: Based on the aerial photography correlation parameters and the surface temperature, the aerial images are divided into several categories, and features are extracted from the aerial images of each category to construct a two-dimensional benchmark feature library; S3: Based on the aerial photography association parameters, perform preprocessing on the aerial photography image, including viewpoint correction, channel weight adjustment and filtering, to obtain a preprocessed image. Call the dual-dimensional benchmark feature library to perform enhancement processing on the preprocessed image, including weighted radiometric correction, freeze-thaw stage feature binding and spectral enhancement, to obtain an enhanced feature map. S4: Based on the background reference features determined by the dual-dimensional benchmark feature library and the spectral difference attention weights calculated during the spectral enhancement process, the enhanced feature map is subjected to confusion suppression processing to obtain the cleaned features. The shallow features, middle features and edge features of the preprocessed image are extracted and cross-layer feature concatenation is performed with the cleaned features to obtain the fused feature map output by the decoding end. S5: Perform Softmax activation on the fused feature map, calculate the probability of each pixel belonging to the risk target and the background, determine the pixels with a probability higher than the threshold as risk targets, and determine the pixels with a probability lower than the threshold as background, and output the pixel-level semantic segmentation mask. S6: Optimize the semantic segmentation mask by performing morphological opening operation, calculate quantitative indicators for permafrost cracks and degraded patches respectively, and generate ecological and environmental risk analysis results for the target permafrost area.

[0007] Optionally, step S1: Obtain aerial monitoring information of the target permafrost area, specifically including: S11: Control the drone to perform cruise monitoring at several preset monitoring points in the target permafrost area; S12: Acquire aerial images, shooting angle parameters, shooting time, and ground temperature collected by the camera equipment on the drone; S13: The shooting angle parameters and shooting time are summarized into aerial photography associated parameters, which, together with the aerial images and surface temperature, constitute the aerial monitoring information of the target permafrost area.

[0008] Optionally, in step S2, the aerial image is divided into several categories based on the aerial photography association parameters and the ground temperature, including: S21: generating 3 sets of scene division indicators based on the aerial photography association parameters and the ground temperature; The scene division indicators include three freeze-thaw stages determined based on the range of surface temperature, three shooting periods determined based on the range of shooting time, and three shooting angle types determined based on shooting angle parameters and shooting time. S22: Based on the freeze-thaw stage, shooting time, and shooting angle of each aerial image, the aerial images collected during the patrol monitoring are divided into 27 categories.

[0009] Optionally, in step S2, feature extraction is performed on aerial images of each category to construct a two-dimensional benchmark feature library, specifically including: S23: Select 200 sample images for each category and label the risk target areas; S24: Extract the average grayscale of the RGB channels of the risk target, the channel variance, the average grayscale of the RGB channels of the background, and the spectral difference between the target and the background for each category of the scene, and construct a two-dimensional benchmark feature library; The expression for the average grayscale of the RGB channels of the risk target is as follows: In the formula, Let be the number of sample images for the category scene corresponding to the shooting conditions at stage k and condition p, and set to 200. This represents the total number of pixels of the risky target in all sample images within this scene category. Let c be the set of risk target pixels in the c-th channel of the i-th image in this category of scene. Let be the grayscale value of the (x,y) pixel and c channel of the i-th image; This represents the average grayscale value of the c-th channel of the target in this category of scenarios, reflecting the basic spectral characteristics of the risk target. The specific expression for the channel variance of the risk objective is as follows: In the formula, The gray-level variance of the c-th channel of the target in this category of scenarios reflects the target risk. Stability of spectral characteristics; The expression for the average grayscale of the RGB channels in the background is as follows: In the formula, For the first stage, Under the shooting conditions, the total number of pixels in the background area of ​​200 sample images; This is the set of background pixels in the c-th channel of the i-th image in this category of scene; For the background of this category of scenes Average gray level of the channel; The expression for the spectral difference between the target and the background is as follows: In the formula, The mean grayscale difference between the target and the background in the c-th channel reflects the spectral distinction between the target and the background.

[0010] Optionally, in step S3, the aerial image is preprocessed according to the aerial photography correlation parameters, including viewpoint correction, channel weight adjustment, and filtering, to obtain a preprocessed image, specifically including: S31: Based on the shooting angle parameter in the aerial photography association parameters, perform viewing angle correction on each pixel of the aerial image to correct the deviation of the ground surface spectral reflection path caused by the shooting angle; S32: Based on the shooting time corresponding to the shooting time in the aerial photography association parameters, call the preset RGB channel gain vector to adjust the channel weights of the image after the viewpoint correction; S33: Using the image after channel weight adjustment as the guide image, guided filtering is used to denoise the image to obtain the preprocessed image.

[0011] Optionally, in step S3, the dual-dimensional benchmark feature library is invoked to perform enhancement processing on the preprocessed image, including weighted radiometric correction, freeze-thaw stage feature binding, and spectral enhancement, to obtain an enhanced feature map, specifically including: S34: Calculate the angle weight and time period weight based on the shooting angle parameter and shooting time corresponding to the shooting time in the aerial photography association parameters, and determine the comprehensive shooting condition weight; S35: Call the average gray level and channel variance of the RGB channels in the dual-dimensional benchmark feature library, and combine them with the global mean and variance of the preprocessed image to correct the spectral deviation of the image through weighted radiometric correction; S36: Call the surface temperature from the aerial monitoring information, normalize the surface temperature, combine the freeze-thaw stage feature coefficients, perform feature modulation on the weighted radiometrically corrected image, and bind the freeze-thaw stage features. S37: Convolve the dynamic spectral features of the binding freeze-thaw stage to extract spectral features, call the spectral difference between the target and the background in the dual-dimensional benchmark feature library, construct attention weights to amplify the spectral difference between the small target and the background, and obtain the enhanced feature map after batch normalization.

[0012] Optionally, step S4: Based on the background reference features determined by the dual-dimensional benchmark feature library and the spectral difference attention weights calculated during the spectral enhancement process, the enhanced feature map is subjected to confusion suppression processing to obtain cleaned features. The shallow, mid-level, and edge features of the preprocessed image are extracted and concatenated with the cleaned features across layers to obtain the fused feature map output from the decoding end. Specifically, this includes: S41: Based on the background reference features in the dual-dimensional baseline feature library and the spectral difference attention weights calculated during the spectral enhancement process, the enhanced feature map is subjected to confusion suppression processing to obtain the cleaned features. The specific expression is as follows: In the formula, It is a 5×5 Gaussian kernel. It is a 3×3 convolution kernel. As a background reference feature, The average gray level of the 3-channel background is represented by λ=0.35, which is the suppression weight. For background reference features in the two-dimensional benchmark feature library, As an anti-attention weight, the lower the spectral difference (the more easily confused the region), the higher the anti-attention weight and the stronger the suppression. The operation is a 3×3 convolution, which extracts the features of the confused region through convolution and then suppresses them through subtraction. The cleaned features output from the encoding end; S42: Perform convolution operations of different lengths on the preprocessed image to extract shallow and mid-level features. Then, use the Canny edge detection algorithm to extract the edge features of the preprocessed image. The expression is:

[0013] In the formula, It is a 3×3 convolution kernel. Extract low-level features such as shallow edges and textures; It is a 3×3 convolution kernel. Extract local contour features from the middle layer; This is a shallow feature map. This is a mid-level feature map. The Canny edge detection algorithm is used to extract image edge information; the edge information is then multiplied by 0.1 for intensity normalization to prevent edge features from overly strong and masking other features. For edge feature maps; S43: Upsample the shallow, mid-level, and edge features, concatenate them with the cleaned features output from the encoder, and then perform 1×1 convolution dimensionality reduction to obtain the final fused feature map from the decoder. The expression is: In the formula, For transposed convolution upsampling operations, the resolution of mid-layer, shallow-layer, and edge features is restored to the same level as... Consistent; For feature splicing operations, deep semantic features, mid-level contour features, shallow edge features, and specialized edge completion features are integrated; Using 1×1 convolution, the concatenated high-dimensional features are reduced to 256 channels; This is the final fused feature map at the decoding end.

[0014] Optionally, in step S5, the pixel-level semantic segmentation mask expression is output as follows: In the formula, The function maps feature values ​​to a probability distribution, with 0.7 as the confidence threshold. This is the semantic segmentation mask for the output.

[0015] Optionally, step S6: Optimize the semantic segmentation mask using morphological opening operations, calculate quantitative indicators for permafrost cracks and degraded patches respectively, and generate ecological environment risk analysis results for the target permafrost area, specifically including: S61: Optimize the semantic segmentation mask by performing morphological opening operation; S62: Based on the optimized mask, core quantitative indicators are calculated for permafrost cracks and vegetation degradation patches respectively, generating ecological and environmental risk analysis results including permafrost crack width, vegetation degradation patch area and dynamic change rate.

[0016] Furthermore, to achieve the above objectives, the present invention also provides an image recognition-based ecological environment risk monitoring and analysis system, comprising: The information acquisition module is used to acquire aerial monitoring information of the target permafrost area; wherein, the aerial monitoring information includes aerial images, aerial correlation parameters and surface temperature collected by the UAV during patrol monitoring; The feature library construction module is used to divide the aerial images into several categories based on the aerial photography association parameters and the surface temperature, extract features from the aerial images of each category, and construct a two-dimensional benchmark feature library. The feature processing module is used to perform preprocessing on the aerial image based on the aerial image association parameters, including viewpoint correction, channel weight adjustment and filtering, to obtain a preprocessed image. The module then calls the two-dimensional benchmark feature library to perform enhancement processing on the preprocessed image, including weighted radiometric correction, freeze-thaw stage feature binding and spectral enhancement, to obtain an enhanced feature map. The fusion feature output module is used to perform confusion suppression processing on the enhanced feature map based on the background reference features determined by the dual-dimensional benchmark feature library and the spectral difference attention weights calculated during the spectral enhancement process, to obtain the cleaned features. The module also extracts the shallow features, middle features and edge features of the preprocessed image and performs cross-layer feature concatenation with the cleaned features to obtain the fusion feature map output by the decoding end. The mask output module is used to perform Softmax activation on the fused feature map, calculate the probability of each pixel belonging to the risk target and the background, determine the pixels with a probability higher than the threshold as risk targets, and determine the pixels with a probability lower than the threshold as background, and output a pixel-level semantic segmentation mask. The risk analysis module is used to optimize the semantic segmentation mask by performing morphological opening operations, calculate quantitative indicators for permafrost cracks and degraded patches respectively, and generate ecological and environmental risk analysis results for the target permafrost area.

[0017] The beneficial effects of this invention are as follows: It proposes an ecological environment risk monitoring and analysis method and system based on image recognition. By acquiring aerial monitoring information of the target permafrost area, the aerial images are divided into several categories and a two-dimensional benchmark feature library is constructed based on the aerial imaging correlation parameters and surface temperature. The aerial images are then preprocessed, including viewpoint correction, channel weight adjustment and filtering, and enhanced, including weighted radiometric correction, freeze-thaw stage feature binding and spectral enhancement. Then, the enhanced feature map is subjected to confusion suppression processing using background reference features and spectral difference attention weights. Shallow features, middle features and edge features are extracted and cross-layer feature splicing is performed with the purified features. The semantic segmentation mask is output and calculated to generate the ecological environment risk analysis results of the target permafrost area. Therefore, this invention uses UAV aerial RGB images and surface temperature as the basis for aerial imaging parameters. Through dual-dimensional scene segmentation and the construction of a benchmark feature library, it applies remote sensing adaptive radiometric correction to low-altitude monitoring optimization. It employs shooting condition weight modulation, freeze-thaw stage feature binding, and micro-target spectral enhancement, combined with enhanced semantic segmentation based on spectral confusion suppression and cross-layer feature fusion, to achieve pixel-level identification of risk targets. This breaks through the limitations of traditional methods in the special scenarios of plateau permafrost regions, ultimately achieving accurate identification and dynamic quantification of risk targets. It provides reliable technical support for ecological environment risk assessment in plateau permafrost regions, helps to accurately grasp the impact of freeze-thaw cycles on the ecological environment, and provides a scientific basis for ecological protection decisions. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the image recognition-based method for monitoring and analyzing ecological and environmental risks according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of the image recognition-based ecological environment risk monitoring and analysis system according to an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0020] This invention provides a method for monitoring and analyzing ecological and environmental risks based on image recognition, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the ecological environment risk monitoring and analysis method based on image recognition, as described in an embodiment of the present invention.

[0021] In this embodiment, an ecological environment risk monitoring and analysis method based on image recognition is provided, the method comprising the following steps: S1: Obtain aerial monitoring information of the target permafrost area; wherein, the aerial monitoring information includes aerial images, aerial correlation parameters, and surface temperature collected by the UAV during patrol monitoring; S2: Based on the aerial photography correlation parameters and the surface temperature, the aerial images are divided into several categories, and features are extracted from the aerial images of each category to construct a two-dimensional benchmark feature library; S3: Based on the aerial photography association parameters, perform preprocessing on the aerial photography image, including viewpoint correction, channel weight adjustment and filtering, to obtain a preprocessed image. Call the dual-dimensional benchmark feature library to perform enhancement processing on the preprocessed image, including weighted radiometric correction, freeze-thaw stage feature binding and spectral enhancement, to obtain an enhanced feature map. S4: Based on the background reference features determined by the dual-dimensional benchmark feature library and the spectral difference attention weights calculated during the spectral enhancement process, the enhanced feature map is subjected to confusion suppression processing to obtain the cleaned features. The shallow features, middle features and edge features of the preprocessed image are extracted and cross-layer feature concatenation is performed with the cleaned features to obtain the fused feature map output by the decoding end. S5: Perform Softmax activation on the fused feature map, calculate the probability of each pixel belonging to the risk target and the background, determine the pixels with a probability higher than the threshold as risk targets, and determine the pixels with a probability lower than the threshold as background, and output the pixel-level semantic segmentation mask. S6: Optimize the semantic segmentation mask by performing morphological opening operation, calculate quantitative indicators for permafrost cracks and degraded patches respectively, and generate ecological and environmental risk analysis results for the target permafrost area.

[0022] It should be noted that current mainstream monitoring relies on drone aerial photography and image processing technology, but this scenario has highly specific pain points: First, the strong ultraviolet radiation at high altitudes easily leads to spectral distortion in RGB images, with high spectral overlap between permafrost, degraded vegetation patches, and non-degraded vegetation; second, the variability in drone shooting conditions exacerbates distortion. Variations in shooting angles (pitch ±15°, roll ±10°) caused by different time periods (40%-60% difference in ultraviolet intensity between morning, noon, and evening) and different flight paths alter the spectral reflection path, and different shooting angles (front lighting / backlighting / side lighting) reduce the spectral difference between the target and the background; third, seasonal freeze-thaw cycles (freezing period / thawing period / stable period) cause dynamic changes in the morphology and spectrum of risk targets, with the grayscale difference of the same target reaching 30%-40% at different stages. Traditional semantic segmentation models struggle to establish a correlation between shooting conditions, freeze-thaw stages, and the true characteristics of the target, resulting in target features being overwhelmed by noise, ultimately leading to high rates of missed detections and false positives for risk targets, large area / width quantification errors, and an inability to accurately support dynamic assessment of ecological risks in high-altitude permafrost regions.

[0023] To address the aforementioned issues, this embodiment acquires aerial monitoring information of the target permafrost region. Based on aerial imaging parameters and surface temperature, the aerial images are categorized into several classes, and a two-dimensional baseline feature library is constructed. The aerial images are then preprocessed, including viewpoint correction, channel weight adjustment, and filtering, and enhanced, including weighted radiometric correction, freeze-thaw stage feature binding, and spectral enhancement. Background reference features and spectral difference attention weights are used to perform obfuscation suppression processing on the enhanced feature maps. Shallow, mid-layer, and edge features are extracted and combined with the purified features for cross-layer feature concatenation. A semantic segmentation mask is output and calculated to generate the ecological and environmental risk analysis results for the target permafrost region. Therefore, this invention uses UAV aerial RGB images and surface temperature as the basis for aerial imaging parameters. Through dual-dimensional scene segmentation and the construction of a benchmark feature library, it applies remote sensing adaptive radiometric correction to low-altitude monitoring optimization. It employs shooting condition weight modulation, freeze-thaw stage feature binding, and micro-target spectral enhancement, combined with enhanced semantic segmentation based on spectral confusion suppression and cross-layer feature fusion, to achieve pixel-level identification of risk targets. This breaks through the limitations of traditional methods in the special scenarios of plateau permafrost regions, ultimately achieving accurate identification and dynamic quantification of risk targets. It provides reliable technical support for ecological environment risk assessment in plateau permafrost regions, helps to accurately grasp the impact of freeze-thaw cycles on the ecological environment, and provides a scientific basis for ecological protection decisions.

[0024] In a preferred embodiment, step S1: obtaining aerial monitoring information of the target permafrost area, specifically including: S11: Control the drone to perform cruise monitoring at several preset monitoring points in the target permafrost area; S12: Acquire aerial images, shooting angle parameters, shooting time, and ground temperature collected by the camera equipment on the drone; S13: The shooting angle parameters and shooting time are summarized into aerial photography associated parameters, which, together with the aerial images and surface temperature, constitute the aerial monitoring information of the target permafrost area.

[0025] In this embodiment, a multi-rotor UAV is used to conduct quarterly aerial photography at pre-set monitoring points in the target permafrost area. The flight altitude is 30-50m above the ground surface, and RGB three-channel images are acquired. Key parameters are recorded simultaneously: shooting time, shooting angle parameters (pitch angle θ, roll angle φ), and ground surface temperature. (The temperature was collected by the temperature measurement component carried by the drone). The flight path was designed to traverse the canyon longitudinally and cover the open area laterally to ensure coverage from different shooting angles. The shooting frequency was set to 1 frame / 5 seconds.

[0026] Based on this, in step S2, the aerial images are divided into several categories based on the aerial photography correlation parameters and the surface temperature, including: S21: Based on the aerial photography correlation parameters and the surface temperature, generate 3 sets of scene division indicators; The scene division indicators include three freeze-thaw stages determined based on the range of surface temperature, three shooting periods determined based on the range of shooting time, and three shooting angle types determined based on shooting angle parameters and shooting time. S22: Based on the freeze-thaw stage, shooting time, and shooting angle of each aerial image, the aerial images collected during the patrol monitoring are divided into 27 categories.

[0027] In this embodiment, three freeze-thaw stages are determined based on the range of surface temperature. The criteria for dividing the freeze-thaw stages are: T1 (freezing period: ≤0℃), T2 (ablation period: 0℃< <10℃), T3 (stable period: ≥10℃); Three shooting time periods are determined based on the range of shooting time. (Morning: 6-10 AM = 1, Noon: 10-4 PM = 2, Evening: 4-8 PM = 3); Determine 3 shooting angle types (front lighting / backlighting / side lighting) based on the shooting angle parameters and the lighting angle corresponding to the shooting time.

[0028] In a preferred embodiment, step S2 involves extracting features from aerial images of each category to construct a two-dimensional benchmark feature library, specifically including: S23: Select 200 sample images for each category and label the risk target areas; S24: Extract the average grayscale of the RGB channels of the risk target, the channel variance, the average grayscale of the RGB channels of the background, and the spectral difference between the target and the background for each category of the scene, and construct a two-dimensional benchmark feature library; The expression for the average grayscale of the RGB channels of the risk target is as follows: In the formula, Let be the number of sample images for the category scene corresponding to the shooting conditions at stage k and condition p, and set to 200. This represents the total number of pixels of the risky target in all sample images within this scene category. Let c be the set of risk target pixels in the c-th channel of the i-th image in this category of scene. Let be the grayscale value of the (x,y) pixel and c channel of the i-th image; This represents the average grayscale value of the c-th channel of the target in this category of scenarios, reflecting the basic spectral characteristics of the risk target. The specific expression for the channel variance of the risk objective is as follows: In the formula, This represents the gray-level variance of the c-th channel of the target in this category of scenarios, reflecting the stability of the target's risk spectral characteristics. The expression for the average grayscale of the RGB channels in the background is as follows: In the formula, For the first stage, Under the shooting conditions, the total number of pixels in the background area of ​​200 sample images; This is the set of background pixels in the c-th channel of the i-th image in this category of scene; For the background of this category of scenes Average gray level of the channel; The expression for the spectral difference between the target and the background is as follows: In the formula, The mean grayscale difference between the target and the background in the c-th channel reflects the spectral distinction between the target and the background.

[0029] In this embodiment, 27 scene categories are divided according to freeze-thaw stage (T1 / T2 / T3) × shooting time (morning / noon / evening) × shooting angle type (front lighting / backlighting / side lighting). By selecting 200 sample images for each scene category, the risk target area is manually labeled and features such as the average gray level of the RGB channels of the risk target, channel variance, the average gray level of the RGB channels of the background, and the spectral difference between the target and the background are extracted to construct a two-dimensional benchmark feature library.

[0030] Therefore, by segmenting the scene, the specificity of the benchmark features is ensured, and the feature mixing under different freeze-thaw stages and shooting conditions is avoided. At the same time, the extracted average gray level and variance provide a basic reference for subsequent spectral correction, and the newly added spectral difference index provides a basis for the enhancement of small target features. The constructed dual-dimensional benchmark feature library breaks through the limitations of traditional radiometric correction with a single benchmark and lack of scene adaptation, and provides a precise feature reference standard for subsequent processing.

[0031] In a preferred embodiment, step S3 involves preprocessing the aerial image according to the aerial photography correlation parameters, including viewpoint correction, channel weight adjustment, and filtering, to obtain a preprocessed image. Specifically, this includes: S31: Based on the shooting angle parameter in the aerial photography association parameters, perform viewpoint correction on each pixel of the aerial image to correct the deviation of the ground surface spectral reflection path caused by the shooting angle. The expression is: In the formula, These are the pixel values ​​of the image after angle normalization; , These are the pitch angle and roll angle parameters in the shooting angle parameters; , This is the angle correction coefficient. The larger the angle, the smaller the cosine value, and the larger the correction amplitude. It is used to correct brightness deviations caused by changes in the reflection path. To compensate for the angle offset and avoid feature loss due to excessively low pixel values ​​after correction.

[0032] Regarding the normalization of shooting angle, this embodiment uses trigonometric function relationships to quantify the impact of shooting angle on spectral reflectance. By dynamically adjusting the pixel grayscale value through angle coefficients, the same target under different shooting angles has a similar brightness base, reducing the differences in spectral reflectance caused by changes in pitch and roll angles, and initially alleviating spectral confusion caused by shooting from different flight paths.

[0033] S32: Based on the shooting time period corresponding to the shooting time in the aerial photography association parameters, call the preset RGB channel gain vector to adjust the channel weights of the image after perspective correction. The expression is: In the formula, For the first The gain vectors of the RGB channels corresponding to the time period are c=1, 2, and 3, which correspond to the R, G, and B channels, respectively. At noon, the ultraviolet radiation is strongest, and the red channel is most affected by ultraviolet radiation, so the gain coefficient is the lowest. Differentiated gain adjustment is used to offset the influence of ultraviolet radiation on different channels at different times.

[0034] Regarding the time-adaptive channel gain adjustment, this embodiment considers the differences in ultraviolet intensity at different times. For the different RGB channels' sensitivity to ultraviolet light (the red channel is the most sensitive and the blue channel is the least sensitive), a differentiated gain vector is designed. Through multiplication adjustment, the spectrum of each channel is equalized, which cancels the spectral distortion caused by the difference in ultraviolet intensity in the morning, noon and evening, and further alleviates the spectral confusion caused by the superposition of multiple shooting conditions.

[0035] S33: Using the channel-weighted adjusted image as the guide image, guided filtering is applied to denoise the image, resulting in a preprocessed image. The expression is: In the formula, ε is the guided filtering function; r=5 is the filtering window radius, balancing noise reduction and detail preservation; ε=0.02 is the smoothing coefficient, controlling the smoothness of the filtering. These are the pixel values ​​of the preprocessed image.

[0036] For guided filtering denoising, this embodiment uses the gain-adjusted image as the guide image and employs guided filtering to denoise the image. This suppresses high-frequency noise caused by airflow turbulence while preserving the edge features of hazardous targets (such as permafrost crack boundaries and vegetation degradation patch outlines). Thus, by controlling the filtering process through the edge structure of the guide image, the filtered image retains the characteristics of the guided image. Figure 1 It effectively removes edge features and avoids edge blurring caused by traditional mean filtering and median filtering. It also effectively removes high-frequency noise caused by airflow turbulence in drone aerial photography while preserving the edge details of small targets, providing a high-quality image data foundation for subsequent feature extraction and semantic segmentation.

[0037] In this embodiment, the RGB image is first normalized based on the collected shooting angle parameters to correct the spectral reflection path deviation. Then, the RGB channel gain is dynamically adjusted according to the ultraviolet intensity difference during the shooting time to specifically counteract ultraviolet interference at different times. Finally, guided filtering is used to remove high-frequency noise during aerial photography, retain the edge features of risky targets, and output the preprocessed image.

[0038] Therefore, a two-dimensional preprocessing strategy of angle normalization and time-time adaptation gain is proposed. The quantitative parameters such as shooting angle and time time are bound to the image preprocessing depth. Differentiated gain vectors are designed for the sensitivity differences of RGB channels to ultraviolet light. Combined with guided filtering, a balance between noise reduction and detail preservation is achieved. The three-step collaborative approach initially alleviates the spectral confusion problem caused by the superposition of multiple shooting conditions. It provides low-noise, high-fidelity image data for subsequent core radiometric correction and feature enhancement, and avoids the superposition effect of complex interference in the original image on subsequent processing.

[0039] In a preferred embodiment, step S3 involves calling the dual-dimensional reference feature library to perform enhancement processing on the preprocessed image, including weighted radiometric correction, freeze-thaw stage feature binding, and spectral enhancement, to obtain an enhanced feature map. Specifically, this includes: S34: Based on the shooting angle parameter and the shooting time corresponding to the shooting period in the aerial photography association parameters, calculate the angle weight and the time period weight, and determine the comprehensive shooting condition weight. The expression is: In the formula, This is the angle weight, with a value range of [1, 1.2]; the larger the angle, the greater the weight. The smaller the value, the greater the angle weight, and the stronger the subsequent correction, ensuring the correction effect when shooting at large angles. The weighting is based on time period. Ultraviolet radiation is strongest at midday, resulting in the most severe spectral distortion, hence the highest weight is given to strengthen the correction during this time. Spectral distortion is weaker in the morning, so the lowest weight is given to avoid over-correction. To comprehensively consider the weighting of shooting conditions and integrate the effects of angle and time period on spectral distortion, dynamic adaptation of correction intensity is achieved.

[0040] During this step, the shooting condition weighting is first performed by calling the pitch angle θ, roll angle φ, and shooting time period. The angle weight and time period weight are calculated, and the comprehensive shooting condition weight is obtained by multiplying the two, which is used to dynamically adjust the intensity of subsequent radiometric correction.

[0041] Therefore, by using weighted modulation based on shooting conditions to quantify the impact of shooting angle and time period on spectral distortion, radiometric correction is upgraded from a uniform intensity to a dynamically adapted intensity. This strengthens the correction of severe distortion at large angles and during periods of strong ultraviolet radiation, while providing moderate correction for slight distortion at small angles and during periods of weak ultraviolet radiation. This overcomes the limitation of traditional radiometric correction in treating different shooting conditions without differentiation, and improves the accuracy of the correction.

[0042] S35: The average gray level and channel variance of the RGB channels in the dual-dimensional benchmark feature library are called, and combined with the global mean and variance of the preprocessed image, the spectral deviation of the image is corrected through weighted radiometric correction. The specific expression is as follows: In the formula, , The global mean and variance of the c-th channel of the preprocessed image are obtained. The feature distribution of the preprocessed image is aligned to the distribution range of the reference feature by variance normalization. Then, the correction amplitude is adjusted by comprehensive weighting. Finally, the reference average gray level is superimposed to achieve spectral restoration. These are the pixel values ​​of the image after weighted radiometric correction.

[0043] For weighted radiometric correction, this embodiment is based on the idea of ​​statistical feature matching. It aligns the spectral features of the preprocessed image with the two-dimensional benchmark features, eliminates the differences in feature distribution through variance normalization, and dynamically adjusts the correction amplitude by comprehensively weighting. This unifies the target spectral features under different shooting conditions and freeze-thaw stages into the benchmark system, which greatly alleviates the spectral confusion caused by the superposition of strong ultraviolet radiation and shooting conditions.

[0044] S36: Retrieve the surface temperature from aerial monitoring information. After normalizing the surface temperature, combine it with freeze-thaw stage feature coefficients to perform feature modulation on the weighted radiometrically corrected image, binding freeze-thaw stage features. The expression is: In the formula, To normalize the surface temperature, The characteristic coefficients for the freeze-thaw stage are T1=0.3, T2=0.5, and T3=0.2 for the three freeze-thaw stages, and are calibrated based on the fluctuation amplitude of the target characteristics at different stages. To bind the dynamic spectral characteristics of the freeze-thaw stage, the feature intensity is modulated by the temperature normalization value to make the target features consistent across different freeze-thaw stages.

[0045] To address the binding of freeze-thaw stage features, this embodiment utilizes the strong correlation between surface temperature and freeze-thaw stage. By normalizing the temperature, the freeze-thaw stage state of the target is quantified, and the feature intensity is dynamically adjusted in combination with the stage feature coefficient to offset the natural fluctuations of the target spectrum under different freeze-thaw stages. This solves the problem that traditional static radiometric correction cannot adapt to the dynamic changes of the target, and achieves dynamic alignment between freeze-thaw stage and target features.

[0046] S37: Convolve the dynamic spectral features of the freeze-thaw phase to extract spectral features, call the spectral difference between the target and the background in the dual-dimensional benchmark feature library, construct attention weights to amplify the spectral difference between small targets and the background, and obtain the enhanced feature map after batch normalization. The expression is: In the formula, A 3×3 convolution kernel is used; stride=8 is the stride, which realizes feature dimensionality reduction and abstract extraction. For the extracted high-dimensional spectral feature map, , The attention weight matrix and bias vector are learned through training. The spectral difference between the target and the background; Using the Sigmoid activation function, the attention weights are mapped to the [0,1] interval; The attention weights are based on spectral differences; the greater the difference, the higher the weight. This is the enhanced feature map after fusion. To perform batch normalization operations, the feature distribution is standardized, thus improving feature stability; This is the final enhanced feature map output.

[0047] For the spectral enhancement of small targets, this embodiment uses convolution to extract the high-dimensional spectral features of the target. The attention mechanism built based on spectral difference can accurately identify small target regions. By multiplying, the feature response is amplified, making the features of small targets that were originally submerged by background noise stand out. This solves the pain points of low pixel ratio and weak features of small targets, and provides strong feature support for subsequent semantic segmentation.

[0048] In this embodiment, firstly, based on the constructed two-dimensional benchmark feature library, the freeze-thaw stage and shooting conditions of the current image are matched. The comprehensive weight is calculated by shooting angle and time period, and the preprocessed image is subjected to weighted radiometric correction. Then, the freeze-thaw stage feature binding is realized by combining the surface temperature, and the target features are dynamically aligned. Next, attention weights are constructed by using the spectral difference between the target and the background to amplify the small target features. Finally, the enhanced feature map is output through convolution and batch normalization.

[0049] Therefore, three scenario-based improvements are made to the traditional remote sensing adaptive radiometric correction technology: First, a new shooting condition weight modulation is added to achieve dynamic adaptation of correction intensity; second, feature binding of the freeze-thaw stage is added to establish a dynamic correlation between target features and the freeze-thaw stage; and third, a micro-target spectral enhancement is designed, and a weak feature is amplified based on the attention mechanism of spectral difference. The three work together to transform the image features from disturbed dynamic features into stable and highly discriminative enhanced features, providing core support for semantic segmentation. This completely solves the core problem of spectral confusion caused by multiple shooting conditions and dynamic feature mismatch driven by freeze-thaw, while also improving the problem of weak features of micro-targets. This provides a feasible solution for introducing the adaptive radiometric correction technology of remote sensing from the mid-to-high altitude to low-altitude UAV monitoring.

[0050] In a preferred embodiment, step S4: Based on the background reference features determined by the dual-dimensional benchmark feature library and the spectral difference attention weights calculated during the spectral enhancement process, the enhanced feature map is subjected to confusion suppression processing to obtain cleaned features. The shallow, mid-level, and edge features of the preprocessed image are extracted and concatenated with the cleaned features across layers to obtain the fused feature map output from the decoding end. Specifically, this includes: S41: Based on the background reference features in the dual-dimensional baseline feature library and the spectral difference attention weights calculated during the spectral enhancement process, the enhanced feature map is subjected to confusion suppression processing to obtain the cleaned features. The specific expression is as follows: In the formula, It is a 5×5 Gaussian kernel. It is a 3×3 convolution kernel. As a background reference feature, The average gray level of the 3-channel background is represented by λ=0.35, which is the suppression weight. For background reference features in the two-dimensional benchmark feature library, As an anti-attention weight, the lower the spectral difference (the more easily confused the region), the higher the anti-attention weight and the stronger the suppression. The operation is a 3×3 convolution, which extracts the features of the confused region through convolution and then suppresses them through subtraction. This refers to the cleaned features output from the encoding end.

[0051] As is easily understood, this embodiment uses background reference features and spectral difference attention weights from a dual-dimensional benchmark feature library to perform obfuscation suppression on the enhanced feature map, eliminating redundant information highly correlated with the background and enhancing the purity of the target features. Thus, by using background features plus anti-attention weights, spectral obfuscation regions are accurately located, and redundant information in these regions is selectively eliminated. This avoids the blind, scene-insensitive nature of traditional obfuscation suppression, further improving the distinguishability between target features and the background, laying the foundation for accurate segmentation at the decoding end.

[0052] S42: Perform convolution operations of different lengths on the preprocessed image to extract shallow and mid-level features. Then, use the Canny edge detection algorithm to extract the edge features of the preprocessed image. The expression is: In the formula, It is a 3×3 convolution kernel. Extract low-level features such as shallow edges and textures; It is a 3×3 convolution kernel. Extract local contour features from the middle layer; This is a shallow feature map. This is a mid-level feature map. The Canny edge detection algorithm is used to extract image edge information; the edge information is then multiplied by 0.1 for intensity normalization to prevent edge features from overly strong and masking other features. This is an edge feature map.

[0053] S43: Upsample the shallow, mid-level, and edge features, concatenate them with the cleaned features output from the encoder, and then perform 1×1 convolution dimensionality reduction to obtain the final fused feature map from the decoder. The expression is: In the formula, For transposed convolution upsampling operations, the resolution of mid-layer, shallow-layer, and edge features is restored to the same level as... Consistent; For feature splicing operations, deep semantic features, mid-level contour features, shallow edge features, and specialized edge completion features are integrated; Using 1×1 convolution, the concatenated high-dimensional features are reduced to 256 channels; This is the final fused feature map at the decoding end.

[0054] As is easily understood, this embodiment extracts shallow and mid-level features by performing convolution operations of different lengths on the preprocessed image, extracts edge features by using the Canny edge detection algorithm, then upsamples the mid-level features, shallow features, and edge features, performs cross-layer feature concatenation with the cleaned features output from the encoding end, and finally obtains the final fused feature map from the decoding end through 1×1 convolution dimensionality reduction.

[0055] Therefore, after feature concatenation, deep enhanced features provide accurate semantic information, while mid- and shallow features provide rich detail information. The newly added edge features specifically compensate for the problem of blurred boundaries of small targets. Cross-layer fusion achieves comprehensive integration of semantic, detail, and boundary information, solving the contradiction of insufficient detail in deep features and unclear semantics in shallow features in traditional semantic segmentation. In particular, it significantly improves the boundary segmentation accuracy of small targets.

[0056] Based on this, in step S5, the pixel-level semantic segmentation mask expression is output as follows: In the formula, The function maps feature values ​​to a probability distribution, with 0.7 as the confidence threshold. This is the semantic segmentation mask for the output.

[0057] As is easily understood, this embodiment performs Softmax activation on the final fused feature map at the decoding end, calculates the probability that each pixel belongs to a risk target and the background, sets a confidence threshold, determines pixels with a probability higher than the threshold as risk targets, and determines pixels with a probability lower than the threshold as background, and outputs a pixel-level semantic segmentation mask. This avoids misjudging background noise as risk targets, ensures the accuracy of the segmentation mask, and provides a reliable pixel-level foundation for subsequent quantitative analysis of risk targets.

[0058] In this embodiment, the enhanced feature map is used as the core input of the encoder. A multi-condition spectral confusion suppression module is added to the encoder. The background features and attention weights of the dual-dimensional benchmark feature library are used to remove redundant information. Shallow, mid-level and edge features are extracted from the preprocessed image. At the decoding end, deep enhanced features, mid-shallow features and edge features are integrated by cross-layer feature concatenation and upsampling. After convolution dimensionality reduction, the pixel-level semantic segmentation mask is output through the Softmax function.

[0059] Therefore, an enhanced semantic segmentation architecture combining multi-conditional spectral confusion suppression, cross-layer feature fusion, and edge completion is proposed. Background features and attention weights from a dual-dimensional benchmark feature library are introduced into the encoding stage for confusion suppression, and a dedicated edge feature completion module is added at the decoding stage to achieve synergistic optimization of semantic cleansing, detail integration, and boundary enhancement. This solves the problems of semantic confusion, detail loss, and boundary blurring in traditional semantic segmentation in high-altitude permafrost regions. Confusion suppression at the encoding stage further cleanses target features, while cross-layer fusion and edge completion at the decoding stage accurately restore the contours of minute targets, ultimately outputting a high-precision segmentation mask, transforming the feature enhancement effects of the preceding steps into actual segmentation results.

[0060] In a preferred embodiment, step S6: Optimize the semantic segmentation mask using morphological opening operations, calculate quantitative indicators for permafrost cracks and degraded patches respectively, and generate ecological environment risk analysis results for the target permafrost region, specifically including: S61: Optimize the semantic segmentation mask by performing a morphological opening operation, the expression being: In the formula, Open the morphological operation function (erosion followed by dilation); It is a 2×2 rectangular structural element; The optimized mask is obtained by removing tiny noise points through erosion and then restoring the original shape of the target area through dilation, thus avoiding shrinkage of the target area.

[0061] In this step, a morphological opening operation is performed on the semantic segmentation mask using a 2×2 rectangular structuring element to remove isolated noise points in the mask, optimize mask quality, and ensure the integrity and purity of the risk target region. This embodiment utilizes erosion operations to eliminate tiny isolated noise, while preserving the integrity of the target region through dilation operations. This solves the problem of sporadic noise interference that may exist in the segmentation mask, ensuring that the object of subsequent quantitative analysis is the real risk target and improving the accuracy of the quantitative results.

[0062] S62: Based on the optimized mask, core quantitative indicators are calculated for permafrost cracks and vegetation degradation patches respectively, generating ecological and environmental risk analysis results including permafrost crack width, vegetation degradation patch area, and dynamic change rate. The expression is: In the formula, This refers to the width of the frozen soil crack. R represents the lengths of the two sides of the minimum bounding rectangle of the permafrost crack; R is the resolution of the aerial image. Area of ​​vegetation degradation patches R² is a pixel counting function that counts the number of target pixels with a value of 1 in the mask; R² is the square of the resolution. , Quantitative parameters for the same monitoring point at the two freeze-thaw stages k1 and k2. or ; This represents the rate of change.

[0063] Specifically, based on the optimized mask, core quantitative indicators are calculated for permafrost cracks and vegetation degradation patches respectively; the size parameters of permafrost cracks are obtained through the minimum bounding rectangle algorithm and converted into the true width using aerial photography resolution; the number of target pixels in vegetation degradation patches is counted by pixel counting and converted into the true area using resolution; and the dynamic change rate is calculated using quantitative parameters from different freeze-thaw stages at the same monitoring point. Thus, this embodiment achieves accurate conversion from pixel size to true geographic size based on pixel-level mask geometric feature extraction and counting, combined with aerial photography resolution; the dynamic change rate, based on the quantitative results of different freeze-thaw stages, intuitively reflects the impact of freeze-thaw cycles on risk targets, transforming image recognition results into quantitative data that can be directly used for ecological risk assessment.

[0064] In this embodiment, morphological opening is performed on the output semantic segmentation mask to remove isolated noise points. The length and width of permafrost cracks are extracted using the minimum bounding rectangle algorithm. The area of ​​vegetation degradation patches is calculated by combining pixel counting with UAV aerial photography resolution. The dynamic change rate of the target is calculated using the quantification results of different freeze-thaw stages at the same monitoring point, and a risk target quantification analysis report is output.

[0065] Therefore, by combining morphological optimization with specific quantitative methods for different types of risk targets, and introducing a dynamic change rate index for the freeze-thaw cycle, the semantic segmentation results obtained in the previous steps are transformed into quantitative data with practical ecological assessment value. Masking optimization ensures the accuracy of the quantitative basis, the classification quantitative method is adapted to the morphological characteristics of different risk targets, and the dynamic change rate index can accurately reflect the impact of the freeze-thaw cycle on the ecological environment, providing specific and comparable core indicators for ecological risk assessment.

[0066] Reference Figure 2 , Figure 2 This is a schematic diagram of the structure of the image recognition-based ecological environment risk monitoring and analysis system according to an embodiment of the present invention.

[0067] like Figure 2 As shown in the embodiments of the present invention, the ecological environment risk monitoring and analysis system based on image recognition includes: The information acquisition module 10 is used to acquire aerial monitoring information of the target permafrost area; wherein, the aerial monitoring information includes aerial images, aerial correlation parameters and surface temperature collected by the UAV during patrol monitoring; The feature library construction module 20 is used to divide the aerial images into several categories based on the aerial photography association parameters and the surface temperature, extract features from the aerial images of each category, and construct a two-dimensional benchmark feature library. The feature processing module 30 is used to perform preprocessing on the aerial image, including viewpoint correction, channel weight adjustment and filtering, according to the aerial image association parameters to obtain a preprocessed image, and to call the dual-dimensional benchmark feature library to perform enhancement processing on the preprocessed image, including weighted radiometric correction, freeze-thaw stage feature binding and spectral enhancement, to obtain an enhanced feature map. The fusion feature output module 40 is used to perform confusion suppression processing on the enhanced feature map based on the background reference features determined by the dual-dimensional benchmark feature library and the spectral difference attention weights calculated during the spectral enhancement process, to obtain the cleaned features, and to extract the shallow features, middle features and edge features of the preprocessed image and perform cross-layer feature concatenation with the cleaned features to obtain the fusion feature map output by the decoding end. The mask output module 50 is used to perform Softmax activation on the fused feature map, calculate the probability of each pixel belonging to the risk target and the background, determine the pixels with a probability higher than the threshold as risk targets, and determine the pixels with a probability lower than the threshold as background, and output a pixel-level semantic segmentation mask. The risk analysis module 60 is used to optimize the semantic segmentation mask by performing morphological opening operations, calculate quantitative indicators for permafrost cracks and degraded patches respectively, and generate ecological and environmental risk analysis results for the target permafrost area.

[0068] Other embodiments or specific implementations of the image recognition-based ecological environment risk monitoring and analysis system of the present invention can refer to the above-described method embodiments, and will not be repeated here.

[0069] It is understood that in the description of this specification, references to terms such as "one embodiment," "another embodiment," "other embodiments," or "first embodiment to Nth embodiment," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0070] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0071] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for monitoring and analyzing ecological and environmental risks based on image recognition, characterized in that, The method includes the following steps: S1: Obtain aerial monitoring information of the target permafrost area; wherein, the aerial monitoring information includes aerial images, aerial correlation parameters, and surface temperature collected by the UAV during patrol monitoring; S2: Based on the aerial photography correlation parameters and the surface temperature, the aerial images are divided into several categories, and features are extracted from the aerial images of each category to construct a two-dimensional benchmark feature library; S3: Based on the aerial photography association parameters, perform preprocessing on the aerial photography image, including viewpoint correction, channel weight adjustment and filtering, to obtain a preprocessed image. Call the dual-dimensional benchmark feature library to perform enhancement processing on the preprocessed image, including weighted radiometric correction, freeze-thaw stage feature binding and spectral enhancement, to obtain an enhanced feature map. S4: Based on the background reference features determined by the dual-dimensional benchmark feature library and the spectral difference attention weights calculated during the spectral enhancement process, the enhanced feature map is subjected to confusion suppression processing to obtain the cleaned features. The shallow features, middle features and edge features of the preprocessed image are extracted and cross-layer feature concatenation is performed with the cleaned features to obtain the fused feature map output by the decoding end. S5: Perform Softmax activation on the fused feature map, calculate the probability of each pixel belonging to the risk target and the background, determine the pixels with a probability higher than the threshold as risk targets, and determine the pixels with a probability lower than the threshold as background, and output the pixel-level semantic segmentation mask. S6: Optimize the semantic segmentation mask by performing morphological opening operation, calculate quantitative indicators for permafrost cracks and degraded patches respectively, and generate ecological and environmental risk analysis results for the target permafrost area.

2. The method for monitoring and analyzing ecological and environmental risks based on image recognition as described in claim 1, characterized in that, Step S1: Obtain aerial monitoring information of the target permafrost area, specifically including: S11: Control the drone to perform cruise monitoring at several preset monitoring points in the target permafrost area; S12: Acquire aerial images, shooting angle parameters, shooting time, and ground temperature collected by the camera equipment on the drone; S13: The shooting angle parameters and shooting time are summarized into aerial photography associated parameters, which, together with the aerial images and surface temperature, constitute the aerial monitoring information of the target permafrost area.

3. The method for monitoring and analyzing ecological and environmental risks based on image recognition as described in claim 2, characterized in that, In step S2, based on the aerial photography correlation parameters and the surface temperature, the aerial images are divided into several categories, including: S21: Based on the aerial photography correlation parameters and the surface temperature, generate 3 sets of scene division indicators; The scene division indicators include three freeze-thaw stages determined based on the range of surface temperature, three shooting periods determined based on the range of shooting time, and three shooting angle types determined based on shooting angle parameters and shooting time. S22: Based on the freeze-thaw stage, shooting time, and shooting angle of each aerial image, the aerial images collected during the patrol monitoring are divided into 27 categories.

4. The method for monitoring and analyzing ecological and environmental risks based on image recognition as described in claim 3, characterized in that, In step S2, features are extracted from aerial images of each category to construct a two-dimensional benchmark feature library, specifically including: S23: Select 200 sample images for each category and label the risk target areas; S24: Extract the average grayscale of the RGB channels of the risk target, the channel variance, the average grayscale of the RGB channels of the background, and the spectral difference between the target and the background for each category of the scene, and construct a two-dimensional benchmark feature library; The expression for the average grayscale of the RGB channels of the risk target is as follows: In the formula, Let be the number of sample images for the category scene corresponding to the shooting conditions at stage k and condition p, and set to 200. This represents the total number of pixels of the risky target in all sample images within this scene category. Let c be the set of risk target pixels in the c-th channel of the i-th image in this category of scene. Let be the grayscale value of the (x,y) pixel and c channel of the i-th image; This represents the average grayscale value of the c-th channel of the target in this category of scenarios, reflecting the basic spectral characteristics of the risk target. The specific expression for the channel variance of the risk objective is as follows: In the formula, This represents the gray-level variance of the c-th channel of the target in this category of scenarios, reflecting the stability of the target's risk spectral characteristics. The expression for the average grayscale of the RGB channels in the background is as follows: In the formula, For the first stage, Under the shooting conditions, the total number of pixels in the background area of ​​200 sample images; This is the set of background pixels in the c-th channel of the i-th image in this category of scene; For the background of this category of scenes Average gray level of the channel; The expression for the spectral difference between the target and the background is as follows: In the formula, The mean grayscale difference between the target and the background in the c-th channel reflects the spectral distinction between the target and the background.

5. The method for monitoring and analyzing ecological and environmental risks based on image recognition as described in claim 4, characterized in that, In step S3, the aerial image is preprocessed according to the aerial photography correlation parameters, including viewpoint correction, channel weight adjustment, and filtering, to obtain a preprocessed image. Specifically, this includes: S31: Based on the shooting angle parameter in the aerial photography association parameters, perform viewing angle correction on each pixel of the aerial image to correct the deviation of the ground surface spectral reflection path caused by the shooting angle; S32: Based on the shooting time corresponding to the shooting time in the aerial photography association parameters, call the preset RGB channel gain vector to adjust the channel weights of the image after the viewpoint correction; S33: Using the image after channel weight adjustment as the guide image, guided filtering is used to denoise the image to obtain the preprocessed image.

6. The method for monitoring and analyzing ecological and environmental risks based on image recognition as described in claim 5, characterized in that, In step S3, the dual-dimensional benchmark feature library is invoked to perform enhancement processing on the preprocessed image, including weighted radiometric correction, freeze-thaw stage feature binding, and spectral enhancement, to obtain an enhanced feature map, specifically including: S34: Calculate the angle weight and time period weight based on the shooting angle parameter and shooting time corresponding to the shooting time in the aerial photography association parameters, and determine the comprehensive shooting condition weight; S35: Call the average gray level and channel variance of the RGB channels in the dual-dimensional benchmark feature library, and combine them with the global mean and variance of the preprocessed image to correct the spectral deviation of the image through weighted radiometric correction; S36: Call the surface temperature from the aerial monitoring information, normalize the surface temperature, combine the freeze-thaw stage feature coefficients, perform feature modulation on the weighted radiometrically corrected image, and bind the freeze-thaw stage features. S37: Convolve the dynamic spectral features of the binding freeze-thaw stage to extract spectral features, call the spectral difference between the target and the background in the dual-dimensional benchmark feature library, construct attention weights to amplify the spectral difference between the small target and the background, and obtain the enhanced feature map after batch normalization.

7. The method for monitoring and analyzing ecological and environmental risks based on image recognition as described in claim 6, characterized in that, Step S4: Based on the background reference features determined by the dual-dimensional benchmark feature library and the spectral difference attention weights calculated during the spectral enhancement process, the enhanced feature map is subjected to confusion suppression processing to obtain cleaned features. The shallow, mid-level, and edge features of the preprocessed image are extracted and concatenated with the cleaned features across layers to obtain the fused feature map output from the decoding end. Specifically, this includes: S41: Based on the background reference features in the dual-dimensional baseline feature library and the spectral difference attention weights calculated during the spectral enhancement process, the enhanced feature map is subjected to confusion suppression processing to obtain the cleaned features. The specific expression is as follows: In the formula, It is a 5×5 Gaussian kernel. It is a 3×3 convolution kernel. As a background reference feature, The average gray level of the 3-channel background is represented by λ=0.35, which is the suppression weight. For background reference features in the two-dimensional benchmark feature library, Inverse attention weights; This is a 3×3 convolution operation; The cleaned features output from the encoding end; S42: Perform convolution operations of different lengths on the preprocessed image to extract shallow and mid-level features. Then, use the Canny edge detection algorithm to extract the edge features of the preprocessed image. The expression is: In the formula, It is a 3×3 convolution kernel. Extract low-level features such as shallow edges and textures; It is a 3×3 convolution kernel. Extract local contour features from the middle layer; This is a shallow feature map. This is a mid-level feature map. The Canny edge detection algorithm is used to extract image edge information. For edge feature maps; S43: Upsample the shallow, mid-level, and edge features, concatenate them with the cleaned features output from the encoder, and then perform 1×1 convolution dimensionality reduction to obtain the final fused feature map from the decoder. The expression is: In the formula, For transposed convolution upsampling operations, the resolution of mid-layer, shallow-layer, and edge features is restored to the same level as... Consistent; For feature splicing operations, deep semantic features, mid-level contour features, shallow edge features, and specialized edge completion features are integrated; It is a 1×1 convolution; This is the final fused feature map at the decoding end.

8. The method for monitoring and analyzing ecological and environmental risks based on image recognition as described in claim 7, characterized in that, In step S5, the pixel-level semantic segmentation mask expression is output as follows: In the formula, The function maps feature values ​​to a probability distribution, with 0.7 as the confidence threshold. This is the semantic segmentation mask for the output.

9. The method for monitoring and analyzing ecological and environmental risks based on image recognition as described in claim 1, characterized in that, Step S6: Optimize the semantic segmentation mask using morphological opening operations, calculate quantitative indicators for permafrost cracks and degraded patches respectively, and generate ecological and environmental risk analysis results for the target permafrost area, specifically including: S61: Optimize the semantic segmentation mask by performing morphological opening operation; S62: Based on the optimized mask, core quantitative indicators are calculated for permafrost cracks and vegetation degradation patches respectively, generating ecological and environmental risk analysis results including permafrost crack width, vegetation degradation patch area and dynamic change rate.

10. An ecological environment risk monitoring and analysis system based on image recognition, characterized in that, include: The information acquisition module is used to acquire aerial monitoring information of the target permafrost area; wherein, the aerial monitoring information includes aerial images, aerial correlation parameters and surface temperature collected by the UAV during patrol monitoring; The feature library construction module is used to divide the aerial images into several categories based on the aerial photography association parameters and the surface temperature, extract features from the aerial images of each category, and construct a two-dimensional benchmark feature library. The feature processing module is used to perform preprocessing on the aerial image based on the aerial image association parameters, including viewpoint correction, channel weight adjustment and filtering, to obtain a preprocessed image. The module then calls the two-dimensional benchmark feature library to perform enhancement processing on the preprocessed image, including weighted radiometric correction, freeze-thaw stage feature binding and spectral enhancement, to obtain an enhanced feature map. The fusion feature output module is used to perform confusion suppression processing on the enhanced feature map based on the background reference features determined by the dual-dimensional benchmark feature library and the spectral difference attention weights calculated during the spectral enhancement process, to obtain the cleaned features. The module also extracts the shallow features, middle features and edge features of the preprocessed image and performs cross-layer feature concatenation with the cleaned features to obtain the fusion feature map output by the decoding end. The mask output module is used to perform Softmax activation on the fused feature map, calculate the probability of each pixel belonging to the risk target and the background, determine the pixels with a probability higher than the threshold as risk targets, and determine the pixels with a probability lower than the threshold as background, and output a pixel-level semantic segmentation mask. The risk analysis module is used to optimize the semantic segmentation mask by performing morphological opening operations, calculate quantitative indicators for permafrost cracks and degraded patches respectively, and generate ecological and environmental risk analysis results for the target permafrost area.

Images