A micro-blog user depression risk identification method based on a large model and remote sensing environment perception

Abstract

The application discloses a microblog user depression risk identification method based on a large model and remote sensing environment perception, and aims to solve the defects of sparse features and ignoring environmental influence in the prior art. After preprocessing and word segmentation of microblog texts, a 13-class depression feature word set is constructed through a Transformer large model, and feature weights are calculated based on information values; user geographic positions are extracted, a 500-meter radius greening rate is calculated by using a GEE platform and Sentinel-2 satellite data, and an environmental adjustment coefficient is constructed; after a negative emotion index is corrected, risk grading is realized according to a normal distribution threshold. Experimental verification shows that the accuracy of the method is 92.05%, the recall rate is 94.12%, the text and environmental features are fused, the multidimensionality and ecological validity of identification are improved, and the user depression risk can be accurately identified.

Description

Technical Field

[0001] This invention belongs to the field of social media user psychological risk identification technology, specifically involving a method for identifying depression risk in Weibo users based on large models and remote sensing environmental perception. Background Technology

[0002] Detecting depression risk in users based on social media information such as Weibo can identify potential depression patients in a timely manner and effectively prevent risky behaviors. Current mainstream detection methods mostly build a text knowledge base based on Weibo posts by individuals with depression, and then use this knowledge base to match and detect the Weibo posts of other users, thereby achieving depression risk prediction.

[0003] However, existing technologies have significant shortcomings: On the one hand, the language expression habits of different social groups differ, resulting in limited transferability of detection models built based on specific groups, leading to low accuracy in predicting new users and users with fewer Weibo posts. The fundamental reason is that existing methods do not effectively sort out parts of speech when constructing sentiment dictionaries, resulting in sparse feature words and susceptibility to interference from emotional and extreme words, affecting the accuracy of judgment. On the other hand, environmental psychology research has confirmed that exposure to the natural environment is closely related to mental health; environments with high green coverage can relieve stress, improve mood, and reduce the risk of depression. However, current methods for identifying depression risk on social platforms ignore this key variable, further restricting the improvement of recognition accuracy. Summary of the Invention

[0004] This invention aims to address the technical problems existing in current methods for identifying depression risk among Weibo users, such as sparse feature words, susceptibility to interference from emotional and extreme words, weak model transferability, and failure to consider the influence of environmental factors. It provides a depression risk identification method that takes into account both textual and environmental features, has high recognition accuracy, and is robust.

[0005] To achieve the above objectives, the present invention employs the following solution:

[0006] A method for identifying the depression risk of Weibo users based on large-scale models and remote sensing environmental perception includes the following steps:

[0007] Step S1: Weibo text preprocessing and word segmentation: obtaining the U value of the depression patient group. dep and ordinary user groups U norm We obtained a sample of Weibo text and preprocessed it.

[0008] Step S2: Semantic expansion and classification system construction of depression feature words: The preprocessed Weibo text is encoded using a large language model based on the Transformer structure. The model attention mechanism is used for feature fusion and clustering to construct a set of depression feature words V containing 13 semantic categories.

[0009] Step S3: Feature word frequency and proportion statistics: Calculate the frequency of each of the 13 feature words for the two groups, as well as the proportion and average proportion of each feature word in the total vocabulary of the corresponding group;

[0010] Step S4: Feature Discriminative Quantification and Weight Calculation: Based on the Information Value (IV) formula, the discriminative power of each type of feature word in the two groups is quantified. A scaling factor α is introduced for smoothing to obtain the final weight of each type of feature word. i ( v );

[0011] Step S5: Environmental Factor Extraction and Adjustment Coefficient Construction: Extract geographic location information from user Weibo data and convert it into standard latitude and longitude coordinates. Calculate the greening rate GR within a set radius R centered on the user coordinates based on the Normalized Difference Vegetation Index (NDVI), and calculate the environmental adjustment coefficient α. env ;

[0012] Step S6: Individual Depression Risk Prediction and Classification: Repeat the word segmentation process of step S1 on the target user's Weibo, complete the semantic matching and classification of feature words through a large language model, calculate the weighted depression words and the original negative emotion index, obtain the adjusted negative emotion index after correction by the environmental adjustment coefficient, set the risk classification standard based on the normal distribution, and realize the depression risk classification of the target user.

[0013] Further optimization, step S1 specifically includes:

[0014] Step S1.1: Sample Acquisition: Clearly define the patient group U of depression dep and ordinary user groups U norm We collected Weibo posts from two groups as training samples, while reserving Weibo posts from the target user as samples to be predicted.

[0015] Step S1.2: Stop word filtering: Using a Chinese stop word library containing more than 4,900 irrelevant words, the text of the Weibo post is traversed by text matching to automatically filter irrelevant information such as grammatical function words, logical connectors, personal pronouns, time-related expressions, and number-related expressions, thus completing text cleaning and standardization.

[0016] Step S1.3: Word segmentation: First, the maximum probability path search algorithm based on prefix dictionary and dynamic programming is used to traverse the cleaned text and determine the preliminary word segmentation results based on the dictionary vocabulary and word occurrence probability; then, through Hidden Markov Model (HMM), based on the preset state transition probability matrix, observation probability matrix and initial probability matrix, state reasoning and recognition are performed on words not included in the prefix dictionary in the preliminary word segmentation, and semantically complete word sequences are output.

[0017] Step S1.4: Total word frequency statistics: Traverse the word sequences of the patient group with depression and the general user group respectively, and accumulate the occurrence frequency of all words in both groups to obtain the total word frequency T of the patient group with depression. dep Total word frequency T for ordinary users norm .

[0018] Further optimization, step S2 specifically includes:

[0019] Step S2.1: Text Encoding: Using a large language model based on the Transformer structure, the preprocessed Weibo text is segmented into a token sequence according to the model input format. Through the model's embedding layer, self-attention layer, and feedforward neural network layer, a high-dimensional semantic vector for each token is generated. This vector contains the contextual semantic information and sentiment features of the words.

[0020] Step S2.2: Feature Fusion and Clustering: Utilize the self-attention mechanism of the large language model to calculate the cosine similarity between the semantic vectors of each token, focus on key tokens with similarity higher than a preset threshold and related to depression, such as "self-denial", "anxiety", and "loneliness", and perform preliminary clustering of key tokens using the K-means clustering algorithm.

[0021] Step S2.3: Classification System Construction: Based on the three main criteria, the preliminary clustering results are optimized and adjusted to construct a depressive feature word set V containing 13 semantic categories.

[0022] The classification process integrates three main bases: ① Depression-related knowledge from psychological research, such as emotional reasoning and attribution of blame; ② Typical depressive symptom dimensions in clinical diagnostic criteria, including affective symptoms, cognitive symptoms, and somatic symptoms; ③ Linguistic style analysis of depressive expressions on social media, such as the use of absolute terms and first-person pronouns.

[0023] The final 13 categories of characteristic words are: pain / uncomfort / torment, sadness / grief / loss, anger / irritability / annoyance, anxiety / worry / tension, strong negative exclamations / curses, fatigue / weakness / drowsiness, disgust / aversion / disdain, fear / fear / panic, self-denial / inferiority / shame, loneliness / alienation / isolation, regret / regret / self-blame, aversion to interpersonal relationships / social anxiety, and other negative expressions that are difficult to classify.

[0024] Step S2.4: New word adaptation: For new words not included in the depression feature word set, the similarity between their semantic vector and the central vector of the 13 categories of depression feature word set is calculated by the large language model. The new words are then assigned to the category with the highest similarity, thus realizing the semantic expansion of the feature words.

[0025] Further optimization, step S3 specifically includes:

[0026] Step S3.1: Calculate the frequency of feature words: Traverse the depression patient group U dep For each user u, the word sequence is analyzed, and the word sequence of each user u in each of the 13 feature word sets V is counted. v count of the number of times the feature words appear u ( v The depression patient population was calculated according to the following formula (1) in category v Feature word frequency f dep ( v ):

[0027] (1)

[0028] Using the same method, iterate through the ordinary user group U. norm The word sequence is calculated according to the following formula (2) for the general user group in the category. v Feature word frequency f norm ( v ):

[0029] (2)

[0030] Step S3.2: Calculate the average proportion of feature words: Calculate the average proportion of depressive feature words in the patient group by applying the following formula (3) to all categories of feature words in the feature word set:

[0031] (3)

[0032] Using the same method, the average proportion of depressive characteristic words in the general user group was calculated according to the following formula (4):

[0033] (4)

[0034] Step S3.3: Calculate the specific proportion of feature words: Calculate the proportion of the first feature word according to the following formula (5). v The specific percentage of category-specific words in the total vocabulary of the depression patient population P dep ( v ):

[0035] (5)

[0036] Using the same method, calculate the categories in the ordinary user group according to the following formula (6). v The specific percentage of feature words P norm (v):

[0037] (6)

[0038] Further optimization, step S4 specifically includes:

[0039] Step S4.1: Calculate the information value IV: Use the P obtained in step S3 dep (v) and P norm (v) Substitute into the following formula (7):

[0040] (7)

[0041] Where, when P dep (v) or P norm When (v) is 0, a minimum value ε is added, and ε = 10 is taken. -8 That is, at this time, P is given dep (v) or P norm (v) Assigning a value of ε avoids the logarithm being meaningless; the IV value is used to quantify categories. v The distribution differences of feature words in the two groups show that the higher the IV value, the stronger the discrimination.

[0042] Step S4.2: Weight Smoothing and Determination: Introduce a scaling factor α to obtain the final weights for each type of feature word. i ( v )=IV( v )×α, this weight directly reflects the category v The contribution of α to the identification of depression risk ranges from 50 to 350.

[0043] Further optimization, step S5 specifically includes:

[0044] Step S5.1: Geographic location information extraction: Extract geographic location information from user Weibo data through Weibo's open interface or text semantic recognition algorithm, including location data recorded by Weibo check-in function, geographic location tags actively marked by users, or location descriptions identified from text content, such as city names, community names, landmark building names, etc.

[0045] Step S5.2: Latitude and longitude conversion: Use geocoding technology (such as Gaode Geocoding API, Baidu Geocoding API, Google Geocoding Service, etc.) to convert the extracted geographic location information into standard latitude and longitude coordinates in the WGS-84 coordinate system.

[0046] Step S5.3: Greening Rate Calculation: Using the Google Earth Engine (GEE) remote sensing data analysis platform, the Level-2A image data of the Sentinel-2 satellite is called, and a circular analysis area with a radius of 500 meters is constructed with the user's latitude and longitude coordinates as the center; based on the satellite image of this area, the Normalized Difference Vegetation Index (NDVI) is calculated using the following formula (8):

[0047] NDVI=(ρ NIR -ρ Red ) / (ρ NIR +ρ Red (8)

[0048] Where, ρ NIR ρ is the reflectance in the near-infrared band. Red The reflectance is defined as the red light band reflectance. Pixels with NDVI > 0.2 are defined as effective vegetation pixels. The ratio of the number of effective vegetation pixels to the total number of pixels in the analysis area is calculated to obtain the greening rate GR around the user.

[0049] Step S5.4: Construction of environmental adjustment coefficient: Statistically calculate the greening rate (GR) of all users in the training sample, and take the median value as the baseline greening rate (GR). baseline Set an empirical adjustment weight parameter β and substitute it into formula (9) to calculate the environmental adjustment coefficient;

[0050] α env =1-β×(GR-GR baseline (9)

[0051] The β value is an empirical value ranging from 0.5 to 1, used to control the adjustment strength of the greening rate on risk assessment. When user geographic location information cannot be obtained, a missing adjustment coefficient is used to ensure stable model operation.

[0052] Further optimization, step S6 specifically includes:

[0053] Step S6.1: Target user text processing: Repeat the stop word filtering and word segmentation process of step S1 for the target user's Weibo text to obtain the target user's word sequence and count the total word frequency N of the target user's Weibo.

[0054] Step S6.2: Feature Matching and Word Frequency Statistics: The large language model from Step S2 is used to encode each word in the target user's word sequence, generating a semantic vector; this semantic vector is then compared with each of the 13 feature word sets V. v The cosine similarity of the center vectors is used to classify words with similarity scores higher than a preset threshold (threshold = 0.7) into their corresponding categories. v Statistical analysis of target users in each category v The number of times the feature words appear on the screen (count) v ).

[0055] Step S6.3: Calculate the negative sentiment index: Calculate the final weights of each type of feature word obtained in Step S4. i ( v ) and the count of the corresponding categories of the target users vThe weighted frequency of depressive words for the target user is calculated according to the following formula (10). W :

[0056] (10)

[0057] The original negative sentiment index (NEI) is calculated according to formula (11). original The adjusted negative sentiment index (NEI) was calculated according to formula (12). adjusted ;

[0058] NEI original =W / N; (11)

[0059] NEI adjusted =NEI original ×α env (12)

[0060] Step S6.4: Risk Classification: Set the Adjusted Negative Sentiment Index (NEI) for all users. adjusted Following a normal distribution, calculate the NEI of all users in the training samples. adjusted The mean μ and standard deviation σ of the target user are calculated, and the depression risk level of the target user is output according to the following risk classification criteria.

[0061] When NEI adjusted When the value is less than μ-0.5σ, the individual is considered to be in the normal population.

[0062] When μ - 0.5σ ≤ NEI adjusted When the value is less than μ + 1.5σ, it is considered a low risk of mild depression.

[0063] When μ + 1.5σ ≤ NEI adjusted When the value is less than μ + 2.5σ, it is considered a moderate risk of depression.

[0064] When NEI adjusted A score ≥μ+2.5σ indicates a risk of major depressive disorder.

[0065] Compared with the prior art, the present invention has the following beneficial effects:

[0066] 1. This invention constructs a depression feature word classification system covering 13 semantic categories. Combining clinical diagnostic criteria and psychological research, it improves the detail and comprehensiveness of the feature word system and effectively solves the problem of feature word sparsity.

[0067] 2. By using a large model based on the Transformer structure for semantic clustering and new word recognition, the model's generalization ability is relied upon to accurately classify new words that were not included in the feature words, thus overcoming the problem of insufficient model transferability caused by differences in the expression habits of different users.

[0068] 3. By adopting a weighted mechanism based on information values ​​and dynamic thresholds, the distinguishability of feature words is optimized, the interference of emotion words and extreme words is reduced, and the accuracy of judgment is improved.

[0069] 4. By introducing remote sensing environmental perception technology, the greening rate is integrated into the risk assessment as an environmental regulation variable, realizing the multimodal fusion of text features and environmental features, and improving the multidimensionality and ecological validity of the identification. Attached Figure Description

[0070] Figure 1 This is a flowchart of the Weibo user depression risk identification method based on large model and remote sensing environment perception described in this invention;

[0071] Figure 2 This is a confusion matrix diagram of the test data for this invention. Detailed Implementation

[0072] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0073] I. Preparation of Experimental Data:

[0074] In this embodiment, 123,013 Weibo posts from 228 users were legally obtained as training data. Among them, 76 users were patients with depression, who posted 37,583 Weibo posts; and 152 users were ordinary users, who posted 85,430 Weibo posts. Another 52 users with depression (10,693 Weibo posts) and 36 ordinary users (31,329 Weibo posts) were selected as test data to verify the effectiveness of the model.

[0075] II. Figure 1 As shown, the specific implementation steps of the Weibo user depression risk identification method based on large model and remote sensing environmental perception are as follows:

[0076] Step S1: Weibo Text Preprocessing and Word Segmentation: The training and test data are cleaned using a Chinese stop word database containing over 4900 words. After removing irrelevant words, preliminary word segmentation is performed using a prefix dictionary and dynamic programming algorithm. Out-of-vocabulary words are identified using an HMM model. The total word frequency of depressed patients and ordinary users in the training set is statistically analyzed to obtain T. dep The value is 228438, T norm The value is 679139.

[0077] Step S2, Feature word semantic expansion and classification system construction: The training data text is encoded using a Transformer-based large language model, and 13 categories of depression feature words are obtained through clustering via an attention mechanism, as described in the invention content.

[0078] Step S3, Feature word frequency and proportion statistics: Based on the training data, calculate f according to the aforementioned formula. norm (v), f dep (v), p norm p dep P norm (v) and P dep (v), the results are shown in Table 1.

[0079] Table 1. Calculation results of frequency and proportion of 13 types of feature words

[0080]

[0081] Step S4, Feature Discriminant Calculation and Weight Calculation: Based on P norm (v) and P dep (v) Calculate the IV value for each category (values ​​below 1 are uniformly calculated as 1), set the scaling factor α=250, and obtain the final weight i(v) for each category as shown in Table 2.

[0082] Table 2. Final weight calculation results for 13 categories of feature words

[0083]

[0084] Step S5, Environmental Factor Extraction and Greening Rate Calculation: Extract geographic location information from user Weibo posts in the training data, convert it into latitude and longitude coordinates, and use the Google Earth Engine platform to call Sentinel-2 satellite image data to calculate the greening rate GR within a 500-meter radius. The baseline greening rate GR_baseline=0.35 is obtained. Set the empirical adjustment weight parameter β=1.5 and construct an environmental adjustment coefficient calculation model.

[0085] Step S6, Model Training and Validation: Based on the feature word weights, environmental adjustment coefficient model, and risk classification criteria obtained from the training data, predict the depression risk of users in the test data. For each user in the test data, repeat the steps of word segmentation, feature extraction, greening rate calculation, and NEIadjusted calculation, and determine their risk level according to the risk classification criteria.

[0086] III. Experimental Results: Test data verification results show that 48 users with depression were correctly predicted, while 33 users with normal depression were correctly predicted. Figure 2As shown, the confusion matrix verification indicates that the prediction accuracy of this method is 92.05%, the recall is 94.12%, and the harmonic mean F1 score of accuracy and recall is 93.20%, which is significantly better than traditional text-based recognition methods, verifying the effectiveness and superiority of the present invention.

[0087] Based on the above-described preferred embodiments of the present invention, and through the foregoing description, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. A method for identifying the depression risk of Weibo users based on a large model and remote sensing environmental perception, characterized in that, Includes the following steps: Step S1: Weibo text preprocessing and word segmentation: obtaining the U value of the depression patient group. dep and ordinary user groups U norm We obtained a sample of Weibo text and preprocessed it; specifically... include: Step S1.1: Sample Acquisition: Clearly define the patient group U of depression dep and ordinary user groups U norm We collected Weibo posts from two groups as training samples, and reserved Weibo posts from the target user as samples to be predicted. Step S1.2: Stop word filtering: Using a Chinese stop word library containing several irrelevant words, the text of the Weibo post is traversed through text matching to automatically filter out grammatical function words, logical connectors, personal pronouns, time-related expressions, and number-related expressions, thus completing text cleaning and standardization; Step S1.3: Word segmentation: First, the maximum probability path search algorithm based on prefix dictionary and dynamic programming is used to traverse the cleaned text and determine the preliminary word segmentation results based on the dictionary vocabulary and word occurrence probability; then, through Hidden Markov Model (HMM), based on the preset state transition probability matrix, observation probability matrix and initial probability matrix, state reasoning and recognition are performed on words not included in the prefix dictionary in the preliminary word segmentation, and semantically complete word sequences are output. Step S1.4: Total word frequency statistics: Iterate through the word sequences of the depression patient group and the general user group respectively, and accumulate the occurrence times of all words in both groups to obtain the corresponding total word frequency, denoted as T. dep T norm； Step S2: Semantic expansion and classification system construction of depression feature words: The preprocessed Weibo text is encoded using a large language model based on the Transformer structure. The model attention mechanism is used for feature fusion and clustering to construct a set of depression feature words V containing 13 semantic categories. Step S3: Feature word frequency and proportion statistics: Calculate the feature word frequency of each of the 13 categories of depression feature words in the two groups, as well as the proportion and average proportion of each type of feature word in the total vocabulary of the corresponding group; Step S4: Feature Discriminative Quantification and Weight Calculation: Based on the Information Value (IV) calculation formula, the discriminative power of each type of feature word in the two groups is quantified. A scaling factor α is introduced for smoothing to obtain the final weight of each type of feature word. i ( v ); Step S5: Environmental Factor Extraction and Adjustment Coefficient Construction: Extract geographic location information from user Weibo data and convert it into standard latitude and longitude coordinates. Calculate the greening rate GR within a set radius R centered on the user coordinates based on the Normalized Difference Vegetation Index (NDVI), and calculate the environmental adjustment coefficient α. env ; Step S6: Individual Depression Risk Prediction and Classification: Repeat the word segmentation process of step S1 on the target user's Weibo, complete the semantic matching and classification of feature words through a large language model, calculate the weighted depression words and the original negative emotion index, obtain the adjusted negative emotion index after correction by the environmental adjustment coefficient, set the risk classification standard based on the normal distribution, and realize the depression risk classification of the target user.

2. The method for identifying the depression risk of Weibo users based on a large model and remote sensing environmental perception as described in claim 1, characterized in that, Step S2 specifically includes: Step S2.1: Text Encoding: Using a large language model based on the Transformer structure, the preprocessed Weibo text is segmented into a token sequence according to the model input format. Through the model's embedding layer, self-attention layer, and feedforward neural network layer, a high-dimensional semantic vector for each token is generated. This vector contains the contextual semantic information and sentiment features of the words. Step S2.2: Feature fusion and clustering: Utilize the self-attention mechanism of the large language model to calculate the cosine similarity between the semantic vectors of each token, focus on key tokens with similarity higher than a preset threshold and related to depression, and perform preliminary clustering of key tokens using the K-means clustering algorithm; Step S2.3: Classification System Construction: Combining psychological knowledge related to depression, clinical diagnostic criteria for depression, and analysis of depressive expression styles on social media, a depressive feature word set V containing 13 semantic categories is constructed: The 13 categories of characteristic words are: pain / uncomfort / torment, sadness / grief / loss, anger / irritability / annoyance, anxiety / worry / tension, strong negative exclamations / curses, fatigue / weakness / drowsiness, disgust / aversion / disdain, fear / fear / panic, self-denial / inferiority / shame, loneliness / alienation / isolation, regret / regret / self-blame, aversion to interpersonal relationships / social anxiety, and other negative expressions that are difficult to classify. Step S2.4: New word adaptation: For new words not included in the depression feature word set, the similarity between their semantic vector and the central vector of the 13 categories of depression feature word set is calculated by the large language model. The new words are then assigned to the category with the highest similarity, thus realizing the semantic expansion of the feature words.

3. The method for identifying the depression risk of Weibo users based on a large model and remote sensing environmental perception according to claim 2, characterized in that, Step S3 specifically includes: Step S3.1: Calculate the frequency of feature words: Traverse the depression patient group U dep For each user u, the word sequence is analyzed, and the position of each user u in the 13 feature word sets is counted. v count of occurrences of class feature words u ( v The depression patient population was calculated according to the following formula (1) in category v Feature word frequency f dep ( v ): ；（1） Using the same method, iterate through the ordinary user group U. norm The word sequence is calculated according to the following formula (2) for the general user group in the category. v Feature word frequency f norm ( v ): ；（2） Step S3.2: Calculate the average proportion of feature words: Calculate the average proportion of depressive feature words in the patient group for all categories of feature words in the feature word set according to the following formula (3): ；（3） Using the same method, the average proportion of depressive characteristic words in the general user group was calculated according to the following formula (4): ；（4） Step S3.3: Calculate the specific proportion of feature words: Calculate the proportion of the first feature word according to the following formula (5). v The specific percentage of category-specific words in the total vocabulary of the depression patient population P dep ( v ): ；（5） Using the same method, calculate the categories in the ordinary user group according to the following formula (6). v The specific percentage of feature words P norm (v): ；（6） Among them, T dep The total word frequency of the patient group with depression; T norm This represents the total word frequency of the general user group.

4. The method for identifying the depression risk of Weibo users based on a large model and remote sensing environmental perception according to claim 3, characterized in that, Step S4 specifically includes: Step S4.1: Calculate the information value IV: Use the P obtained in step S3 dep (v) and P norm (v) Substitute into the following formula (7): ；（7） Where, when P dep (v) or P norm When (v) is 0, the logarithm is avoided by adding a minimum value ε; the IV value is used to quantify the categories. v The distribution differences of feature words between the two groups show that the higher the IV value, the stronger the discrimination. Step S4.2: Weight Smoothing and Determination: Introduce a scaling factor α to obtain the final weights for each type of feature word. i ( v )=IV( v )×α.

5. The method for identifying the depression risk of Weibo users based on a large model and remote sensing environmental perception according to claim 4, characterized in that, Step S5 specifically includes: Step S5.1: Geographic location information extraction: Extract geographic location information from user Weibo data through Weibo's open interface or text semantic recognition algorithm, including location data recorded by Weibo check-in function, geographic location tags actively marked by users, or location descriptions identified from text content; Step S5.2: Latitude and longitude conversion: Using geocoding technology, the extracted geographic location information is converted into standard latitude and longitude coordinates in the WGS-84 coordinate system; Step S5.3: Greening Rate Calculation: Using the Google Earth Engine remote sensing data analysis platform, the Level-2A image data of the Sentinel-2 satellite is called, and a circular analysis area with a radius of 500 meters is constructed with the user's latitude and longitude coordinates as the center; based on the satellite image of this area, the Normalized Difference Vegetation Index (NDVI) is calculated using formula (8): NDVI=(r NIR -r Red ) / (ρ NIR +r Red );(8) Where, ρ NIR ρ is the reflectance in the near-infrared band. Red The reflectance is defined as the red light band reflectance. Pixels with NDVI > 0.2 are defined as effective vegetation pixels. The ratio of the number of effective vegetation pixels to the total number of pixels in the analysis area is calculated to obtain the greening rate GR around the user. Step S5.4: Construction of environmental adjustment coefficient: Statistically calculate the greening rate (GR) of all users in the training sample, and take the median value as the baseline greening rate (GR). baseline Set an empirical adjustment weight parameter β and substitute it into formula (9) to calculate the environmental adjustment coefficient; when the user's geographical location information cannot be extracted, the preset missing adjustment coefficient is automatically called; α env =1-β×(GR-GR baseline );(9).

6. The method for identifying the depression risk of Weibo users based on a large model and remote sensing environmental perception according to claim 5, characterized in that, Step S6 specifically includes: Step S6.1: Target user text processing: Repeat the stop word filtering and word segmentation process of step S1 for the target user's Weibo text to obtain the target user's word sequence and count the total word frequency N of the target user's Weibo. Step S6.2: Feature Matching and Word Frequency Statistics: The large language model from Step S2 is used to encode each word in the target user's word sequence, generating a semantic vector; this semantic vector is then compared with each of the 13 feature word sets V. v The cosine similarity of the center vectors is used to classify words with similarity higher than a preset threshold into their corresponding categories. v Statistical analysis of target users in each category v The number of times the feature words appear on the screen (count) v ); Step S6.3: Calculate the negative sentiment index: Calculate the final weights of each type of feature word obtained in Step S4. i ( v ) and the count of the corresponding categories of the target users v The weighted frequency of depressive words for the target user is calculated according to the following formula (10). W : ；（10） The original negative sentiment index (NEI) is calculated according to formula (11). original The adjusted negative sentiment index (NEI) was calculated according to formula (12). adjusted ; NOW original =W / N;(11) NEI adjusted =NEI original ×α env (12) Step S6.4: Risk Classification: Set the Adjusted Negative Sentiment Index (NEI) for all users. adjusted Following a normal distribution, calculate the NEI of all users in the training samples. adjusted The mean μ and standard deviation σ of the target user are calculated, and the depression risk level of the target user is output according to the following risk classification criteria. This NEI adjusted <μ-0.5σ, normal group; When μ - 0.5σ ≤ NEI adjusted When the value is less than μ + 1.5σ, it is considered a low risk of mild depression. When μ + 1.5σ ≤ NEI adjusted When the value is less than μ + 2.5σ, it is considered a moderate risk of depression. When NEI adjusted A score ≥μ+2.5σ indicates a risk of major depressive disorder.

Images