A weather recognition and short-time prediction method based on multi-vision feature fusion and time sequence analysis

Abstract

The application discloses a weather recognition and short-time prediction method based on multi-vision feature fusion and time sequence analysis, and belongs to the technical field of computer vision intelligent meteorological prediction. The application realizes high-precision recognition of six typical outdoor weathers and 0-15 minute short-time weather trend prediction through image self-adaptive preprocessing, multi-dimensional vision physical feature extraction, feature threshold fusion discrimination and time sequence frame difference analysis. The application breaks through the technical bottleneck of traditional single-frame recognition, no prediction capability and high misjudgment rate in complex scenes, does not need professional meteorological sensors, and can complete intelligent meteorological monitoring and prediction by relying on ordinary vision equipment. The algorithm is light in weight, strong in robustness, low in deployment cost, suitable for various outdoor intelligent monitoring, smart home and traffic meteorological auxiliary scenes, and has extremely high engineering application and industrialization value.

Description

Technical Field

[0001] This invention relates to the fields of computer vision, deep learning image processing, and intelligent weather forecasting, specifically to an intelligent weather recognition and short-term trend forecasting method based on multi-visual feature fusion and temporal frame difference analysis. Background Technology

[0002] Most existing weather image recognition technologies rely on single deep learning classification or single color feature discrimination, which have significant drawbacks: 1. Single-frame image recognition cannot capture dynamic weather changes, can only achieve static recognition, and lacks predictive capabilities; 2. It relies solely on AI models to extract features, lacking traditional visual physical feature constraints, resulting in a high misidentification rate in complex scenes such as backlight, fog, rain, and snow; 3. It does not specifically extract texture features for rain, snow, fog, and clouds, resulting in low feature differentiation between different weather types; 4. It lacks an adaptive image enhancement mechanism, making recognition functions prone to failure in backlight, low light, and fog scenes.

[0003] Currently, there is no complete technical solution that combines traditional visual physical features, lightweight AI models, and temporal difference prediction. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a weather identification and short-term forecasting method based on multi-visual feature fusion and temporal analysis, so as to achieve high-precision weather classification and short-term dynamic forecasting.

[0005] This invention comprises four core innovations: 1. Multi-dimensional physical visual feature fusion technology: Simultaneously extracting three physical features—mean brightness, texture variance, and white pixel ratio—to accurately distinguish between six weather types: sunny, cloudy, rainy, foggy, snowy, and nighttime, avoiding the poor generalization ability of pure AI models; 2. Adaptive CLAHE image enhancement preprocessing algorithm: Performing adaptive brightness equalization and noise reduction for outdoor strong light, backlight, and low light scenes, significantly improving the system's robustness under complex conditions; 3. Temporal frame difference short-term prediction algorithm: By caching continuous multi-frame recognition results and combining segmented mean difference analysis, it enables short-term trend prediction of weather changes such as clearing, cloudy, rainfall intensity, fog, and snowfall duration; 4. Lightweight AI + traditional feature dual-fusion discrimination mechanism: Using traditional visual physical features as the discrimination benchmark and employing a lightweight MobileNetV2 model as an auxiliary correction method, it balances recognition accuracy and real-time performance, and can run offline on edge devices. Attached Figure Description

[0006] This application includes five accompanying drawings to explain the technical solution and implementation process of the present invention:

[0007] Figure 1: Flowchart of the overall method architecture of the present invention;

[0008] Figure 2: Flowchart of the image adaptive preprocessing steps of the present invention;

[0009] Figure 3: Block diagram of parallel extraction logic for multiple visual features in this invention;

[0010] Figure 4: Flowchart of the temporal difference short-time prediction algorithm of the present invention;

[0011] Figure 5: Schematic diagram of the system operation interface of the present invention. Detailed Implementation

[0012] The present invention will be further described in detail with reference to the accompanying drawings.

[0013] As shown in Figure 1, a weather identification and short-term forecasting method based on multi-visual feature fusion and temporal analysis includes the following steps:

[0014] Step S1, Image Acquisition: Acquire real-time outdoor video streams using camera equipment, or read locally stored outdoor image data;

[0015] Step S2, Adaptive Image Preprocessing: As shown in Figure 2, Gaussian denoising and CLAHE local brightness adaptive equalization are performed on the original image in sequence to eliminate the interference caused by backlight, dark light and fog noise. Finally, image size unification and data normalization are completed.

[0016] Step S3, Multi-visual feature extraction: As shown in Figure 3, the preprocessed image is converted to the HSV color space, and the overall brightness mean of the image, the texture variance obtained by the Laplacian operator, and the proportion of high and low brightness white pixels are calculated respectively. A three-dimensional weather feature vector is constructed based on the three types of data.

[0017] Step S4, Multi-feature Fusion Weather Judgment: Based on the threshold range corresponding to the three-dimensional feature vector, match the corresponding weather category: low brightness is judged as night, high white pixel ratio + high texture variance is judged as snowy, high texture variance + medium white pixel ratio is judged as rainy, medium-low brightness + low texture variance is judged as cloudy, high brightness + low texture variance is judged as sunny, and medium-low brightness + medium white pixel ratio is judged as foggy.

[0018] Step S5, Time-series data caching: Establish a fixed-length frame buffer queue to continuously store real-time weather identification results for multiple consecutive frames;

[0019] Step S6, Time Series Differential Trend Prediction: As shown in Figure 4, the data in the cache queue is divided into a front-end time series dataset and a back-end time series dataset. The average value of the two datasets is calculated respectively. The weather evolution trend is determined by the difference between the front and back average values, and the prediction results such as weather continuation, intensification, weakening, clearing up, and turning cloudy are output.

[0020] Step S7, Result Fusion Output: Integrate the three types of data: current weather category, recognition confidence level, and short-term forecast trend, as shown in Figure 5. The results are displayed through a visualization interface, and the recognition log is saved at the same time.

[0021] 6. Beneficial effects

[0022] Compared with existing technologies, this invention has the following advantages: it adopts a pure vision solution, requiring no professional meteorological sensors and resulting in extremely low deployment costs; it relies on multiple physical features for constraints, achieving a significantly higher recognition accuracy than a single AI model; it features a unique time-series analysis logic, enabling short-term weather forecasting; the overall algorithm has a lightweight design, supporting offline deployment on embedded devices; and it is adaptable to various complex outdoor lighting and weather scenarios, demonstrating strong practicality and broad application prospects.

Claims

1. A weather identification and short-term forecasting method based on multi-visual feature fusion and temporal analysis, characterized in that, Includes the following steps: S1. Acquire outdoor environmental images or real-time video stream data; S2. Perform Gaussian denoising and CLAHE adaptive brightness equalization preprocessing on the acquired image to complete image size normalization and format conversion; S3. Extract multi-dimensional visual features from the preprocessed image, including the image brightness mean, texture variance, and white pixel ratio. S4. Based on the threshold range of multi-dimensional visual features, the current outdoor weather type is determined by fusion. S5. Establish a time-series frame buffer queue to continuously store multiple frames of weather recognition results; S6. Perform segmented mean-difference operations on the cached time-series data to analyze the short-term weather evolution trend; S7. Output the current weather category, identification confidence level, and short-term weather forecast results.

2. The method according to claim 1, characterized in that: In step S3, brightness information is extracted using the HSV color space, the complexity of image texture is calculated using the Laplacian operator, and the proportion of white pixels in fog and snow is calculated using color thresholding.

3. The method according to claim 1, characterized in that: In step S4, the weather types that can be identified include six scenarios: sunny, cloudy, rainy, foggy, snowy, and night / evening.

4. The method according to claim 1, characterized in that: In step S6, the time-series cache data is divided into a front-end time-series set and a back-end time-series set. The trend of weather continuity, intensification, weakening, clearing up, and turning cloudy is determined by the sign and magnitude of the mean difference.

5. The method according to claim 1, characterized in that: A dual-fusion discrimination method combining traditional visual physical features and a lightweight MobileNetV2 deep learning model is adopted to achieve high-precision weather recognition.

6. The method according to claim 1, characterized in that: The overall algorithm features a lightweight design, supporting offline real-time operation on PC terminals and embedded devices, without the need for cloud servers and professional weather sensors.

Images