A method for predicting outdoor surface atmospheric temperature based on visible light

Abstract

The present invention discloses a method for predicting outdoor surface atmospheric temperature based on visible light. The method includes three steps: training sample construction, model construction, and surface temperature prediction. First, surface image data under natural light is collected, along with surface atmospheric temperature data at the center of the image region at the corresponding moment. A holdout method is used to construct training samples. Second, a CNN-GRU network is used to extract image features, which are input into a regression network composed of a fully connected neural network. The training samples are used to train a surface atmospheric temperature prediction model. Finally, model parameters are optimized, and the trained model is used to predict the surface atmospheric temperature hourly. The present invention predicts surface temperature by extracting surface visible light image features and using a transfer learning method to quickly apply the model to similar landforms to implement surface atmospheric temperature prediction. This reduces the cost of the Internet of Things system, expands the application of deep learning methods in surface temperature monitoring, and enhances the perception capability of visible light cameras.

Description

Technical Field

[0001] The present invention relates to the field of agricultural and forestry information technology, and in particular to a method for predicting outdoor surface atmospheric temperature based on visible light. Background Art

[0002] Atmospheric temperature influences the habits and distribution of plants and animals, and is of significant importance in agriculture, forestry, and climate. Traditionally, IoT devices such as image acquisition and temperature and humidity sensors regularly collect data and transmit it to data centers for environmental monitoring. With the development of smart agriculture and smart forestry, the granularity of ambient atmospheric temperature monitoring is becoming increasingly demanding. The increase in sensor types and quantity not only increases the cost and energy consumption of forest IoT systems, but also restricts the application scope and upgradeability of existing IoT systems. Therefore, how to monitor and predict ambient atmospheric temperature while reducing costs has become a critical technical challenge that needs to be addressed in the field of agricultural and forestry information technology.

[0003] In recent years, with the development of breath data collection technology, a variety of atmospheric temperature monitoring methods have emerged, mainly including near-infrared and temperature sensors. These methods involve the update and upgrade of collection equipment. In theory, there is a certain correlation between atmospheric temperature and sunlight, and the atmospheric temperature can be calculated through changes in surface illumination. It is feasible to use machine learning methods to extract high-dimensional features in visible light images and thus calculate the ambient atmospheric temperature. Summary of the Invention

[0004] The purpose of the present invention is to solve the deficiencies of the prior art and provide a method for predicting land surface temperature based on visible light images.

[0005] In order to achieve the above technical effects, the present invention is implemented by the following technical solutions: a method for predicting outdoor surface atmospheric temperature based on visible light images, characterized in that it includes the following steps:

[0006] S1. Build training dataset:

[0007] Data acquisition: A visible light camera is used to collect visible light radiation data reflected from the ground in an outdoor environment, i.e., a visible light image, and a temperature sensor is used to collect the surface atmospheric temperature at the center of the corresponding image area;

[0008] S2, a combination of data classification labels, data preparation, and target temperature setting;

[0009] Data preprocessing:

[0010] Assign image classification labels according to environmental meteorological information,

[0011] Category label 1: According to the natural law of sunrise and sunset, the collected images are divided into three light categories according to time period: morning low light, medium-high light, and evening low light, and labeled IA, IB, and IC respectively. It can also be divided into more detailed categories, and the labels are increased accordingly. For example, if the sunrise time in a certain place is 6:00 and the sunset time is 18:00, the images can be divided into three categories: 30 minutes before sunrise to 3 hours after sunrise, 3 hours after sunrise to 3 hours before sunset, and 3 hours before sunset to 30 minutes after sunset, namely IA: [5:30, 9:00), IB: [9:00, 15:00), IC: [15:00, 18:30]. Alternatively, a neural network can be used to classify the average light intensity within the △T time, where △T is 30 minutes, and the number of categories is not limited to three.

[0012] Category label 2: The dataset is divided into S1, S2, ...S24 according to the 24 solar terms.

[0013] Data preparation: t is the image sampling period. Setting parameters N and M indicates that N images within a continuous N*t time range are used to predict the temperature at the next M*t time. t0 represents the time of the last image in N images. The image data of the time window [t0-(N-1)*t, t0] is used as the input of the model's visible light image.

[0014] Target temperature of training data: Calculate the average temperature within the time range of t0+M*t, and use the result as the target temperature of training data;

[0015] Wherein, N>1, M≥0; when M=0, it indicates the expected output temperature at time t0; when M>0, it indicates the expected output temperature at time t0+M*t;

[0016] Temperature prediction data set division: The holdout method is used to construct training samples, with 80% of the collected data as the training data set and the remaining 20% ​​as the test data set;

[0017] S3. Model construction: CNN-GRU, Embedding, and regression networks are used to construct a temperature prediction model. The CNN-GRU network is used to extract illumination features from visible light images. Embedding is used to classify meteorological features and map them into a fixed-length feature space. A fully connected neural network with three hidden layers and one output node is used to construct a temperature regression neural network. The regression neural network is optimized using Adam with a learning rate of lr = 0.001. The surface atmospheric temperature prediction model is then trained using training samples.

[0018] S4. Model parameter tuning: Set the temperature error threshold parameters α and β, where α = 0.2°C and β = 3°C, which can be adjusted as needed. Use test data to evaluate the model prediction effect. The model outputs the predicted temperature as x and the actual temperature as y. When |xy| ≤ α, the predicted temperature is correct, otherwise it is wrong. When |xy| ≥ β, it means that the predicted value deviates too much and needs to be optimized. From the training data set, retrieve the image data 10 minutes before and after the moment, use image rotation, flipping and other methods to expand the training samples, and then train the model.

[0019] S5. Actual prediction:

[0020] Data preprocessing: Assign classification labels 1 and 2 in the same way as the labels in S1, and prepare data using the same parameters N and M as in S1;

[0021] Temperature prediction: Load the S3 optimized model parameters, input preprocessed data, and output the predicted target temperature.

[0022] Furthermore, the data acquisition in S1 is specifically as follows: a visible light camera at a fixed position and angle collects visible light radiation data modulated by the ground surface and ground cover, i.e., a visible light image.

[0023] Furthermore, the images captured by the visible light camera in S1 are three-channel images with a pixel size greater than 2048*1080 pixels, and the sky accounts for less than 1 / 2 of the captured images.

[0024] Furthermore, CNN-GRU, Embedding and regression networks together constitute the surface atmospheric temperature prediction model. The CNN network in the model adopts the feature extraction structure of image processing networks such as VGG or Restnet, or uses different convolution kernels to establish a four-layer convolutional neural network.

[0025] Furthermore, when using training samples to train the surface atmospheric temperature prediction model, the GRU network input dimension is the same as the dimension of the CNN network output after Flattening; Embedding is used to map the average meteorological features into a 64-dimensional vector; the GRU network output and the Embedding result are concatenated as the input of the regression network, the number of nodes in the regression hidden layer is set to half of the number of nodes in the previous input layer, and the output layer outputs the temperature prediction result.

[0026] The beneficial effects of the present invention are:

[0027] The present invention predicts surface atmospheric temperature from a new perspective, using artificial intelligence methods to establish a correlation between visible light illumination and ground atmospheric temperature through training data. Compared with existing methods and technologies, this method does not rely on temperature sensing equipment. Temperature prediction only requires high-definition cameras and algorithms deployed in data centers to centrally process and predict the atmospheric temperature of each node, effectively reducing the types of IoT devices for environmental monitoring, enhancing the perception capabilities of IoT devices, and saving energy and costs. It facilitates the upgrade and transformation of IoT systems for environmental monitoring, expands the surface atmospheric temperature prediction method, and provides a reference for simplifying surface atmospheric temperature prediction. At the same time, the present invention has good applicability and strong generalization capabilities under conditions of medium and low illumination. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0029] Figure 1 This is a flow chart of the surface temperature prediction method based on visible light images according to the present invention;

[0030] Figure 2 This is a schematic diagram of heat exchange between the earth's surface and the atmosphere according to the present invention;

[0031] Figure 3 Schematic diagram of the equipment deployment method of the present invention. DETAILED DESCRIPTION

[0032] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present invention.

[0033] Example 1

[0034] See Figure 1 As shown, a surface temperature prediction method based on visible light images specifically includes the following steps:

[0035] Step 1: Build a training dataset:

[0036] A visible light camera is used to collect visible light radiation data reflected from the ground in outdoor environments. The camera is set to collect surface images in real time at a fixed azimuth and elevation angle, or images are captured from videos. The pixels of the collected images must be larger than a 2048*1080 pixel three-channel image, and the sky ratio in the captured image must be less than 1 / 2. The collection time interval can be set from a few seconds to tens of minutes.

[0037] A temperature sensor is used to collect the surface atmospheric temperature at the center of the corresponding image area, or the temperature in the local meteorological data is used as the temperature of the collected image.

[0038] Data preprocessing:

[0039] Assign image classification labels according to environmental meteorological information,

[0040] Category label 1: According to the natural law of sunrise and sunset, the collected images are divided into three light categories according to time period: morning low light, medium-high light, and evening low light, and labeled IA, IB, and IC respectively. It can also be divided into more detailed categories, and the labels are increased accordingly. For example, if the sunrise time in a certain place is 6:00 and the sunset time is 18:00, the images can be divided into three categories: 30 minutes before sunrise to 3 hours after sunrise, 3 hours after sunrise to 3 hours before sunset, and 3 hours before sunset to 30 minutes after sunset, namely IA: [5:30, 9:00), IB: [9:00, 15:00), IC: [15:00, 18:30]. Alternatively, a neural network can be used to classify the average light intensity within the △T time, where △T is 30 minutes, and the number of categories is not limited to three.

[0041] Category label 2: The dataset is divided into S1, S2, ...S24 according to the 24 solar terms.

[0042] Temperature prediction data set division: Alternatively, the parameter setting method of using the image data of the time window [t0-(N-1)*t, t0] to calculate the predicted temperature at time t0+M*t is used to divide the samples into a training data set and a test data set.

[0043] Step 2: Model construction:

[0044] The temperature prediction model is divided into two parts: feature extraction and regression network. The CNN-GRU network is used to extract illumination features, and the illumination type characteristics are integrated as illumination features. A fully connected neural network with three hidden layers and one output node is used to build a temperature regression neural network. Then, the training samples are used to train the surface atmospheric temperature prediction model.

[0045] The CNN network in the feature extraction network uses 7*7, 5*5, 4*4, and 4*4 convolution kernels and ReLU activation functions to extract image features in sequence. The output of the CNN network is flattened into a one-dimensional vector and input into the GRU network. The average meteorological feature map corresponding to the input image is mapped into a 64-dimensional vector using Embedding. The GRU network output and the Embedding result are concatenated into a 4096+64-dimensional vector and input into the regression network. The number of regression network nodes is set to 4160, 2080, 1040, 510, and 1, respectively, with a dropout rate of 0.1. The network is optimized using Adam, with lr = 0.001.

[0046] Step 3: Model parameter tuning:

[0047] Set the temperature error threshold parameters α and β, where α = 0.2°C and β = 3°C, and adjust as needed. Use test data to evaluate the model prediction effect. The model outputs the predicted temperature as x and the actual temperature as y. When |xy| ≤ α, the predicted temperature is correct, otherwise it is wrong. When |xy| ≥ β, it means that the predicted value deviates too much and needs to be optimized. Retrieve image data 10 minutes before and after the moment from the training dataset, use image rotation, flipping, and other methods to expand the training samples, and then train the model.

[0048] Step 4: Actually bring in the prediction:

[0049] Data preprocessing: Assign classification labels 1 and 2 in the same way as in step 1, and prepare data using the same parameters N and M as in step 1;

[0050] Temperature prediction: Load the model parameters optimized in step 3, input the preprocessed data, and output the predicted target temperature.

[0051] The electronic device of the present invention can be deployed at an IoT endpoint or at a convergence node of an IoT system.

Claims

1. A method for predicting outdoor surface atmospheric temperature based on visible light images, characterized in that: The following steps are involved: S1. Build training dataset: Data acquisition: A visible light camera is used to collect visible light radiation data reflected from the ground in an outdoor environment, i.e., a visible light image, and a temperature sensor is used to collect the surface atmospheric temperature at the center of the corresponding image area; S2, a combination of data classification labels, data preparation, and target temperature setting; Data preprocessing: Assign image classification labels according to environmental meteorological information, Category label 1: Based on the natural laws of sunrise and sunset, the collected images are divided into three lighting categories: morning low light, medium-high light, and evening low light, and labeled IA, IB, and IC respectively; Category label 2: The dataset is divided into S1, S2, ...S24 according to the 24 solar terms. Data preparation: t is the image sampling period. Setting parameters N and M indicates that N images within a continuous N*t time range are used to predict the temperature at the next M*t time. t0 represents the time of the last image in N images. The image data of the time window [t0-(N-1)*t, t0] is used as the input of the model's visible light image. Target temperature of training data: Calculate the average temperature within the time range of t0+M*t, and use the result as the target temperature of training data; Wherein, N>1, M≥0; when M=0, it indicates the expected output temperature at time t0; when M>0, it indicates the expected output temperature at time t0+M*t; Temperature prediction data set division: The holdout method is used to construct training samples, with 80% of the collected data as the training data set and the remaining 20% ​​as the test data set; S3. Model construction: CNN-GRU, Embedding, and regression networks are used to construct a temperature prediction model. The CNN-GRU network is used to extract illumination features from visible light images. Embedding is used to classify meteorological features and map them into a fixed-length feature space. A fully connected neural network with three hidden layers and one output node is used to construct a temperature regression neural network. The regression neural network is optimized using Adam with a learning rate of lr = 0.

001. The surface atmospheric temperature prediction model is then trained using training samples. S4. Model parameter tuning: Set the temperature error threshold parameters α and β, where α = 0.2°C and β = 3°C, and adjust as needed. Use test data to evaluate the model prediction effect. The model outputs the predicted temperature as x and the actual temperature as y. When |xy| ≤ α, the predicted temperature is correct, otherwise it is wrong. When |xy| ≥ β, it means that the predicted value deviates too much and needs to be optimized. From the training data set, retrieve the image data 10 minutes before and after the moment, use methods such as image rotation and flipping to expand the training samples, and then train the model. S5. Actual prediction: Data preprocessing: Assign classification labels 1 and 2 in the same way as the labels in S1, and prepare data using the same parameters N and M as in S1; Temperature prediction: Load the S3 optimized model parameters, input preprocessed data, and output the predicted target temperature.

2. The method for predicting outdoor surface atmospheric temperature based on visible light according to claim 1, characterized in that: The data acquisition in S1 is specifically as follows: a visible light camera at a fixed position and angle collects visible light radiation data modulated by the ground surface and ground cover, namely a visible light image.

3. The method for predicting outdoor surface atmospheric temperature based on visible light according to claim 1, characterized in that: The images captured by the visible light camera in S1 are three-channel images with a resolution greater than 2048*1080 pixels, and the sky accounts for less than 1 / 2 of the captured images.

4. The method for predicting outdoor surface atmospheric temperature based on visible light according to claim 1, characterized in that: CNN-GRU, Embedding and regression networks together constitute the surface atmospheric temperature prediction model. The CNN network in the model adopts the feature extraction structure of image processing networks such as VGG or Restnet, or uses different convolution kernels to establish a four-layer convolutional neural network.

5. The method for predicting outdoor surface atmospheric temperature based on visible light according to claim 1, characterized in that: When using training samples to train the surface atmospheric temperature prediction model, the GRU network input dimension is the same as the dimension of the CNN network output after flattening; Embedding is used to map the average meteorological features into a 64-dimensional vector; the GRU network output and the Embedding result are concatenated as the input of the regression network, the number of nodes in the regression hidden layer is set to half of the number of nodes in the previous input layer, and the output layer outputs the temperature prediction result.

Images