The aim of Intelligent parking systems in cities is to enable quick identification of available parking spaces. In this project, an intelligent method for vehicle counting using satellite imagery and machine learning was developed. An existing approach was analyzed, and the Amazon SageMaker platform which is based on Geodatas of Ternopil, Ukraine was chosen for implementation.
Analyzing the number of cars in big cities, can see that it is constantly growing. The transport
industry is one of the large sectors which determines the development of the industry as a whole and
agriculture in any country, including the European Union (EU). One car consumes an average of 1 ton
of oxygen per year and emits about 600-800 kg of carbon dioxide, 40 kg of nitrogen oxides, and 200
kg of unburned hydrocarbons [
The relevance of this topic is in the fact that the development of the intelligent method for counting cars from satellite images based on machine learning will give the possibility to easily and quickly determine the number of cars in a locality, and calculate the amount of emissions.
This article is presented in the following structure: Section 2 discusses the analysis of related works, and Section 3 presents an Intelligent method for counting cars from satellite images based on machine learning. Section 4 presents the implementation of the method, and Section 5 is conclusions.
Given to account the increase in urban population and traffic jams, smart parking is always a
strategic issue [
2023 Copyright for this paper by its authors.
rounded image drawn in a parking space and provides information about empty parking spaces.
Authors of this work [
The authors of [
In [
For intelligent vehicle detection and counting, works such as [
Paper [
Also, companies that own servers are trying to develop platforms that allow automating the
machine learning process: AWS SageMaker [
An intelligent method for counting cars from satellite images based on machine learning can be described in several steps:
Step 1. Data preprocessing (Figure 1): This stage includes the collection of satellite images of the selected area.The images are then divided into squares with defined boundaries to simplify processing and analysis. Data preprocessing, especially when working with satellite imagery, is an important process that usually includes the following steps:
1.1. Image collection: First of all, a series of satellite images are collected from information providers such as NASA, ESA or commercial providers such as Google Earth.
1.2. Geometric correction: Images require geometric correction to correct any distortions caused by satellite movement, shooting angle, etc. This process usually involves image registration algorithms that align the image with a map or other images. 1.3. Normalisation: It is also important to normalise the intensity of the pixels, especially when using images captured at different times of the day or year when the lighting may vary. This typically involves calculating the mean and standard deviation of the pixel intensity and transforming each pixel so that its intensity is relative to the mean intensity.
Step 1. Data preprocessing Image collection
Geometric correction
Image delineation
Normalisation
Dividing into squares 1.4. Dividing into squares: In order to simplify data processing and improve the efficiency of algorithms, large satellite images are divided into smaller squares, or "tiles". The size of these tiles can vary greatly depending on the task and computing resources, but it is usually chosen so that each tile contains enough information for analysis (e.g. 256x256 pixels).
The images processed using this method can then be used to train a machine learning model or perform other analytical procedures.
Step 2: Training the machine learning model (Figure 2): Using a limited number of marked images (i.e. images where cars have already been identified and marked by humans), a machine learning model (e.g. deep neural networks such as convolutional neural networks) is trained.
This model is trained to detect the presence of a car in the image. The process of training a machine learning model, in particular a deep neural network (DNN) such as a convolutional neural network (CNN), can be considered as follows:
2.1. Marking the data: First of all, a marked data set is required, where each image has a corresponding label indicating whether a car is present in the image. Marks can be binary (e.g., 1 if the car is present and 0 if it is not), or they can be more complex, such as indicating the position of the car in the image. 2.2. Creating a model: A CNN model is included several layers. The first layers, commonly known as "convolutional layers", are used to identify key features in the ima ge, such as edges, textures, colors, etc. Each convolutional layer consists of several filters, each of which defines a specific feature. Next layers, known as "fully connected layers", are used to combine these features and determine the final mark.
2.3 Model training: The model is trained by inputting the marked images into the model and adjusting the weighting factors to minimize the difference between the predicted marks of the model and the actual marks. This process is known as "backward error propagation" and usually involves the use of optimization algorithms such as stochastic gradient descent.
2.4. Model testing: After training the model on the training dataset, it needs to be validated on a separate dataset known as the "test set". This helps to ensure that the model can generalize its learning to new data.
2.5. Model evaluation: The model is evaluated using various metrics such as precision (proportion of correctly classified images), recall (proportion of true positive cases that were correctly identified), accuracy (proportion of cases that were correctly identified as positive), F1 metric (harmonic mean between precision and recall), and others.
Step 3. Vehicle detection (Figure 3): Once the model has been properly trained, it can be applied to a large set of unlabeled satellite images.
The model looks at each square of the image, determines whether a car is present, and counts the number of cars. Once trained, a machine learning model (e.g., a convolutional neural network) can be applied to a new set of satellite images to detect cars. The process of using the model for car detection can be described as follows:
3.1. Data preparation: New images to be applied to the model are prepared in the same way as the training images. This means that the images are divided into squares (or "tiles") of the same size as in the training set.
3.2. Model implementation: The machine learning model is then applied to each tile. It takes the tile as input and produces an output that indicates the probability of a car being present on the tile.
3.3 Classification threshold: The probability produced by the model is usually compared to a certain threshold to determine whether a tile contains a car. For example, if the model produces a probability of 0.7 and the threshold is 0.5, then the tile is classified as containing a car.
3.4. Counting cars: The number of cars in the image is calculated by counting the number of tiles that were classified as containing a car.
This process is repeated for all new images.
Step 4. Emissions analysis and forecasting (Figure 4): Once the number of cars on the maps is determined, this information can be used to calculate the current level of emissions.
This can be done using different emission models for different types of vehicles. The data (Table 1) can also be used to predict future emissions levels based on trends in vehicle numbers, whether they are increasing or decreasing.
Calculating emissions
Collecting data for forecasting
Training the Gradient
Boosting model
Forecasting future emissions
Emissions forecasting and analysis can be done using statistical methods and machine learning algorithms such as Gradient Boosting. The steps of this stage are presented below: 4.1 Calculating emissions: To calculate current emissions, a model is applied that relates the number of cars to CO2 emissions. This model can be based on various parameters, such as the average age of cars, average mileage, engine type, etc. For example, can use data from official sources or scientific research to determine the average CO2 emissions per car per year.
4.2. Collecting data for forecasting: To forecast future emissions, a dataset is used that includes a time series of vehicle counts (or calculated CO2 emissions) for previous time periods. Other parameters can also be included, such as economic indicators, demographic information, policy and climate information, and so on.
4.3. Training the Gradient Boosting model: Gradient Boosting is a machine learning technique used to make predictions based on previous data. It creates an ensemble of simple models (often decision trees), each of which attempts to correct the errors of the previous model. The model learns by minimising the loss through gradient descent. The sequence of steps is described below: 4.3.1 Initialisation: Starts with a simple model. This can be any model, but often a very simple model is used that simply predicts the average value of the target variable. It serves as a baseline for subsequent models. 4.3.2 Boosting: In the next step, create a new model, that attempts to correct the mistakes made by the previous model. That provides training a new model on the "leftovers", or errors, of the previous model. For example, if the previous model predicted that CO2 emissions would be 10 tonnes, but the actual value was 8 tonnes, the residuals would be -2 tonnes.
4.3.3 Combining models: The new model is combined with the previous model to create a more accurate model. This is realized by adding the predictions of the new model to the predictions of the previous models. In addition, each new model is trained in small steps (often known as the "learning rate") to avoid overtraining.
4.3.4 Iteration: This process is repeated many times (often hundreds or thousands of iterations), each time creating a new model that attempts to correct the errors of previous models. The final model is simply the sum of the predictions of all the individual models.
Gradient Boosting takes into account model losses or errors and uses a gradient descent method to minimise these losses. During the gradient descent process, the model parameters are continuously adjusted to reduce the overall model error until the best possible parameters are reached.
4.4. Forecasting future emissions: Once the model has been trained, it can be used to predict future emissions based on the latest available data.
This approach allows for a variety of factors and patterns to be incorporated into the predictions, making it a very powerful tool for emissions analysis and prediction and vehicle detection.
This method can be useful for a variety of applications, including determining the effectiveness of emission reduction policies, monitoring urbanization trends, and developing traffic management systems.
To implement the described method, satellite images of Ternopil City were collected from 2004 (Fig. 5) to 2019 (Fig. 6). Each image was divided into sectors (or squares). This was chosen to simplify the analysis, as each sector can be processed separately.
To implement the described method, the Amazon SageMaker platform was chosen. Satellite images of Ternopil City from 2004 to 2019 were uploaded to Amazon S3 for further processing. Amazon SageMaker tools and services, such as Jupyter Notebook and Amazon Sagemaker Ground Truth, were used to divide the images into sectors and mark cars on the images.
After that, a machine learning model was built using Amazon SageMaker Autopilot, which allowed the model to automatically train to recognise cars on satellite images.
Using a trained machine learning model on the number of marked cars in each sector recognized cars in each sector (Figure 7). The number of cars was tracked over the period 2004-2019.
Based on the car count data, a change index was built for each sector. This index shows how much the number of cars in each sector has changed over the period. The change index was visualized on a map (Figure 8), where darker squares correspond to larger changes. Once the number of cars has been determined for each sector, these data can be summed to give the total number of cars in Ternopil in 2004 (Figure 9). As a result, 28,579 cars were recognised in 2004. In this way, it is possible to track how the total number of cars has changed over time and to make predictions about the future number of cars in the city.
The Gradient Boosting Regressor model was used to forecast the amount of CO2 emissions and the number of cars in Ternopil for the period from 2020 to 2030.
The following conclusions can be drawn from the analysis of the predicted number of cars (Figure 10) and CO2 emissions (Figure 11) for the years 2004-2025: The number of cars and CO2 emissions were based on actual data from 2004 to 2019. The forecasted vehicle numbers for the years 2020-2025 are increasing over time. The forecasted CO2 emissions also increase over time, although some years may have smaller increases compared to other years.
The forecasted CO2 emissions for 2020-2025 show a general upward trend but with deviations in different years. Emission values may vary from year to year due to the introduction of CO2 emission standards, such as EU regulations, which force car manufacturers to reduce emissions.
Given to account the forecasts, can expect an increase in the number of cars and, consequently, an increase in CO2 emissions in Ternopil in the future. The values of these forecasts can be used for planning and decision-making on sustainable development of the transport system and reducing environmental impact. The developed method combines two key functions - object recognition on satellite images and forecasting the number of cars in the context of emissions accounting and analysis. Thus, the developed method opens up new possibilities for accurate vehicle counting and emissions forecasting, which has significant potential in many areas requiring traffic flow monitoring and environmental sustainability.
In this work, a method was developed that combines car detection in satellite images with car count and CO2 emissions forecasting. For this purpose, satellite images of the city of Ternopil from 2004 to 2019 were used, which were divided into quadrants for further analysis. In addition, the Gradient Boosting Regressor model based on time series analysis was used to forecast CO2 emissions. The actual and projected data show an increase in the number of cars and CO2 emissions in Ternopil. These predicted values show a stable trend to increase but with variations in different years. Thus, the developed method can be useful for car counting and forecasting emissions based on satellite images. This can be important for monitoring traffic flow, infrastructure planning, and developing environmental strategies to reduce the environmental impact of vehicles. Possible further research could include analysing the impact of traffic on air quality and developing emission reduction strategies, studying the impact of infrastructure on car traffic, and investigating the effectiveness of environmental policies in the city.