Detection of green areas in the City of Buenos Aires — Open Government Week
In the Open Government Week event organized by the Government of the City of Buenos Aires, we offered an in-person workshop where we addressed the detection of green areas in cities through the use of satellite imagery, machine learning and open data with our tools, Satproc and Unetseg. In this post, we give you more details about this issue and our detection project applied to the City of Buenos Aires.
Geospatial imagery, machine learning and open data
Technology increasingly provides us more tools to monitor key features of cities, thus promoting quality public policies and efficient services. Currently, thanks to the price decline, it is possible to access high-quality geospatial imagery and analyze them using different artificial intelligence techniques.
Fortunately, the number of studies which use satellite imagery as a source of information is continuously growing. Likewise, the implementation of machine learning and deep learning tools is expanding, as well as its use by different sectors, including public administration. The possibility of performing scalable, prompt and affordable complex analysis of large areas is particularly appealing for this sector.
Moreover, there is a global tendency among governments to encourage open data availability. This information is not only a critical input for the growth and improvement of private developments and an increased government transparency, but also a great aid to encourage innovation and engagement among new stakeholders.
Technology and public policies: how to improve access to green areas in cities
As mentioned before, the price decline enabled an increased access to technology both for the private and public sectors. The increasing use of technology is allowing governments to perform more accurate and detailed analyses, thereby becoming more efficient in allocating resources and issuing public policies of higher quality.
One of the main issues that the sector is facing is the need to gather information about large areas. This not only implies a complex and expensive deployment of resources, but also it can be subject to potential human errors during measurement.
Measuring and monitoring the number of green areas in a city, for example, can be much more efficient if satellite imagery is used. But… Why are green areas important? Why is it important to map them?
Over the last few years, the lack of green areas in big cities has been identified as a key aspect to improve in order to keep fighting against climate change. Among the purposes of green areas, its role in decreasing the likelihood of floods, preventing high temperatures and improving air quality is highlighted.
The first step to address the issue is knowing where they are located, understanding how they are distributed, its use and which distance a person has to travel to access them, depending on where they live, just to name a few examples.
Mapping green areas in four steps:
The mapping of large areas, or in this case, of green spaces, consists in four big steps: pre-processing, model training, prediction and post-processing. Here we outline each step and, at the end of this article, you can find a link to the code to run it yourself.
1- During the pre-processing step, we generate the dataset that will be used to train the model, based on images and masks located where we already know that there are green areas, thereby defining the area of interest.
This is the most time-consuming step. To streamline the process, we created the Satproc package, a tool developed as a Python library that processes, in a scalable way, a great number of images, which helps us to create the appropriate datasets for the segmentation models. In this case, its function is to identify the pertaining class for each image: park or vegetation.
2- During the training step, the model learns to recognize the patterns which define the object of interest. To that end, we use a machine learning model based on U-net, a convolutional neural network architecture used for the semantic classification and segmentation of objects. For this step, we will use Unetseg and we will divide the dataset in three parts: one for training, another one for validation and the last one for testing.
3- During the prediction step, the already trained model processes new images to find the patterns which define the object of interest. Based on each image, a new image where each pixel represents the probability of having detected the object of interest is generated.
4- Finally, during the post-processing step, a probability threshold is applied to the prediction obtained, in order to keep only the pixels with high probability of being hits.
If you want to learn more about the mapping of green areas, you can access Google Colab. In the next post, we will explain in detail how to run the code and we will focus on more technical aspects.
If you have any doubts or if you are interested in applying the code to your project or business, do not hesitate to contact us at contact@dymaxionlabs.com or through our social media.
