Machine Learning + Geography = ❤
Utilizing neural networks in spatial data analysis
Introduction
Earth Sciences need to process, often in real-time, datasets of complex internal structure, various resolutions, a multitude of formats, and significant volume. Spatial data may originate from observations made during field research, aerial and satellite imaging, wearable sensors, and more. Ever-increasing data stream imposes a decent amount of automation in data analysis. This ensures that geographers can access up-to-date and reliable data. In this short presentation, you will learn about several use cases of utilizing machine learning, mainly deep learning, in supporting geographical space analysis.
What is ML?
We say that a computer program learns as it gains experience on a particular class of tasks when performance improves with additional data. In other words, the machine learning model, which is the program's main component, adapts to the input data. Machine learning is discovering an approximation of a specific function based on a large portion of input and output reference data (β).
ML + Geography
Machine learning in geography is an exciting combination of multiple disciplines of science and technology. To start your adventure, you must be reasonably proficient in Earth Science. The basic understanding of processes and phenomena in the geographical space is crucial to understanding data utilized during model training. Identifying data sources should not be a problem. Multiple public portals offer high-quality remote sensing imagery. Knowledge of mathematics is also essential. Not only it’s needed to understand the algorithms used, but to be able to follow the scientific literature and stay up to date. Finally, coding allows us to materialize our ideas for solving research problems through a computer program.
Machine learning is a vast discipline. Nowadays, apart from classic ML algorithms, neural networks are also used. We call it deep learning (DL). Understanding the principle of neural networks’ operation requires mathematical foundations. In a nutshell, a neural network in an iterative process (called training) of model reconfiguration concerning the error it commits in performing the assigned task.
Semantic segmentation
The purpose of segmentation is to divide the image into characteristic, disjoint areas called regions. Each region is assigned to at least one of the strictly defined categories. Segmentation supports the solution of several exciting research problems. An excellent example is recognizing land cover types (LULC). In the example above, satellite imagery and U-Net were used to find forested areas.
Classification
The purpose of classification is to group the input material by assigning each element to a particular category. A set of categories or classes is strictly defined. An example of the use of DCNN to support the interpretation of geographical space is the study of the occurrence of fallow land in the Łódź Voivodeship.
Time series classification
The classification does not have to be based only on remote sensing imagery. Nothing stands in the way of using data from wearable sensors to classify the type of activity performed by mobile phone users.
Data augmentation
Generative adversarial networks (GAN) opened new image processing and analysis possibilities. Inpainting, dataset augmentation using artificial samples, or the increasing resolution of aerial imagery are only a few notable examples of utilizing GANs in remote sensing. Here GAN has been applied to enhance panchromatic images with Normalized Difference Vegetation Index (NDVI).
GAN can also perform advanced spatial queries based on spatial feature similarity when appropriately applied. In the example, several patches are grouped by a neural network by their similarity in visible features.
This approach can be further extended to generate artificial but realistic patches.
Summary
I hope this short presentation gave you a general idea of how ML can be used in research work. In my opinion, Earth Sciences, especially geography, can gain a lot from introducing the developed methods and techniques to its methodological workshop. If you want to know more or need help starting your adventure in ML in Geography, feel free to contact me.
