Raster4ML: A Geospatial Raster Processing Library for Machine Learning

Machine learning has been playing a key role in different domains of geospatial science, such as, natural resource management, hydrology, agricultural monitoring, land cover dynamics, and so on. Researchers and scientists often use raster data derived from satellites, airplanes or unmanned aerial vehicles (UAVs) coupled with novel machine learning algorithms to estimate physical, chemical, or biological parameters and explain the underlying processes of different phenomenon. However, geospatial raster data is different from natural images often seen in computer vision applications. For example, a common task in utilizing machine learning for raster data is to derive hand-crafted features (such as, vegetation indices or texture features) based on different disciplines or research questions. Such features can explain certain characteristics, which cannot be interpreted by the individual bands or channels. To date, there has been numerous vegetation indices or texture features reported in literature. Therefore, it is difficult for researchers or scientists to derive the required features from raster data and extract the values for sample locations to perform tasks in machine learning. To ease these manual tasks, we propose a Python-package called "Raster4ML", which helps the users to easily create machine learning ready dataset from given geospatial data. The package can automatically calculate more than 350 vegetation indices and various texture features. It can also derive the statistical outputs from areas of interests for multiple raster data at once, which reduces the manual processing time for users. On the other hand, the package provides supports for dynamic visualization of the geospatial data that can help the users interacting with the inputs and outputs. The package can assist the geospatial community by efficiently handling the feature engineering portion, whereas the scientists can focus more on algorithm development, training, and reproducibility of the machine learning application.