Machine learning is everywhere, not just in application areas that we hear about every day. Learning systems have the potential to positively impact society through their application to fields ranging from finance and public health to the earth sciences. In Machine learning for earth sciences, Maurizio Petrelli has delivered a guide to the application of machine learning for earth sciences that assumes some familiarity with the application area, but no fluency in machine learning itself. By showing specific code examples and interactions with the tools, he walks the reader through the steps of implementing simple machine learning models, both supervised and unsupervised, before exploring a few specific applications in the earth sciences.

Petrelli begins by laying out and walking through the complete development of simple machine learning models, from data import and preparation to modeling using many of the common techniques. The implementations are built on Python, NumPy, sklearn, and pandas, and for more complex models, PyTorch. This section of the book is incredibly broad and, inevitably, somewhat shallow. The author provides extensive references rather than including elaborate explanations of the functions being used. This approach is compact, but perhaps a bit too compact. For example, the section on imputation only mentions constant value and mean-value imputation of missing values, while many other techniques are quite approachable and often give better results.

The chapters on modeling (one each for unsupervised and supervised learning) are very good. Each section begins with one chapter on the underlying methods (for example, clustering, principal component analysis, mixture models) and follows with chapters on their application to earth science problems such as petrology and multispectral image data. The chapter on unsupervised methods is especially clear and useful, interleaving concise discussions of the methods with practical code examples. Throughout the book, the code is clean and useful--far better than the notorious online resources in the machine learning field. The brief discussion of scaling models is well done, containing material on cloud-based learning systems.

This book is a compact presentation of how to begin implementing machine learning models in the earth sciences--but again, perhaps too compact. This is especially true of the discussion on deep learning, which I found to be just too sparse. The top-level descriptions of many deep learning approaches are good, but I think it would be difficult to begin implementing a deep learning approach using only the material herein. While reading the book, I often wished that the author had gone a bit more slowly.

This book is essential for anyone planning to apply machine learning to earth science data (including multispectral and hyperspectral imaging). For maximum benefit, the reader should treat it as both an extensive tutorial as well as a bibliography: be prepared to code along with the examples and to look up the references.