Earth Observation Satellite Imagery Information Based Decision Support Using Machine Learning

Abstract

This paper presented a review on the capabilities of machine learning algorithms toward Earth observation data modelling and information extraction. The main purpose was to identify new trends in the application of or research on machine learning and Earth observation—as well as to help researchers positioning new development in these domains, considering the latest peer-reviewed articles. A review of Earth observation concepts was presented, as well as current approaches and available data, followed by different machine learning applications and algorithms. Special attention was given to the contribution, potential and capabilities of Earth observation-machine learning approaches. The findings suggested that the combination of Earth observation and machine learning was successfully applied in several different fields across the world. Additionally, it was observed that all machine learning categories could be used to analyse Earth observation data or to improve acquisition processes and that RF, SVM, K-Means, NN (CNN and GAN) and A2C were among the most-used techniques. In conclusion, the combination of these technologies could prove to be crucial in a wide range of fields (e.g., agriculture, climate and biology) and should be further explored for each specific domain.

Studying How Machine Learning Maps Mangroves in Moderate-Resolution Satellite Images

“…In supervised classification, learning is performed on labeled training datasets, where the output i.e. the variable to be classified or predicted is already known [32]. For land cover classification, training datasets are typically collected through field surveys or ground truth.…”

“…Supervised machine learning techniques used for mangrove mapping include decision tree (DT) [13], random forest (RF) [33], maximum likelihood classification (MLC) [19], support vector machine (SVM) [34], and classification and regression trees (CART) [14]. In unsupervised classification, learning is conducted on training datasets that have not been labeled, where the output variable is unidentified [32]. Unsupervised classification is particularly useful when prior knowledge of field data is not available or when there are no experienced analysts.…”