Achieving sustainable development goals coupling earth observation data with machine learning

Abstract

The main purpose of this work is to assess and understand the achievement of Sustainable Development Goals by means of Earth Observation (EO) data and Machine Learning (ML) technologies. Thus, this study intends to promote and suggest the use of EO and ML in benefits to the Sustainable Development Goals (SDGs) to support and optimize the actual industry and field processes and moreover provide new insights (techniques) on EO approaches and applicability as well as ML techniques. A review on the Sustainable Development concept and its goals is presented followed by EO data and methods and its approaches relevant to this field, giving special attention to the contribution of ML methods and algorithms as well as their potential and capabilities to support the achievement of SDGs. Additionally, different ML approaches and techniques are reviewed (i.e., Classification and Regression techniques, Non-Binary Decision Tree (NBDT), and two novel methods are proposed, designated as: Random Forest built based on the Non-Binary Decision Tree (NBRF) and Fusion of techniques). Both developed methods are applied, optimized and validated to two case studies also aligned with specific SGDs: Case study I – Identification and mapping of healthy or infected crops, tackling SDGs 2, 8, 9 and 12; and Case study II - Deep-sea mining exploitation SDGs 8, 9, 12 and 14). Such is achieved by using data provided by European satellites or programs that allows to also contribute to the goals for Europe’s Space strategy. For Case study I, the parameters analysed to achieve the respective SDGs correspond to: several vegetation indices as well as the values of the spectral bands. Such parameters have been extracted by means of EO data (from Sentinel-2) and validated with different ML approaches. The results obtained from the different ML approaches suggest that for Case study I, the best classification technique (overall accuracy of 92.87%) as well as the best regression (Root mean square error of 0.148) corresponds to the Fusion of techniques All the applied techniques, however, show their applicability on this case study with good results, disregarding the NBDT which is the “weakest” one (best result on all tests: accuracy of 57.07%). For Case study II, the parameters analysed to achieve the respective SDGs correspond to the topography of the seafloor and, physical and biogeochemical ocean’s parameters. Such parameters have been extracted by means of EO data (from CMEMS and GEBCO) and validated with different ML approaches. The results of these approaches suggests that the best technique corresponds to the Fusion of techniques with a root mean square error of 0.196. However, not all the techniques proved to be appropriated, where the NBDT present the worst results (best result on all tests: accuracy 60.62%). Overall, it is observed that EO plays a key role in the monitoring and achievement of the SDGs given its cost-effectiveness pertaining to data acquisition on all scales and information richness, and the success of ML upon EO data analysis. The applicability of ML techniques allied to EO data has proven, by both case studies, that can contribute to the SDGs and can be extrapolated to other applications and fields.

Keywords: Sustainable Development Goals; Earth Observation; Europe Space Strategy; Machine Learning; Deep-sea Mining; Agriculture.