Machine Learning Algorithms for Satellite Image Classification Using Google Earth Engine and Landsat Satellite Data: Morocco Case Study

Earth observation data have proven to be a valuable resource of quantitative information that is more consistent in time and space than traditional land-based surveys. Remote sensing plays a vital role in collecting data in many aspects of life, whether scientific, economic, or political. Land cover information is very important in supporting urban planning and decision-making and provides many opportunities for mapping and monitoring urban areas. Multiple data sources exist, including satellite data of different resolutions ranging from very high to medium resolution and aerial and drone image acquisitions. Today, accurate land cover information is in high demand. Using satellite imagery and remote sensing techniques for planning and development is becoming a common study conducted by many researchers to find practical solutions to the many problems affecting our planet. The recovery, management, and analysis of these large amounts of satellite imagery pose considerable challenges. The classification of satellite images is a very popular and complex topic. In classification studies over the last decade, researchers have been frequently studying only those three machine learning algorithms RF, CART, and SVM, applied in cities or countries except Morocco which poses a great lack of information on the land use of Morocco. To solve these challenges, six machine learning algorithms were applied and compared to each other based on several evaluation metrics, and then, to avoid the problems of data download and storage space, we used Google Earth Engine, a geospatial processing platform that operates in the cloud. It provides free access to substantial satellite data and free computations to monitor, visualize, and analyze environmental features at the petabyte scale. In this paper, we used Landsat 8 satellite data to perform a land cover classification of Morocco, applying machine learning algorithms, which is a subfield of artificial intelligence. This paper proposes an experimental study of six supervised machine learning algorithms, namely Support Vector Machine (SVM), Random Forest (RF), Classification and Regression Trees (CART), Minimum Distance (MD), Decision Tree (DT) and Gradient Tree (GTB), to classify the water areas, built-up areas, cultivated areas, sandy areas, barren areas and forest areas on the Moroccan territory to deduce at the end the best-performing classifier which has higher accuracy. The classification results are displayed using a set of accuracy indicators, including overall accuracy (OA), Kappa, user accuracy (UA), and producer accuracy (PA). We obtained the best accuracy of 0.93 for the minimum distance (MD) algorithm, but the worst result is 0.74 for the support vector machine (SVM) algorithm. To improve these results, we added different indices such as normalized difference vegetation index (NDVI), normalized difference accumulation index (NDBI), bare soil index (BSI), and modified normalized difference water index (MNDWI). In general, the addition of these indices improves accuracy. When comparing these classifiers before and after the addition of these indices, Minimum Distance yields nearly 93% better accuracy. Therefore, we conclude that it was the best-performing classifier among the other classifiers that can quickly produce more accurate land cover maps, especially for hard-to-reach areas.