Landsat 8Bands’ 1 to 7 spectral vectors plus machine learning to improve land use/cover classification using Google Earth Engine

This paper explores a spectral vector-based methodology on Landsat 8 bands of the visible wavelengths, that is deep-blue (1) to shortwave infrared (7), to improve the urban land features classification. Using two different ratio models, based on two and three bands’ combinations in the cloud environment of Google Earth Engine, the Uncertainty reducing Spectral Vector (USV_r), the Onward Continuous Spectral Vector (OSV_c) and the Onward Discontinuous Spectral Vector (OSV_d) are proposed as new entries for the land use land cover (LULC) classification. Two different sizes of arrays are built, i.e. 42 vectors and 15 vectors corresponding to the same number of derivative bands and new pixels′ values. A decision tree is built in J.48 and applied to select the most suitable derivative bands for the analysis. Hereafter, the selected ones are stacked and submitted to five machine learning classifiers using a supervised process, namely, Classification and Regression Trees (CART), Random Forest (RF) Gradient Boosting (GBR), Support Vector Machine (SVM) and Minimum Distance (MD). This method was tested in the two cities of Bamenda and Foumban in west-Cameroon highlands, due to their good representativeness of tropical hilly urban areas’ spatial heterogeneity. The results are satisfying for 4/5 classifiers, up to 87% Overall Accuracy, OA, for 0.82 kappa coefficient, KC, in Bamenda, while combining SVM/OSV_d. Whereas, in Foumban, the classifiers perform up to 85%OA and 0.78 KC for the combination SVM/USV_r. Only the MD classifier has always performed below 80%OA. The process has been found better than performing classifiers directly on the multispectral (MS) image, by providing more possibilities of hidden spectral indices not yet explored, as far as we know, to detect and discriminate between LULC features, plus an accurate extraction of human settlements.