Enhancing traditional machine learning methods using concatenation two transfer learning for classification desert regions

Abstract

Deserts cover a significant portion of the earth and present environmental and economic difficulties owing to their harsh conditions. Satellite remote sensing images (SRSI) have evolved into an important tool for monitoring and studying these regions as technology has advanced. Machine learning (ML) is critical in evaluating these images and extracting valuable information from them, resulting in a better knowledge of hard settings and increasing efforts toward sustainable development in desert regions. As a result, in this study, four ML approaches were enhanced by hybridizing them with pre-training methods to achieve multi model learning. Two pre-training approaches (Xception and DeneseNet201) were used to extract features, which were concatenated and fed into ML algorithms light gradient boosting model (LGBM), decision tree (DT), k-nearest neighbors (KNN), and naïve Bayes (NB). In addition, an ensemble voting was used to improve the outcomes of ML algorithms (DT, NB, and KNN) and overcome their flaws. The models were tested on two datasets and hybrid LGBM outperformed other traditional ML methods by 99% in accuracy, precision, recall, and F1 score, and by 100% in area under the curve (AUC)-receiver operating characteristic (ROC).