Remote Sensing and GIS Applications in Environmental Sciences
Classification of agricultural crops using the ensemble structure of machine learning models and time series of remote sensing data (Agricultural lands of Mahabad city)
Document Type : Original Article
Authors
1 Department of Photogrammetry and Remote Sensing, Faculty of Geodesy and Geomatics , Khajeh Nasiraddin Toosi University of Technology, Tehran, Iran.
2 Department of Photogrammetry and Remote Sensing, Faculty of Geodesy and Geomatics, Khajeh Nasiraddin Toosi University of Technology, Tehran, Iran
Abstract
Effective identification of crop types is fundamental for safeguarding food security. In this context, satellite-based remote sensing (RS) data has emerged as a promising tool, offering wide spatial coverage and high temporal frequency. However, due to the significant intra- and inter-class variability among crop types, there remains a growing need for more precise classification methods using RS data. This study proposes a novel supervised parallel-cascaded architecture (Ensemble models) targeting seven crop classes. The parallel component consists of five independent branches, each generating probability maps for different target classes using multi-temporal Sentinel-1 and Landsat-8/9 imagery. These maps were produced through various machine learning and deep learning classification models. Subsequently, the generated probability maps from the parallel branches were stacked and used as a new feature set fed into a meta-model in the cascaded structure. The results demonstrated that the Convolutional Neural Network (CNN) algorithm achieved superior performance in both parallel and cascaded components, making it the optimal choice for both the base and meta models. The proposed architecture achieved an overall accuracy (OA) of 94.79% and a Kappa coefficient of 0.937. These results highlight the significant advantage of the parallel-cascaded structure over conventional classification methods that simply stack RS data and feed it into a single model leading to a 6% increase in overall accuracy. Future research could use higher-resolution distance assessments and more accurate synthetic hybrid models.
Keywords
Main Subjects
)