Assessing CNN and Semantic Segmentation Models for Coarse Resolution Satellite Image Classification in Subcontinental Scale Land Cover Mapping

Based on studies using high-medium resolution images, convolutional neural networks (CNNs) and semantic segmentation have shown superiority over classical machine learning (ML), particularly in small-scale mapping. However, few/no studies have assessed the techniques on coarse resolution image classification for extensive area land cover mapping. In this study, we evaluated the performance and feasibility of three CNN models (1-D CNN, 2-D CNN, and 3-D CNN), and U-net for coarse-resolution satellite image classification and compared them to a random forest (RF) classifier. We utilized time-series, coarse resolution (1 km) composite imageries acquired by FengYun-3C visible and infrared radiometer. Labeled datasets were collected as shapefiles and split into three independent datasets: training, validation, and test datasets, and preprocessed to meet each model's input format requirements. We conducted several experiments to optimize models and select the best models. Then, the best models were evaluated on an unseen dataset. Among the DL models, one-dimensional (1-D) CNN achieved the highest overall accuracy (OA) 0. 87 and kappa (<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">k</i>) 0.84, 2% higher than the best results attained by 2-D CNN, 3-D CNN, and U-net models. However, 1-D CNN is outperformed by RF which achieved 0.89 (OA) and 0.87 (k). Achieving the best and the second-best results using RF and 1-D CNN models, respectively, indicates the superiority of the pixel-based method and the insignificance of spatial information in coarse-resolution image classification. Furthermore, although the DL models can yield high accuracy, especially 1-D CNN, they are less feasible than RF classifiers for coarse-resolution satellite image classification in extensive area land cover mapping.