HiResNet: A Low-to-High Deep Learning Framework for Updating Large-Scale Land Cover Inventory

This Large-scale land cover (LC) mapping is critical for monitoring Earth's surface and managing environmental changes. While machine learning (ML) and deep learning (DL) methods have significantly advanced LC mapping efforts, the reliance on high-quality training labels remains a major limitation due to their time-intensive and costly collection process. To address this challenge, we propose HiResNet, a novel DL framework that combines a multi-scale feature extractor (MSFE) within convolutional neural networks (CNNs), a vision transformer (ViT) architecture, and a semi-supervised loss function, to produce a 10 m LC map using an existing low resolution (LR) 30 m LC map. In particular, this framework reduces dependence on high-resolution (HR) training labels while generating HR (10 m) LC maps using Sentinel-2 imagery and LR 30 m labels. The capability of the proposed method is evaluated in province of Ontario (ON), Canada, a region characterized by diverse landscapes and different LC types, covering approximately 1.07 million km2. The results demonstrate a remarkable ability of the proposed model to enhance spatial resolution and achieve high classification accuracy, with an overall accuracy (OA) of 89% for large-scale LC classification. Compared to conventional methods and advanced DL architectures, such as SegFormer, ViT, Random Forest (RF), DeepLabV3, L2HNet, and HRNet, the proposed method exhibits a superior performance by handling class imbalances and extracting complex features. Furthermore, uncertainty estimation using Shannon entropy and class disagreement metrics demonstrates that the highest levels of uncertainty are associated with object edges, narrow features (e.g., rivers and roads), and complex LC classes like forests.