High-Resolution Land-Cover Mapping Based on a Cross-Resolution Deep Learning Framework and Available Low-Resolution Labels

Abstract

High-resolution land cover mapping (LCM) is pivotal in numerous disciplines but still challenging to be acquired because traditional supervised methods require a substantial number of high-resolution labels that is labouring and expensive. To this issue, abundantly available low-resolution land-cover maps are regarded as alternative label sources, but the mismatch of spatial resolution and the misidentification of different land-cover categories introduced an amount of noisy labeled samples. This study introduces a novel cross-resolution deep learning framework, termed CRNN, to generate high-resolution LCM by leveraging low-resolution mapping products. First, a high-resolution backbone is proposed to safeguard the preservation of output resolution while simultaneously retaining the deep feature extraction capability of the network. Furthermore, an attention module is incorporated into the CRNN framework to alleviate the adverse impact of imbalanced samples. More importantly, to address the label noise issue, a weakly supervised loss based on feature similarity is proposed and calculated for obtaining dependable supervision information from low-resolution LCM products. The qualitative and quantitative results demonstrate that the CRNN framework surpasses several state-of-the-art methods. Moreover, based on the CRNN, the 10-m resolution land-cover maps of Beijing and Shanghai for 2020 are produced using 30-m resolution LCM products as reference data. As a further application, CRNN provides a viable and promising approach for reusing existing data products, contributing to some extent toward achieving sustainability goals.

Keywords

• Attention module
• deep learning (DL)
• land cover mapping (LCM)
• noisy label
• weakly supervised learning