Conditional Gaussian Enhanced Dense Correlation Matching for Cross-Category Land Cover Classification

As the requirements for the downstream tasks of land cover classification (LCC) continue to increase, the category system used for LCC is constantly being refined. This causes previous land cover products and manually annotated training samples to become quickly outdated. Meanwhile, manually annotating samples with more refined categories is extremely time-consuming. To address this impasse, a cross-category knowledge transfer process is needed that can directly generate land cover products under a fine-scale category system using existing training samples under a large-scale category system. Accordingly, this article proposes a cross-category LCC method called conditional Gaussian enhanced dense correlation matching (CGE-DCM). CGE-DCM uses samples under a large-scale category system for training. It then uses only one annotated example of each fine-scale category to achieve fine-scale LCC. In cases with very little sample support, the problems caused by different spectra of the same object and different objects of the same spectrum in complex scenes can be particularly severe. To solve this problem and improve the accuracy of classifications of complex objects, CGE-DCM offers a dense correlation matching (DCM) strategy. In addition, context distribution is different under fine-scale category systems than it is under large-scale category systems. For this reason, CGE-DCM features a conditional Gaussian enhancement mechanism and designs different loss functions for degenerated and nondegenerated scenarios to ensure the stability of the model. Extensive experiments on Gaofen-2 and Orbita hyperspectral satellite (OHS) images demonstrate the effectiveness of each module in CGE-DCM and its superiority over existing methods.

I. Introduction

Remote sensing-based land cover classification (LCC) has received much attention because of its wide-ranging applications [1]. LCC assigns a land cover category to each pixel in a remote sensing image, which helps us better understand the relationship between humans and nature [2]. On the one hand, from the perspective of data sources, the literature on LCC has variously focused on the mapping of multitemporal [3], multimodal [4], and single-form data [5]. On the other hand, from a methodological perspective, the extant research can be divided into fully supervised [6], weakly supervised [5], [7], and unsupervised domain adaptation [8] classifications. These studies have achieved remarkable results in improving the intelligence of LCC methods by enhancing their generalization ability when there are insufficient samples or differences in resolutions, data sources, or geographic locations. However, we have found that, with the development of the geographic information industry, the categories of LCC products are required to be more and more refined. Therefore, some high-resolution LCC datasets, such as the gaofen image dataset (GID), provide hierarchical category systems: the first-level system contains only large-scale categories, and the second-level system contains fine-scale categories [9]. Table I presents the hierarchical category systems of GID. Here, the finer the classification, the higher the cost of manual labeling. In this case, it is important to understand how categories can be updated to form a more refined system if the land cover samples contain only large-scale categories. This article defines this task as the cross-category LCC problem.TABLE IHierarchical Category Systems of GID

Large-scale categories	Fine-scale categories
Built-up	Industrial land Urban residential Rural residential Traffic land
Farmland	Paddy field Irrigated land Dry cropland Garden plot
Forest	Arbor woodland Shrub land
Meadow	Natural grassland Artificial grassland
Water	River Lake Pond