Geographical Supervision Correction for Remote Sensing Representation Learning

Four 1 km global land cover products are currently available to the scientific community: the University of Maryland (UMD) global land cover product, the International Geosphere–Biosphere Programme Data and Information System Cover (IGBP‐DISCover), the MODerate resolution Imaging Spectrometer (MODIS) global land cover product and Global Land Cover 2000 (GLC2000). Because of differences in data sources, temporal scales, classification systems and methodologies, it is important to compare and validate these global maps before using them for a variety of studies at regional to global scales. This study aimed to perform the validation and comparison of the four global land cover datasets, and to examine the suitability and accuracy of different coarse spatial resolution datasets in mapping and monitoring cropland across China. To meet this objective, we compared the four global land cover products with the National Land Cover Dataset 2000 (NLCD‐2000) at three scales to evaluate the accuracy of estimates of aggregated cropland areas in China. This was followed by a spatial comparison to assess the accuracies of the four products in estimating the spatial distribution of cropland across China. A comparative analysis showed that there are varying levels of apparent discrepancies in estimating the cropland of China between these four global land cover datasets, and that both area totals and spatial (dis)agreement between them vary from region to region. Among these, the MODIS dataset has the best fit in depicting China's croplands. The coarse spatial resolution and the per pixel classification approach, as well as landscape heterogeneity, are the main reasons for the large discrepancies between the global land cover datasets tested and the reference data.

Global land cover (GLC) products have been widely used in land use and land cover change (LUCC). It is an urgent need to investigate the harmonization and comparative methods for different products derived from various satellite imagery and classification schemes. MCD12Q1 and GlobCover were comparatively used to identify the changes of land cover at a provincial scale in China. The European Space Agency (ESA) GlobCover and the Terra and Aqua combined Moderate Resolution Imaging Spectroradiometer (MODIS) Land Cover Type (MCD12Q1) datasets of 2005 and 2009 were used and compared comprehensively. The two products were harmonized and reclassified into six types at 500-m resolution, namely, forestland, grassland, cropland, wetland, artificial area, and others. Four aspects of land cover changing rate (K), land use integrated index (Ld), land use dynamic degree (LUDD), and transfer analysis were respectively identified and analyzed. The analysis results show that the two datasets have similar mapping capabilities in monitoring the LUCC, but there are also obvious differences due to the classification schemes, especially for the forestland and grassland. The K of forestland and grassland for GlobCover is respectively − 4.94 and 2.90%, while they are 7.63 and − 47.73% for MCD12Q1. The Ld difference is, respectively, 10 and 1 for MCD12Q1 and GlobCover. The LUDD in Anhui Province from 2005 to 2009 is 3.06% for MCD12Q1 and 0.90% for GlobCover. Additionally, the transfer matrix is also constructed to explore the land cover changes. The significant differences between the two datasets are attributed to forestland and grassland. We conclude that the differences are primarily ascribed to the classification systems. The study provides an important reference for identifying provincial LUCC by jointly using different GLC products.

Abstract. Over past decades, a lot of global land-cover products have been released; however, these still lack a global land-cover map with a fine classification system and spatial resolution simultaneously. In this study, a novel global 30 m land-cover classification with a fine classification system for the year 2015 (GLC_FCS30-2015) was produced by combining time series of Landsat imagery and high-quality training data from the GSPECLib (Global Spatial Temporal Spectra Library) on the Google Earth Engine computing platform. First, the global training data from the GSPECLib were developed by applying a series of rigorous filters to the CCI_LC (Climate Change Initiative Global Land Cover) land-cover and MCD43A4 NBAR products (MODIS Nadir Bidirectional Reflectance Distribution Function-Adjusted Reflectance). Secondly, a local adaptive random forest model was built for each 5∘×5∘ geographical tile by using the multi-temporal Landsat spectral and texture features and the corresponding training data, and the GLC_FCS30-2015 land-cover product containing 30 land-cover types was generated for each tile. Lastly, the GLC_FCS30-2015 was validated using three different validation systems (containing different land-cover details) using 44 043 validation samples. The validation results indicated that the GLC_FCS30-2015 achieved an overall accuracy of 82.5 % and a kappa coefficient of 0.784 for the level-0 validation system (9 basic land-cover types), an overall accuracy of 71.4 % and kappa coefficient of 0.686 for the UN-LCCS (United Nations Land Cover Classification System) level-1 system (16 LCCS land-cover types), and an overall accuracy of 68.7 % and kappa coefficient of 0.662 for the UN-LCCS level-2 system (24 fine land-cover types). The comparisons against other land-cover products (CCI_LC, MCD12Q1, FROM_GLC, and GlobeLand30) indicated that GLC_FCS30-2015 provides more spatial details than CCI_LC-2015 and MCD12Q1-2015 and a greater diversity of land-cover types than FROM_GLC-2015 and GlobeLand30-2010. They also showed that GLC_FCS30-2015 achieved the best overall accuracy of 82.5 % against FROM_GLC-2015 of 59.1 % and GlobeLand30-2010 of 75.9 %. Therefore, it is concluded that the GLC_FCS30-2015 product is the first global land-cover dataset that provides a fine classification system (containing 16 global LCCS land-cover types as well as 14 detailed and regional land-cover types) with high classification accuracy at 30 m. The GLC_FCS30-2015 global land-cover products produced in this paper are free access at https://doi.org/10.5281/zenodo.3986872 (Liu et al., 2020).

This paper introduces a reference sample database that is being developed by the Remote Sensing Unit of the Portuguese Geographic Institute for the accuracy assessment of medium scale land cover products in Portugal. The goal is to provide the worldwide remote sensing community with sufficient data for the accurate estimation of overall and per class proportions of correctly classified area in regional and global land cover products at this part of the globe. This is a massive database that encloses various descriptive attributes for each sample observation, namely primary and alternate reference land cover labels, nominally scored interpretation ratings and location confidence ratings. We present in detail the attached land cover nomenclature and the database design, i.e. the sampling design used to collect sample observations, as well as the process used to identify the most pertinent reference land cover label for each observation. In addition, we briefly describe some statistics about attribute information that was recorded for each observation and that can be used as auxiliary information for the accuracy assessment of land cover products.

In this study, using the common classification systems of IGBP-17, IGBP-9, IPCC-5 and TC (vegetation, wetlands and others only), we studied spatial and areal inconsistencies in the three most recent multi-resource land cover products in a complex mountain-oasis-desert system and quantitatively discussed the uncertainties in classification system conversion. This is the first study to compare these products based on terrain and to quantitatively study the uncertainties in classification system conversion. The inconsistencies and uncertainties decreased from high to low levels of aggregation (IGBP-17 to TC) and from mountain to desert areas, indicating that the inconsistencies are not only influenced by the level of thematic detail and landscape complexity but also related to the conversion uncertainties. The overall areal inconsistency in the comparison of the FROM-GLC and GlobCover 2009 datasets is the smallest among the three pairs, but the smallest overall spatial inconsistency was observed between the FROM-GLC and MODISLC. The GlobCover 2009 had the largest conversion uncertainties due to mosaic land cover definition, with values up to 23.9%, 9.68% and 0.11% in mountainous, oasis and desert areas, respectively. The FROM-GLC had the smallest inconsistency, with values less than 4.58%, 1.89% and 1.2% in corresponding areas. Because the FROM-GLC dataset uses a hierarchical classification scheme with explicit attribution from the second level to the first, this system is suggested for producers of map land cover products in the future.

Major forest changes and land cover transitions based on plant functional types derived from the ESA CCI Land Cover product

Satellite-based land cover mapping plays an important role in understanding changes in ecosystems and biodiversity. There are global land cover products available, however for region specific studies of drivers of infectious disease patterns, these can lack the spatial and thematic detail or accuracy required to capture key ecological processes. To overcome this, we produced our own Landsat derived 30 m maps for three districts in India's Western Ghats (Wayanad, Shivamogga and Sindhudurg). The maps locate natural vegetation types, plantation types, agricultural areas, water bodies and settlements in the landscape, all relevant to functional resource use of species involved in infectious disease dynamics. The maps represent the mode of 50 classification iterations and include a spatial measure of class stability derived from these iterations. Overall accuracies for Wayanad, Shivamogga and Sindhudurg are 94.7 % (SE 1.2 %), 88.9 % (SE 1.2 %) and 88.8 % (SE 2 %) respectively. Class classification stability was high across all three districts and the individual classes that matter for defining key interfaces between human habitation, forests, crop, and plantation cultivation, were generally well separated. A comparison with the 300 m global ESA CCI land cover map highlights lower ESA CCI class accuracies and the importance of increased spatial resolution when dealing with complex landscape mosaics. A comparison with the 30 m Global Forest Change product reveals an accurate mapping of forest loss and different dynamics between districts (i.e., Forests lost to Built-up versus Forests lost to Plantations), demonstrating an interesting complementarity between our maps and the % tree cover Global Forest Change product. When studying infectious disease responses to land use change in tropical forest ecosystems, we recommend using bespoke land cover/use classifications reflecting functional resource use by relevant vectors, reservoirs, and people. Alternatively, global products should be carefully validated with ground reference points representing locally relevant habitats.

Remote sensing based land cover mapping at large scale is time consuming when using either supervised or unsupervised classification approaches. This article used a fast clustering method—Clustering by Eigen Space Transformation (CBEST) to produce a land cover map for China. Firstly, 508 Landsat TM scenes were collected and processed. Then, TM images were clustered by combining CBEST and K-means in each pre-defined ecological zone (50 in total for China). Finally, the obtained clusters were visually interpreted as land cover types to complete a land cover map. Accuracy evaluation using 2159 test samples indicates an overall accuracy of 71.7% and a Kappa coefficient of 0.64. Comparisons with two global land cover products (i.e., Finer Resolution Observation and Monitoring of Global Land Cover (FROM-GLC) and GlobCover 2009) also indicate that our land cover result using CBEST is superior in both land cover area estimation and visual effect for different land cover types.

Achieving food security in particular in Africa continues to pose a major challenge to humankind. It is clear that the future agricultural potential of Africa plays a critical role in meeting this challenge. Although crop yield can be estimated with a degree of reliability using a limited sample of ground observations, the exact crop acreage and the spatial distribution are rarely available. Even though remote sensing offers the ability to produce a rapid and up-to-date land use and land cover database for agricultural monitoring, there are only a few countries in Africa where higher resolution satellite data such as Landsat have been used for land cover map production at the national level. However a number of global products have been produced which contain information on cropland extent. This paper will outline a comparison of four sources of land cover data to determine which product is the most suitable for agricultural monitoring and for the subsequent development of a crop mask. The land cover products used are the Global Land Cover Map (GLC-2000), the Moderate Resolution Imaging Spectroradiometer (MODIS) land cover product (MOD12V1), the SAGE (Center for Sustainability and the Global Environment) and the AFRICOVER dataset from the Food and Agriculture Organisation (FAO). Both the GLC-2000 and MODIS land cover products are at a resolution of 1 km2 while AFRICOVER, based on visual interpretation of 30-m resolution Landsat images, is available at a much finer resolution. The four land cover products are first aggregated to the same resolution so that they can be compared. The legend categories of the four land cover products in this study are reconciled using the method developed in Fritz and See [Fritz, S. and See, L., 2005, Comparison of land cover maps using fuzzy agreement. International Journal of Geographic Information Science, 19, pp. 787–807.] and See and Fritz [See, L.M. and Fritz, S., 2005, A user-defined fuzzy logic approach to comparing global land cover products. In 14th European Colloquium on Theoretical and Quantitative Geography, 9–12 September 2005, Lisbon, Portugal.] that allows overlap between legend definitions to be taken into account. Once the legend definitions between the different land cover products are reconciled, the maps are then compared with national and sub-national statistics. Analysis is undertaken at both continental and national scales as well as sub-national for Sudan and Eritrea. The study generally concludes that MODIS has the tendency to underestimate cropland cover when compared with FAO statistics or AFRICOVER data, whereas GLC-2000 tends to overestimate cropland cover in those countries that are located at the northern transition zone of subtropical shrubland and semi-desert areas. In this area MODIS and SAGE show a relatively similar cropland distribution. Even though the SAGE database has been calibrated with national statistics, it does not perform better than the other two datasets overall, and has highlighted the fact that the SAGE data show regional weaknesses and should be replaced in certain regions by more recent datasets such as GLC-2000 and MODIS, or ideally by a hybrid product that combines the best of the three products, depending upon the region and country.

Deciphering the evolving trajectories of China's megaregions from 1992 to 2020: A novel morphological approach based on global land cover products

Agroforestry, the integration of trees on productive agricultural land (Mosquera-Losada et al., 2018), can be identified through remote sensing methods by the combination of land cover and tree cover maps. Previous work has classified agroforestry as agricultural land with greater than 5% tree cover (Lawson et al., 2025; Zomer et al., 2016).However, there exist several regional, European, and global land cover and tree cover products that could be suitable for agroforestry identification, but these products vary in resolution, data inputs, and methodology of production. Our work benchmarked the performance of four land cover maps(Büttne et al., 2021; Karvatte et al., 2021; Schultz et al., 2025; UKCEH, 2022) and nine tree cover maps(Brandt et al., 2024; Copernicus, 2023, 2025; Hunter et al., 2025; Lang et al., 2023; Tolan et al., 2024; Weinstein et al., 2020) that were capable of mapping trees outside of woodlands. We evaluated the datasets’ ability to identify agroforestry on 25 agroforestry sites across the United Kingdom, including a mix of silvoarable and silvopastoral systems, as well as planting ages, densities, and species, as well as nearby agricultural (no trees) and woodland (no agriculture) control fields.We found that a number of datasets used in previous studies underperformed when distinguishing agroforestry from control fields as well as the previously utilised pixel-based approaches being unsuitable to identify agroforestry fields as a whole. Datasets with coarse resolutions (>10m) often confused proximal small woodlands for trees within agricultural fields. Many datasets struggled to map trees in silvoarable systems, likely due to their linear arrangement differing from that of other trees outside of woodlands.In addition, the majority of datasets were unable to identify agroforestry sites planted since 2000, suggesting a 20-year lag in identification. The only tree identification method capable of identifying young sites was the fine-tuning of the DeepForest tree detection model (Weinstein et al., 2020) with local data, suggesting a need for tree cover datasets that are capable of identifying seedlings and saplings.We used the best-performing datasets (balanced accuracy 83-87%) to create a map of agroforestry, with quality flags to signify agreement between datasets. We conclude that there is 517,300 ha (3% of UAA) of area under agroforestry in the United Kingdom. Our map could have further use cases for calculating the uptake of agroforestry, as well as its benefits to people and nature, such as carbon storage, biodiversity impacts, farm income, and health of crops and livestock. We also conclude the need for model training on agroforestry trees, to both identify young trees and those with complex planting arrangements.

Two global 1 km land cover data sets derived from 1992-1993 Advanced Very High Resolution Radiometer (AVHRR) data are currently available, the International Geosphere-Biosphere Programme Data and Information System (IGBP-DIS) DISCover and the University of Maryland (UMd) 1 km land cover maps. This paper makes a preliminary comparison of the methodologies and results of the two products. The DISCover methodology employed an unsupervised clustering classification scheme on a per-continent basis using 12 monthly maximum NDVI composites as inputs. The UMd approach employed a supervised classification tree method in which temporal metrics derived from all AVHRR bands and the NDVI were used to predict class membership across the entire globe. The DISCover map uses the IGBP classification scheme, while the UMd map employs a modified IGBP scheme minus the classes of permanent wetlands, cropland/natural vegetation mosaic and ice and snow. Global area totals of aggregated vegetation types are very similar and have a per-pixel agreement of 74%. For tall versus short/no vegetation, the per-pixel agreement is 84%. For broad vegetation types, core areas map similarly, while transition zones around core areas differ significantly. This results in high regional variability between the maps. Individual class agreement between the two 1 km maps is 49%. Comparison of the maps at a nominal 0.5 resolution with two global ground-based maps shows an improvement of thematic concurrency of 46% when viewing average class agreement. The absence of the cropland mosaic class creates a difficulty in comparing the maps, due to its significant extent in the DISCover map. The DISCover map, in general, has more forest, while the UMd map has considerably more area in the intermediate tree cover classes of woody savanna/ woodland and savanna/wooded grassland.

Classification and change detection of high -resolution remote sensing images using low-resolution labels is a challenging issue in remote sensing community. In this paper, we propose a semi-supervised method based on deep learning to generate high-resolution change maps of Maryland using low-resolution NLCD labels in the Multitemporal Semantic change Detection challenge track (Track MSD) of the 2021 IEEE Data Fusion Contest. Firstly, we refined the NLCD labels using five global land cover products. Subsequently, Fully Convolutional Network (FCN) was trained for the classification of NAIP images with the refined NLCD labels and then the network training was continued with the pseudo-labels from the previous classification results and Modified Normalized Difference Water Index (MNDWI) extracted from Landsat-8 images as the new features. Finally, after the decision-level fusion of the two training periods, change detection results were generated from the bi-temporal classification maps with a post-processing to improve the performance. This algorithm achieved an average Intersection-over-Union (IoU) of 0.6657 on the test dataset and won the 2 <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">nd</sup> place of the contest.

Representation Learning (RL) aims to convert data into low-dimensional and dense real-valued vectors, so as to realize reasoning in vector space. RL is one of the important research contents in the analysis of health data. This paper systematically reviews the latest research on Electronic Health Records (EHR) RL. We searched the Web of Science, Google Scholar, and Association for Computing Machinery Digital Library for papers involving EHR RL. On the basis of literature review, we propose a new taxonomy to categorize the state-of-the-art EHR RL methods into three categories: statistics learning-based RL methods, knowledge RL methods and graph RL methods. We analyze and summarize their characteristics according to the input data form and underlying learning mechanisms. In addition, we provide evaluation strategies to verify the quality of EHR representations from both intrinsic and extrinsic perspectives. Finally, we put forward three promising research directions to promote future research. Overall, this survey aims to provide a profound overview of state-of-the-art developments in the field of EHR RL and to help researchers find the most appropriate methods.

India, a mega-diverse country, possesses a wide range of climate and vegetation types along with a varied topography. The present study has classified forest types of India based on multi-season IRS Resourcesat-2 Advanced Wide Field Sensor (AWiFS) data. The study has characterized 29 land use/land cover classes including 14 forest types and seven scrub types. Hybrid classification approach has been used for the classification of forest types. The classification of vegetation has been carried out based on the ecological rule bases followed by Champion and Seth's (1968) scheme of forest types in India. The present classification scheme has been compared with the available global and national level land cover products. The natural vegetation cover was estimated to be 29.36% of total geographical area of India. The predominant forest types of India are tropical dry deciduous and tropical moist deciduous. Of the total forest cover, tropical dry deciduous forests occupy an area of 2,17,713 km(2) (34.80%) followed by 2,07,649 km(2) (33.19%) under tropical moist deciduous forests, 48,295 km(2) (7.72%) under tropical semi-evergreen forests and 47,192 km(2) (7.54%) under tropical wet evergreen forests. The study has brought out a comprehensive vegetation cover and forest type maps based on inputs critical in defining the various categories of vegetation and forest types. This spatially explicit database will be highly useful for the studies related to changes in various forest types, carbon stocks, climate-vegetation modeling and biogeochemical cycles.