Land cover information is vital for research and applications concerning natural resources and environmental modeling. Accuracy assessment is an important dimension in use and production of land cover information. GlobeLand30 is a relatively new global land cover information product with a fine spatial resolution of 30 m and is potentially useful for many applications. This paper describes the methods for and results from the first country-wide and statistically based accuracy assessment of GlobeLand30 2010 land cover dataset over China. For this, a total of 8400 validation sample pixels were collected based on a sampling design featuring two levels of stratification (ten geographical regions, each with nine or eight land-cover classes). Validation sample data with reference class labels were acquired from visual interpretation based on Google Earth high-resolution satellite images. Error matrices for individual regions and entire China were estimated properly based on the sampling design adopted, with the former aggregated to get the latter through suitable weighting. Results were obtained, with agreement at a sample pixel defined both as a match between the map (class) label and either the primary or alternate reference label therein and, more strictly, as a match between the map label and the primary reference label only. Based on the former definition of agreement, the overall accuracy of GlobeLand30 2010 land cover for China was assessed to be 84.2%. User’s accuracy and producer’s accuracy were both greater than 80% for cultivated land, forest, permanent snow and ice, and bareland, with user’s accuracy for water bodies estimated 94.2% (82.1% for wetland, 79.8% for artificial surface) and producer’s accuracy for grassland estimated 89.0%. These indicate that GlobeLand30 2010 depicts land cover circa 2010 in China quite accurately, although estimates of accuracy indicators based on the latter definition of agreement were lower as expected with an estimated national overall accuracy of 81.0%. Regional and class variations in accuracy were revealed and examined in the light of their associations with land cover distributions and patterns. Implications for use and production of GlobeLand30 land cover information were discussed, so were commonality and lack of it between GlobeLand30 and other fine-resolution land cover products.

It is increasingly recognized that classification accuracy should be characterized locally at the level of individual pixels to depict its spatial variability to better inform users and producers of land-cover information than by conventional error-matrix-based methods. Local or per-pixel accuracy is usually estimated through empirical modelling, such as logistic regression, which often proceeds in a class-aggregated or a class-stratified way, with the latter being generally more accurate due to its accommodation for between-class inhomogeneity in accuracy-context relations. As an extension to class-stratified modelling, class-heterogeneity-stratified modelling, in which logistic models are built separately for contextually heterogeneous vs. homogeneous sub-strata in individual strata of map classes, is proposed in this paper for proper handling of within-class inhomogeneity in accuracy-context relations to increase accuracy of estimation. Unlike in existing literature where sampling is usually approached separately, the double-stratification method is also adopted in sampling design so that more sample data are likely allocated to heterogeneous sub-strata (which are more prone to misclassifications than homogeneous ones). This class-heterogeneity-stratified method furnished for sampling and modelling jointly thus constitutes an integrative framework for accuracy estimation and information refinement. As the first step in building up such a framework, this paper investigates the proposed double-stratification method’s performance and sensitivity to sample size regarding local accuracy estimation in comparison with those of existing methods through a case study concerning Globeland30 2010 land cover over Wuhan, China. A detailed review of existing methods for analyses, estimation, and use of local accuracy was provided, helping to put the proposed research in a broader context. Candidate explanatory variables for logistic regression included sample pixels’ map classes, positions, and contextual features that were computed in different-sized moving windows. Relative performances of these methods were evaluated based on an independent reference sample, with all methods found reliable. It was confirmed that the proposed method is in general the most accurate, as observed with varying sample sizes. The proposed method’s competitive performance is thus proved, reinforcing its potential for information refinement. Extensions to and uncertainty aspects of the proposed method were discussed, with further research proposed.

Multi-feature, especially multi-temporal, remote-sensing data have the potential to improve land cover classification accuracy. However, sometimes it is difficult to utilize all the features efficiently. To enhance classification performance based on multi-feature imagery, an improved rotation forest, combining Principal Component Analysis (PCA) and a boosting naïve Bayesian tree (NBTree), is proposed. First, feature extraction was carried out with PCA. The feature set was randomly split into several disjoint subsets; then, PCA was applied to each subset, and new training data for linear extracted features based on original training data were obtained. These steps were repeated several times. Second, based on the new training data, a boosting naïve Bayesian tree was constructed as the base classifier, which aims to achieve lower prediction error than a decision tree in the original rotation forest. At the classification phase, the improved rotation forest has two-layer voting. It first obtains several predictions through weighted voting in a boosting naïve Bayesian tree; then, the first-layer vote predicts by majority to obtain the final result. To examine the classification performance, the improved rotation forest was applied to multi-feature remote-sensing images, including MODIS Enhanced Vegetation Index (EVI) imagery time series, MODIS Surface Reflectance products and ancillary data in Shandong Province for 2013. The EVI imagery time series was preprocessed using harmonic analysis of time series (HANTS) to reduce the noise effects. The overall accuracy of the final classification result was 89.17%, and the Kappa coefficient was 0.71, which outperforms the original rotation forest and other classifier ensemble results, as well as the NASA land cover product. However, this new algorithm requires more computational time, meaning the efficiency needs to be further improved. Generally, the improved rotation forest has a potential advantage in remote-sensing classification.

Assessing model performance is a continuous challenge for modelers of land use change. Comparing land use models with two neutral models, including the random constraint match model (RCM) and growing cluster model (GrC) that consider the initial land use patterns using a variety of evaluation metrics, provides a new way to evaluate the accuracy of land use models. However, using only two neutral models is not robust enough for reference maps. A modified neutral model that combines a density-based point pattern analysis and a null neutral model algorithm is introduced. In this case, the modified neutral model generates twenty different spatial pattern results using a random algorithm and mid-point displacement algorithm, respectively. The random algorithm-based modified neutral model (Random_MNM) results decrease regularly with the fragmentation degree from 0 to 1, while the mid-point displacement algorithm-based modified neutral model (MPD_MNM) results decrease in a fluctuating manner with the fragmentation degree. Using the modified neutral model results as benchmarks, a new proposed land use model, the Dynamics of Land System (DLS) model, for Jilin Province of China from 2003 to 2013 is assessed using the Kappa statistic and Kappa_in-out statistic for simulation accuracy. The results show that the DLS model output presents higher Kappa and Kappa_in-out values than all the twenty neutral model results. The map comparison results indicate that the DLS model could simulate land use change more accurately compared to the Random_MNM and MPD_MNM. However, the amount and spatial allocation of land transitions for the DLS model are lower than the actual land use change. Improving the accuracy of the land use transition allocations in the DLS model requires further investigation.