Towards a common validation sample set for global land-cover mapping

Validating land-cover maps at the global scale is a significant challenge. We built a global validation data-set based on interpreting Landsat Thematic Mapper (TM) and Enhanced TM Plus (ETM+) images for a total of 38,664 sample units pre-determined with an equal-area stratified sampling scheme. This was supplemented by MODIS enhanced vegetation index (EVI) time series data and other high-resolution imagery on Google Earth. Initially designed for validating 30 m-resolution global land-cover maps in the Finer Resolution Observation and Monitoring of Global Land Cover (FROM-GLC) project, the data-set has been carefully improved through several rounds of interpretation and verification by different image interpreters, and checked by one quality controller. Independent test interpretation indicated that the quality control correctness level reached 90% at level 1 classes using selected interpretation keys from various parts of the USA. Fifty-nine per cent of the samples have been verified with high-resolution images on Google Earth. Uncertainty in interpretation was measured by the interpreter’s perceived confidence. Only less than 7% of the sample was perceived as low confidence at level 1 by interpreters. Nearly 42% of the sample units located within a homogeneous area could be applied to validating global land-cover maps whose resolution is 500 m or finer. Forty-six per cent of the sample whose EVI values are high or with little seasonal variation throughout the year can be applied to validate land-cover products produced from data acquired in different phenological stages, while approximately 76% of the remaining sample whose EVI values have obvious seasonal variation was interpreted from images acquired within the growing season. While the improvement is under way, some of the homogeneous sample units in the data-set have already been used in assessing other classification results or as training data for land-cover mapping with coarser-resolution data.     

Exploring the potential role of feature selection in global land-cover mapping

Global land cover has been acknowledged as a fundamental variable in several global-scale studies for environment and climate change. Recent developments in global land-cover mapping focused on spatial resolution improvement with more heterogeneous features to integrate the spatial, spectral, and temporal information. Although the high dimensional input features as a whole lead to discriminatory strengths to produce more accurate land-cover maps, it comes at the cost of an increased classification complexity. The feature selection method has become a necessity for dimensionality reduction in classification with large amounts of input features. In this study, the potential of feature selection in global land-cover mapping is explored. A total of 63 features derived from the Landsat Thematic Mapper (TM) spectral bands, Moderate Resolution Imaging Spectroradiometer (MODIS) time series enhanced vegetation index (EVI) data, digital elevation model (DEM), and many climate-ecological variables and global training samples are input to k-nearest neighbours (k-NN) and Random Forest (RF) classifiers. Two filter feature selection algorithms, i.e. Relieff and max-min-associated (MNA), were employed to select the optimal subsets of features for the whole world and different biomes. The mapping accuracies with/without feature selection were evaluated by a global validation sample set. Overall, the result indicates no significant accuracy improvement in global land-cover mapping after dimensionality reduction. Nevertheless, feature selection has the capability of identifying useful features in different biomes and improves the computational efficiency, which is valuable in global-scale computing.     

An assessment of the effectiveness of a rotation forest ensemble for land-use and land-cover mapping

Increasing the accuracy of thematic maps produced through the process of image classification has been a hot topic in remote sensing. For this aim, various strategies, classifiers, improvements, and their combinations have been suggested in the literature. Ensembles that combine the prediction of individual classifiers with weights based on the estimated prediction accuracies are strategies aiming to improve the classifier performances. One of the recently introduced ensembles is the rotation forest, which is based on the idea of building accurate and diverse classifiers by applying feature extraction to the training sets and then reconstructing new training sets for each classifier. In this study, the effectiveness of the rotation forest was investigated for decision trees in land-use and land-cover (LULC) mapping, and its performance was compared with performances of the six most widely used ensemble methods. The results were verified for the effectiveness of the rotation forest ensemble as it produced the highest classification accuracies for the selected satellite data. When the statistical significance of differences in performances was analysed using McNemar's tests based on normal and chi-squared distributions, it was found that the rotation forest method outperformed the bagging, Diverse Ensemble Creation by Oppositional Relabelling of Artificial Training Examples (DECORATE), and random subspace methods, whereas the performance differences with the other ensembles were statistically insignificant.     

An all-season sample database for improving land-cover mapping of Africa with two classification schemes

High-quality training and validation samples are critical components of land-cover and land-use mapping tasks in remote sensing. For large area mapping it is much more difficult to build such sample sets due to the huge amount of work involved in sample collection and image processing. As more and more satellite data become available, a new trend emerges in land-cover mapping that takes advantage of images acquired beyond the greenest season. This has created the need for constructing sample sets that can be used in classifying images of multiple seasons. On the other hand, seasonal land-cover information is also becoming a new demand in land and climate change studies. Here we produce the first training and validation data sets with seasonal labels in order to support the production of seasonal land-cover data for entire Africa. Nonetheless, for the first time, two classification systems were created for the same set of samples. We adapted the finer resolution observation and monitoring of global land cover (FROM-GLC) and the Food and Agriculture Organization (FAO) Land Cover Classification System legends. Locations of training-sample units of FROM-GLC were repurposed here. Then we designed a process to enlarge the training-sample units to increase the density of samples in the feature space of spectral characteristics of Moderate Resolution Imaging Spectroradiometer (MODIS) time-series and Landsat imagery. Finally, we obtained 15,799 training-sample units and 7430 validation-sample units. The land-cover type at each point was recorded at the time of maximum greenness in addition to four seasons in a year. Nearly half of the sample units were also suitable for 500 m resolution MODIS data. We analysed the representativeness of the training and validation sets and then provided some suggestions about their use in improving classification accuracies of Africa.     

Assessing effects of temporal compositing and varying observation periods for large-area land-cover mapping in semi-arid ecosystems: Implications for global monitoring

Land-cover is an important parameter in analyzing the state and dynamics of natural and anthropogenic terrestrial ecosystems. Land-cover classes related to semi-arid savannas currently exhibit among the greatest uncertainties in available global land cover datasets. This study focuses on the Kalahari in northeastern Namibia and compares the effects of different composite lengths and observation periods with class-wise mapping accuracies derived from multi-temporal MODIS time series classifications to better understand and overcome quality gaps in mapping semi-arid land-cover types. We further assess the effects of precipitation patterns on mapping accuracy using Tropical Rainfall Measuring Mission (TRMM) observation data. Botanical field samples, translated into the UN Land Cover Classification System (LCCS), were used for training and validation. Different sets of composites (16-day to three-monthly) were generated from MODIS (MOD13Q1) data covering the sample period from 2004 to 2007. Land-cover classifications were performed cumulatively based on annual and inter-annual feature sets with the use of random forests. Woody vegetation proved to be more stable in terms of omission and commission errors compared to herbaceous vegetation types. Generally, mapping accuracy increases with increasing length of the observation period. Analyses of variance (ANOVA) verified that inter-annual classifications significantly improved class-wise mapping accuracies, and confirmed that monthly composites achieved the best accuracy scores for both annual and inter-annual classifications. Correlation analyses using piecewise linear models affirmed positive correlations between cumulative mapping accuracy and rainfall and indicated an influence of seasonality and environmental cues on the mapping accuracies. The consideration of the inter-seasonal variability of vegetation activity and phenology cycling in the classification process further increases the overall classification performance of savanna classes in large-area land-cover datasets. Implications for global monitoring frameworks are discussed based on a conceptual model of the relationship between observation period and accuracy.     

Exploring the addition of Landsat 8 thermal band in land-cover mapping

One focus of remote-sensing studies is obtaining highly accurate land-cover maps, which is essential for quantifying and monitoring changes in the environment. However, thermal data, which can provide auxiliary information, is often ignored in land-cover classification. In this study we compare the performance of different remote-sensing feature combinations with and without the Landsat 8 thermal band (band 10). The results show that overall the thermal feature had a positive effect on mapping land cover. A combination of spectral features, indices and the thermal feature maximized the improvement in accuracy. The proposed classifier was applied to map land cover in an area in Egypt. The thermal feature significantly reduced the confusion between cropland and wetland. The improvement in accuracy obtained by adding the thermal feature was analysed in a time series spanning 1 year. We found that the thermal feature improved the classification accuracy when temperature variations occurred among the different land-cover types. The effect of the thermal feature was also influenced by the land cover; in cloudless conditions, warmer weather can enhance the accuracy improvement of the thermal feature.     

Land-use and land-cover mapping and change detection in Tripura using satellite LANDSAT data

In the present paper efforts have been made by the authors to study the changes in land use and land cover in Tripura using LANDSAT images of two different dates and to see how well data obtained help in the study of geographical phenomena with special reference to land use and land cover. The methodology for land-use mapping and limitations of LANDSAT data have been described in detail. The LANDSAT computer compatible tapes (CCTs) were analysed on the sophisticated interactive Multispectral Data Analysis System (MDAS) with the help of training sets of each category collected during field visits. Ninety per cent accuracy of these categories has been achieved when compared with existing data compiled on ground surveys by the working plan of the Division of Forest Department. The area of each land-use category was also calculated for monitoring land-use changes.     

An image-based texture mapping technique for apparel products exhibition and interior design

In this paper, we propose an image-based texture mapping technique for textile products exhibition that does not require geometric representation of 3D models. Under this technique, a texture grid is built interactively for a target area in an original image. This grid acts as the intermediate space between planar space and texture space. The texture coordinate for each pixel in the target area can be calculated based on this grid, and the 3D effect can be successfully realized by further fine adjustment of the grid. This technique can be applied in apparel products exhibition and interior design.     

Adaptively weighted decision fusion in 30 m land-cover mapping with Landsat and MODIS data

Although the combined use of Landsat and Moderate Resolution Imaging Spectroradiometer (MODIS) time-series data in land-cover classification has been widely adopted, the majority of such use of Landsat and MODIS data is done at the pixel level or feature input level in land-cover classification. We propose in this research a new method to make integrated use of different satellite data by adaptively weighted decision-level fusion. Training and validation samples were collected independently. Training samples were obtained from 329 regions and validation samples from 439 randomly distributed single-point positions. A Support Vector Machine (SVM) classifier was applied to the Landsat 8 data for classification and probability estimation. A Random Forests (RF) classifier was applied to the MODIS time-series data for probability estimation. Weight values were computed based on decision credibility, and reliability values were computed based on data quality. Three decision fusion procedures were performed. In the first procedure, decisions obtained from a Landsat 8 pixel and its corresponding MODIS pixel were fused for improvements (FUSION1). In the second, decisions obtained from the spatial neighbours of the Landsat 8 pixel were added to FUSION1 (FUSION2). In the third, decision fusion only among the Landsat 8 pixel and its spatial neighbours was performed (FUSION3) for comparison. Overall accuracies for the results with Landsat data only, FUSION1, FUSION2, and FUSION3 are 74.0%, 79.3%, 80.6%, and 75.6%, respectively. As a comparison, we also experimented on the use of Landsat and MODIS data by concatenating their features directly. Two classifiers, SVM and RF, were trained and validated on the concatenated features. The overall accuracies were 72.9% and 75.4%, respectively. Results show that the proposed method can utilize information selectively, so that considerable improvements can be obtained and fewer errors introduced. Moreover, it can be easily extended to handle more than two types of data source.     

Effects of terrain-induced shade removal using global DEM data sets on land-cover classification

Free access to global data sets of satellite images and digital elevation models (DEMs) such as Aster Global DEM (GDEM) and Shuttle Radar Topography Mission (SRTM) digital topography are expected to contribute to various study areas that deal with land cover and land use. To assess the capabilities of these global DEM data sets and to provide guidelines for performing shade removal under various terrain and illumination conditions, we evaluated the results of shade removal using the Minnaert correction and C-correction. These corrections were applied, using the GDEM (versions 1 and 2), SRTM, and a DEM derived from a local map (local DEM), to 30 sample images from 20 scenes of 10 path-rows in global land survey (GLS) Landsat-TM/ETM+ images, in terms of statistical indices and the accuracy of land-cover discrimination. The analysis indicated that the results of shade removal depended mainly on the correlation between the cosine of the sunshine incidence angle (cos(i)) and the radiance before shade removal, except in some cases with inferior illumination conditions. Of the three global DEMs, GDEM version 2 had the highest performance in shade removal. However, this study also indicated that successful shade removal was only one of the several factors that increased the accuracy of land-cover classification. In practical applications, shade removal can be recommended only for images where the terrain shade obviously disturbs the original spectral reflection characteristics of each land-cover type and no significant dependence of the land-cover distribution on the slope aspect is assumed. In such cases, also global DEMs evaluated in this study can be used for shade removal.     

Using MERIS fused images for land-cover mapping and vegetation status assessment in heterogeneous landscapes

In this paper we evaluate the potential of ENVISAT–Medium Resolution Imaging Spectrometer (MERIS) fused images for land-cover mapping and vegetation status assessment in heterogeneous landscapes. A series of MERIS fused images (15 spectral bands; 25 m pixel size) is created using the linear mixing model and a Landsat Thematic Mapper (TM) image acquired over the Netherlands. First, the fused images are classified to produce a map of the eight main land-cover types of the Netherlands. Subsequently, the maps are validated using the Dutch land-cover/land-use database as a reference. Then, the fused image with the highest overall classification accuracy is selected as the best fused image. Finally, the best fused image is used to compute three vegetation indices: the normalized difference vegetation index (NDVI) and two indices specifically designed to monitor vegetation status using MERIS data: the MERIS terrestrial chlorophyll index (MTCI) and the MERIS global vegetation index (MGVI).

Results indicate that the selected data fusion approach is able to downscale MERIS data to a Landsat-like spatial resolution. The spectral information in the fused images originates fully from MERIS and is not influenced by the TM data. Classification results for the TM and for the best fused image are similar and, when comparing spectrally similar images (i.e. TM with no short-wave infrared bands), the results of the fused image outperform those of TM. With respect to the vegetation indices, a good correlation was found between the NDVI computed from TM and from the best fused image (in spite of the spectral differences between these two sensors). In addition, results show the potential of using MERIS vegetation indices computed from fused images to monitor individual fields. This is not possible using the original MERIS full resolution image. Therefore, we conclude that MERIS–TM fused images are very useful to map heterogeneous landscapes.     

Segmentation parameter selection for object-based land-cover mapping from ultra high resolution spectral and elevation data

Segmentation is the primary task for image analysis in many practical applications, such as object-based image analysis. Segmentation algorithms need to have properly estimated parameters to provide efficient performance and reliable results. Due to the fact that some features have different shapes and spectral characteristics, it is hard to find the proper parameters for the whole image. In this article, we propose a new method for resolving this issue through the building of a hierarchy of segmentations, based on the number of land-cover classes in the image, namely segmentation scale space (SSS). Both spectral and elevation data are employed in order to enhance the SSS and to obtain a single segmentation for the image. The performance of the proposed algorithm is evaluated using two data sets, which consist of ultra-high resolution aerial images and elevation data with ground sampling distance of 5 and 9 cm, respectively. The experiments demonstrate the efficiency of enhanced segmentation with respect to over and under segmentation cases. Finally, the comparative analysis shows that the accuracy of the proposed method is superior to the classical methods.     

Comparison of multi-temporal NOAA-AVHRR and SPOT-XS satellite data for mapping land-cover dynamics in the west African Sahel

Abstract

Multi-resolution and multi-temporal remote sensing data (SPOT-XS and AVHRR) were evaluated for mapping local land cover dynamics in the Sahel of West Africa. The aim of this research was to evaluate the agricultural information that could be derived from both high and low spatial resolution data in areas where there is very often limited ground information. A combination of raster-based image processing and vector-based geographical information system mapping was found to be effective for understanding both spatial and spectral land-cover dynamics. The SPOT data proved useful for mapping local land-cover classes in a dominantly recessive agricultural region. The AVHRR-LAC data could be used to map the dynamics of riparian vegetation, but not the changes associated with recession agriculture. In areas where there was a complex mixture of recession and irrigated agriculture, as well as riparian vegetation, the AVHRR data did not provide an accurate temporal assessment of vegetation dynamics.     

Vision-based global localization and mapping for mobile robots

We have previously developed a mobile robot system which uses scale-invariant visual landmarks to localize and simultaneously build three-dimensional (3-D) maps of unmodified environments. In this paper, we examine global localization, where the robot localizes itself globally, without any prior location estimate. This is achieved by matching distinctive visual landmarks in the current frame to a database map. A Hough transform approach and a RANSAC approach for global localization are compared, showing that RANSAC is much more efficient for matching specific features, but much worse for matching nonspecific features. Moreover, robust global localization can be achieved by matching a small submap of the local region built from multiple frames. This submap alignment algorithm for global localization can be applied to map building, which can be regarded as alignment of multiple 3-D submaps. A global minimization procedure is carried out using the loop closure constraint to avoid the effects of slippage and drift accumulation. Landmark uncertainty is taken into account in the submap alignment and the global minimization process. Experiments show that global localization can be achieved accurately using the scale-invariant landmarks. Our approach of pairwise submap alignment with backward correction in a consistent manner produces a better global 3-D map.     

An assessment of impacts of land-cover changes on root-zone soil moisture

Land cover exerts considerable control over the exchange of energy, water, and carbon dioxide and other greenhouse gases between land surface and the atmosphere. In China, dramatic land-cover changes have occurred along with rapid economic development in the past 30 years. However, research specifically on whether such land-cover changes have any influence on root-zone soil moisture in the region has started only in very recent few years. In this study, the performance of selected land-surface models (Noah 2.7.1, Noah 3.2, Common Land Model (CLM version 2.0), and Mosaic) implemented in National Aeronautics and Space Administration (NASA)’s Land Information System (LIS version 6.1.6) is first tested using quality-controlled soil moisture observations from 108 in situ sites of the China Meteorological Administration. The best-performing model (CLM2.0) is selected to estimate the influence of land-cover changes on root-zone soil moisture, as well as drought occurrence in Yunnan Province in China. Both the 1992–1993 Advanced Very High Resolution Radiometer (AVHRR) and 2007–2010 Moderate Resolution Imaging Spectroradiometer Collection 5 (MODIS) land-cover products at 1 km resolution are employed to represent 1990 and 2010 land-cover status, respectively. These are verified using the local ground records of Yunnan Province over the two time periods. Their differences are considered roughly as land-cover changes occurring during the period 1990–2010. It is found that land-cover changes from primeval forest to grassland may increase root-zone soil moisture, thus reducing drought, while changes from grassland and primeval forest to cropland or reforested areas have increased the likelihood of drought.     

A hybrid data fusion approach to the hierarchical mapping of land-cover types in the urban core

It is difficult to map land covers in the urban core due to the close proximity of high-rise buildings. This difficulty is overcome with a proposed hybrid, the hierarchical method via fusing PAN-sharpened WorldView-2 imagery with light detection and ranging (lidar) data for central Auckland, New Zealand, in two stages. After all features were categorized into ‘ground’ and ‘above-ground’ using lidar data, ground features were classified from the satellite data using the object-oriented method. Above-ground covers were grouped into four types from lidar-derived digital surface model (nDSM) based on rules. Ground and above-ground features were classified at an accuracy of 94.1% (kappa coefficient or κ = 0.913) and 93.7% (κ = 0.873), respectively. After the two results were merged, the nine covers achieved an overall accuracy of 93.7% (κ = 0.902). This accuracy is highly comparable to those reported in the literature, but was achieved at much less computational expense and complexity owing to the hybrid workflow that optimizes the efficiency of the respective classifiers. This hybrid method of classification is robust and applicable to other scenes without modification as the required parameters are derived automatically from the data to be classified. It is also flexible in incorporating user-defined rules targeting hard-to-discriminate covers. Mapping accuracy from the fused complementary data sets was adversely affected by shadows in the satellite image and the differential acquisition time of imagery and lidar data.     

Stacked Autoencoder-based deep learning for remote-sensing image classification: a case study of African land-cover mapping

Land-cover mapping is an important research topic with broad applicability in the remote-sensing domain. Machine learning algorithms such as Maximum Likelihood Classifier (MLC), Support Vector Machine (SVM), Artificial Neural Network (ANN), and Random Forest (RF) have been playing an important role in this field for many years, although deep neural networks are experiencing a resurgence of interest. In this article, we demonstrate early efforts to apply deep learning-based classification methods to large-scale land-cover mapping. Based on the Stacked Autoencoder (SAE), one of the deep learning models, we built a classification framework for large-scale remote-sensing image processing. We adjusted and optimized the model parameters based on our test samples. We compared the performance of the SAE-based approach with traditional classification algorithms including RF, SVM, and ANN with multiple performance analytics. Results show that the SAE classifier trained with an entire set of African training samples achieves an overall classification accuracy of 78.99% when assessed by test samples collected independently of training samples, which is higher than the accuracies achieved by the other three classifiers (76.03%, 77.74%, and 77.86% of RF, SVM, and ANN, respectively) based on the same set of test samples. We also demonstrated the advantages of SAE in prediction time and land-cover mapping results in this study.