Sources of error in accuracy assessment of thematic land-cover maps in the Brazilian Amazon

Valid measures of map accuracy are critical, yet can be inaccurate even when following well-established procedures. Accuracy assessment is particularly problematic when thematic classes lie along a land-cover continuum, and boundaries between classes are ambiguous. In this study, we examined error sources introduced during accuracy assessment of a regional land-cover map generated from Landsat Thematic Mapper (TM) data in Rondônia, southwestern Brazil. In this dynamic, highly fragmented landscape, the dominant land-cover classes represent a continuum from pasture to second growth to primary forest. We used high spatial resolution, geocoded videography as a reference, and focused on second-growth forest because of its potential contribution to the regional carbon balance. To quantify subjectivity in reference data labeling, we compared reference data produced by five trained interpreters. We also quantified the impact of other error sources, including geolocation errors between the map and reference data, land-cover changes between dates of data collection, heterogeneous reference samples, and edge pixels.Interpreters disagreed on classification of almost 30% of the samples; mixed reference samples and samples located in transitional classes accounted for a majority of disagreements. Agreement on second-growth forest labels between any two interpreters averaged below 50%, while agreement on primary forest was over 90%. Greater than 30% of disagreement between map and reference data was attributed to geolocation error, and 2.4% of disagreement was attributed to change in land cover between dates. After geocorrection, 24% of remaining disagreements corresponded to reference samples with mixed land cover, and 47% corresponded to edge pixels on the classified map. These findings suggest that: (1) labels of continuous land-cover types are more subjective and variable than commonly assumed, especially for transitional classes; however, using multiple interpreters to produce the reference data classification increases reference data accuracy; and (2) validation data sets that include only non-mixed, non-edge samples are likely to result in overly optimistic accuracy estimates, not representative of the map as a whole. These results suggest that different regional estimates of second-growth extent may be inaccurate and difficult to compare.     

Improving spatiotemporal reflectance fusion using image inpainting and steering kernel regression techniques

A novel spatiotemporal reflectance fusion method integrating image inpainting and steering kernel regression fusion model (ISKRFM) is proposed to improve the fusion accuracy for remote-sensing images with different temporal and spatial characteristics in this article. This method first detects the land-cover changed regions and then fills them with unchanged similar pixels by an exemplar-based inpainting technique. Furthermore, a steering kernel regression (SKR) is used to adaptively determine the weightings of local neighbouring pixels to predict high spatial resolution image. Accordingly, the main contributions of this method are twofold. One is to address the land-cover change issues in the spatiotemporal fusion, and the other is to establish an adaptive weighting assignment according to the pixel locations and the radiometric properties of the local neighbours to account for the effect of neighbouring pixels. To validate the proposed method, two actual Enhanced Thematic Mapper Plus (ETM+) and Moderate Resolution Imaging Spectroradiometer (MODIS) acquisitions at southeast China were implemented and compared with the baseline spatial and temporal adaptive reflectance fusion model (STARFM). The experimental results demonstrate that addressing the land-cover changes in spatiotemporal fusion has positive effects on the fused image, and the proposed ISKRFM method significantly outperforms STARFM in terms of both visual and quantitative measurements.     

Combining medium and coarse spatial resolution satellite data to improve the estimation of sub-pixel NDVI time series

The potential of multitemporal coarse spatial resolution remotely sensed images for vegetation monitoring is reduced in fragmented landscapes, where most of the pixels are composed of a mixture of different surfaces. Several approaches have been proposed for the estimation of reflectance or NDVI values of the different land-cover classes included in a low resolution mixed pixel. In this paper, we propose a novel approach for the estimation of sub-pixel NDVI values from multitemporal coarse resolution satellite data. Sub-pixel NDVIs for the different land-cover classes are calculated by solving a weighted linear system of equations for each pixel of a coarse resolution image, exploiting information about within-pixel fractional cover derived from a high resolution land-use map. The weights assigned to the different pixels of the image for the estimation of sub-pixel NDVIs of a target pixel i are calculated taking into account both the spatial distance between each pixel and the target and their spectral dissimilarity estimated on medium-resolution remote-sensing images acquired in different periods of the year. The algorithm was applied to daily and 16-day composite MODIS NDVI images, using Landsat-5 TM images for calculation of weights and accuracy evaluation.Results showed that application of the algorithm provided good estimates of sub-pixel NDVIs even for poorly represented land-cover classes (i.e., with a low total cover in the test area). No significant accuracy differences were found between results obtained on daily and composite MODIS images. The main advantage of the proposed technique with respect to others is that the inclusion of the spectral term in weight calculation allows an accurate estimate of sub-pixel NDVI time series even for land-cover classes characterized by large and rapid spatial variations in their spectral properties.     

An approach based on discrete wavelet transform to unsupervised change detection in multispectral images

Change vector analysis (CVA) and Spectral Angle Mapper (SAM) are widely used for change detection in multitemporal multispectral images. CVA and SAM describe the difference from the perspective of vector magnitude and spectral angle, respectively. It has been proved that three change categories may occur in a changed pixel; however, CVA or SAM alone can only detect two of the three change categories properly. Hence, we propose a novel approach integrating the advantages of them to acquire a better change map. This approach, based on discrete wavelet transform (ABDWT, i.e. approach based on discrete wavelet transform), obtains two difference images by using CVA and SAM, and then yields a novel difference image by fusing them in the coefficients domains of discrete wavelet transform. Experimental results from a simulated and two real data sets validate the effectiveness of the proposed approach. In the first real data set, the proposed approach can identify 14,916 changed pixels while the best result of other methods is 14,806. In the second real data set, the proposed approach detects 3203 changed pixels, while the maximum of other methods is 3189.     

Quantifying processes of land-cover change by remote sensing: Resettlement and rapid land-cover changes in south-eastern Zambia

The objectives of this study are to quantify, based on remote sensing data, processes of land-cover change and to test a Markov-based model to generate short-term land-cover change projections in a region characterised by exceptionally high rates of change. The region of Lusitu, in the Southern Province of Zambia, has been a land-cover change 'hot spot' since the resettlement of 6000 people in the Lusitu area and the succession of several droughts. Land-cover changes were analysed on the basis of a temporal series of three multispectral SPOT images in three steps: (i) land-cover change detection was performed by combining the postclassification and image differencing techniques; (ii) the change detection results were examined in terms of proportion of land-cover classes, change trajectories and spatio-temporal patterns of change; (iii) the process of land-cover change was modelled by a Markov chain to predict land-cover distributions in the near future. The remote sensing approach allowed: (i) to quantify land-cover changes in terms of percentage of area affected and rates of change; (ii) to qualify the nature of changes in terms of impact on natural vegetation; (iii) to map the spatial pattern of land-cover change. 44% of the area has been affected by at least one change in land cover during the period 1986 to 1997. The average annual rate of land-cover change was 4.0%. Agricultural expansion was the dominant change process. Land-cover change trajectories highlighted the dynamic character of changes. The results obtained by applying a Markov chain for projecting future evolutions showed the continuing upward trend of bare soils and cultivated land, and the rapid downward trend of forests and other natural vegetation covers.     

An automated two-step NDVI-based method for the production of low-cost historical burned area map records over large areas

The recognition and understanding of long-term fire-related processes and patterns, such as the possible connection between the increased frequency of wildfires and global warming, requires the study of historical data records. In this study, a methodology was proposed for the automated production of long historical burned area map records over large-scale regions. The methodology was based on remotely sensed, high temporal resolution, normalized difference vegetation index (NDVI) data that could be easily acquired at medium or low spatial resolution. The proposed methodology was used to map the burned areas of the wildfires that occurred over the Peloponnese peninsula, Greece, during the summer of 2007. The method was built upon the NDVI data collected by the Moderate Resolution Imaging Spectroradiometer (MODIS) and Système Pour l’Observation de la Terre (SPOT)-VEGETATION. The higher spatial resolution data of MODIS resulted in higher burned area user accuracy (91.10%) and kappa (0.85) values than the respective ones for VEGETATION (79.29% and 0.77). The majority of classification errors were located along the perimeter of the burned areas and were mainly attributed to spatial resolution limitations of the remotely sensed data. The commission errors located away from the fire perimeter were primarily attributed to topographically shaded areas and land-cover types spectrally similar to burned areas. The omission errors resulted primarily from the small size and elongated shape of remote burned areas. Owing to their geometry, they have a high proportion of mixed pixels, whose spectral properties failed to meet the strict set of criteria for core fire pixels. The benefits of the proposed methodology are maximized when applied to data of the highest available spatial resolution, such as those collected by MODIS and the Project for On-Board Autonomy – Vegetation (PROBA-V) and when land-cover types spectrally similar to burned areas are masked prior to its application.     

The potential use of high-resolution Landsat satellite data for detecting land-cover change in the Greater Horn of Africa

To assess the potential of high-resolution satellite data for land-cover monitoring in the Greater Horn of Africa, we used a regular sampling grid of 170 sites (each measuring 20 km?×?20 km) located at the confluence of the latitudes and meridians across the study area. For each of these sites, Landsat Thematic Mapper (TM) and Enhanced Thematic Mapper (ETM) satellite data were acquired for the years 1990 and 2000. A dot grid visual interpretation was used to assess land-cover change between the two dates in each of the sites. With only two acquisition dates, we found that these data were insufficient for accurately determining land-cover change and degradation in arid areas where non-woody biomass dominates. We were nevertheless able to detect land-cover modifications in three areas: increases in agriculture on the coastal plain near Mogadishu, increases in agriculture at the western fringes of our study area where there is higher rainfall, and finally a reduction in woodlands and shrublands in areas close to refugee camps on the Somali–Kenya border.     

Improving super-resolution mapping through combining multiple super-resolution land-cover maps

Super-resolution mapping (SRM) is an ill-posed problem, and different SRM algorithms may generate non-identical fine-spatial resolution land-cover maps (sub-pixel maps) from the same input coarse-spatial resolution image. The output sub-pixels maps may each have differing strengths and weaknesses. A multiple SRM (M-SRM) method that combines the sub-pixel maps obtained from a set of SRM analyses, obtained from a single or multiple set of algorithms, is proposed in this study. Plurality voting, which selects the class with the most votes, is used to label each sub-pixel. In this study, three popular SRM algorithms, namely, the pixel-swapping algorithm (PSA), the Hopfield neural network (HNN) algorithm, and the Markov random field (MRF)-based algorithm, were used. The proposed M-SRM algorithm was validated using two data sets: a simulated multispectral image and an Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) hyperspectral image. Results show that the highest overall accuracies were obtained by M-SRM in all experiments. For example, in the AVIRIS image experiment, the highest overall accuracies of PSA, HNN, and MRF were 88.89, 93.81, and 82.70%, respectively, and these increased to 95.06, 95.37, and 85.56%, respectively for M-SRM obtained from the multiple PSA, HNN, and MRF analyses.     

Super-resolution land-cover mapping using multiple sub-pixel shifted remotely sensed images

Super-resolution land-cover mapping is a promising technology for prediction of the spatial distribution of each land-cover class at the sub-pixel scale. This distribution is often determined based on the principle of spatial dependence and from land-cover fraction images derived with soft classification technology. However, the resulting super-resolution land-cover maps often have uncertainty as no information about sub-pixel land-cover patterns within the low-resolution pixels is used in the model. Accuracy can be improved by incorporating supplemental datasets to provide more land-cover information at the sub-pixel scale; but the effectiveness of this is limited by the availability and quality of these additional datasets. In this paper, a novel super-resolution land-cover mapping technology is proposed, which uses multiple sub-pixel shifted remotely sensed images taken by observation satellites. These satellites take images over the same area once every several days, but the images are not identical because of slight orbit translations. Low-resolution pixels in these remotely sensed images therefore contain different land-cover fractions that can provide useful information for super-resolution land-cover mapping. We have constructed a Hopfield Neural Network (HNN) model to solve it. Maximum spatial dependence is the goal of the proposed model, and the fraction maps of all images are constraints added to the energy function of HNN. The model was applied to synthetic artificial images as well as to a real degraded QuickBird image. The output maps derived from different numbers of images at different zoom factors were compared visually and quantitatively to the super-resolution map generated from a single image. The resulting land-cover maps with multiple remotely sensed images were more accurate than was the single image map. The use of multiple remotely sensed images is therefore a promising method for decreasing the uncertainty of super-resolution land-cover mapping. Moreover, remotely sensed images with similar spatial resolution from different satellite platforms can be used together, allowing a fusion of information obtained from remotely sensed imagery.     

Super-resolution mapping using Hopfield Neural Network with panchromatic imagery

Land-cover proportions of mixed pixels can be predicted using soft classification. From the land-cover proportions, a hard land-cover map can be predicted at sub-pixel spatial resolution using super-resolution mapping techniques. It has been demonstrated that the Hopfield Neural Network (HNN) provides a suitable method for super-resolution mapping. To increase the detail and accuracy of the sub-pixel land-cover map, supplementary information at an intermediate spatial resolution can be used. In this research, panchromatic (PAN) imagery was used as an additional source of information for super-resolution mapping. Information from the PAN image was captured by a new PAN reflectance constraint in the energy function of the HNN. The value of the new PAN reflectance constraint was defined based on forward and inverse models with local end-member spectra and local convolution weighting factors. Two sets of simulated and degraded data were used to test the new technique. The results indicate that PAN imagery can be used as a source of supplementary information to increase the detail and accuracy of sub-pixel land-cover maps produced by super-resolution mapping from land-cover proportion images.     

A hypergraph-based context-sensitive representation technique for VHR remote-sensing image change detection

This article presents a novel semi-supervised change detection approach for very-high-resolution (VHR) remote-sensing images. The proposed approach aims at extracting the change information by making full use of the context-sensitive relationships among pixels in the images. This is accomplished via a context-sensitive image representation technique based on hypergraph model. First, each temporal image is modelled as a hypergraph that utilizes a set of hyperedges to capture the context-sensitive properties of pixels in the image. Second, the difference in the bi-temporal images is measured by both the similarity and the consistency between the two hypergraphs. Finally, the changes are separated from the unchanged ones by a hypergraph-based semi-supervised classifier on the difference image. Experimental results obtained on different VHR remote-sensing data sets demonstrate the effectiveness of the proposed approach.     

线性模型的遥感图像时空融合

目的 时空融合是解决当前传感器无法兼顾遥感图像的空间分辨率和时间分辨率的有效方法。在只有一对精细-粗略图像作为先验的条件下，当前的时空融合算法在预测地物变化时并不能取得令人满意的结果。针对这个问题，本文提出一种基于线性模型的遥感图像时空融合算法。方法 使用线性关系表示图像间的时间模型，并假设时间模型与传感器无关。通过分析图像时间变化的客观规律，对模型进行全局和局部约束。此外引入一种多时相的相似像素搜寻策略，更灵活地选取相似像素，消除了传统算法存在的模块效应。结果 在两个数据集上与STARFM（spatial and temporal adaptive reflectance fusion model）算法和FSDAF（flexible spatiotemporal data fusion）算法进行比较，实验结果表明，在主要发生物候变化的第1个数据集，本文方法的相关系数CC（correlation coefficient）分别提升了0.25%和0.28%，峰值信噪比PSNR（peak signal-to-noise ratio）分别提升了0.153 1 dB和1.379 dB，均方根误差RMSE（root mean squared error）分别降低了0.05%和0.69%，结构相似性SSIM（structural similarity）分别提升了0.79%和2.3%。在发生剧烈地物变化的第2个数据集，本文方法的相关系数分别提升了6.64%和3.26%，峰值信噪比分别提升了2.086 0 dB和2.510 7 dB，均方根误差分别降低了1.45%和2.08%，结构相似性分别提升了11.76%和11.2%。结论 本文方法根据时间变化的特点，对时间模型进行优化，同时采用更加灵活的相似像素搜寻策略，收到了很好的效果，提升了融合结果的准确性。     

Vectorial boundary-based sub-pixel mapping method for remote-sensing imagery

This article presents a vectorial boundary-based sub-pixel mapping (VBSPM) method to obtain the land-cover distribution with finer spatial resolution in mixed pixels. With inheritance from the geometric SPM (GSPM), VBSPM first geometrically partitions a mixed pixel using polygons, and then utilizes a vectorial boundary extraction model (VBEM), rather than the rasterization method in GSPM, to determine the location and length of each edge in the polygon, while these edges are located at the boundary of and within the interior of the mixed pixel. Furthermore, VBSPM uses a decay function to manage the mixed pixels along the image boundary region due to the missing parts of their neighbours. Finally, a ray-crossing algorithm is employed to determine the land-cover class of each sub-pixel in terms of vectorial boundaries. The experiments with artificial and remotely sensed images have demonstrated that VBSPM can reduce the inconsistency between the boundaries of different land-cover classes, approximately calculating errors with an odd zoom factor, and achieve more accurate sub-pixel mapping results than the hard classification methods and GSPM.     

Requirements for labelling forest polygons in an object-based image analysis classification

The ability to spatially quantify changes in the landscape and create land-cover maps is one of the most powerful uses of remote sensing. Recent advances in object-based image analysis (OBIA) have also improved classification techniques for developing land-cover maps. However, when using an OBIA technique, collecting ground data to label reference units may not be straightforward, since these segments generally contain a variable number of pixels as well as a variety of pixel values, which may reflect variation in land-cover composition. Accurate classification of reference units can be particularly difficult in forested land-cover types, since these classes can be quite variable on the ground. This study evaluates how many prism sample locations are needed to attain an acceptable level of accuracy within forested reference units in southeastern New Hampshire (NH). Typical forest inventory guidelines suggest at least 10 prism samples per stand, depending on the stand area and stand type. However, because OBIA segments group pixels based on the variance of the pixels, fewer prism samples may be necessary in a segment to properly estimate the stand composition. A bootstrapping statistical technique was used to find the necessary number of prism samples to limit the variance associated with estimating the species composition of a segment. Allowing for the lowest acceptable variance, a maximum of only six prism samples was necessary to label forested reference units. All polygons needed at least two prism samples for classification.     

Detecting temporal changes in the extent of annual flooding within the Cambodia and the Vietnamese Mekong Delta from MODIS time-series imagery

This paper presents the methodology used to detect temporal changes in the extent of annual flooding within the Cambodia and the Vietnamese Mekong Delta (VMD) based on MODIS time-series imagery (Wavelet-based Filter for detecting spatio-temporal changes in Flood Inundation; WFFI). This methodology involves the use of a wavelet-based filter to interpolate missing information and reduce the noise component in the time-series data, as proposed in a previous study. The smoothed time profiles of Enhanced Vegetation Index (EVI), Land Surface Water Index (LSWI), and the Difference Value between EVI and LSWI (DVEL) are obtained from MOD09 8-day composite time-series data (resolution: 500 m; time period: 2000-2005). The proposed algorithm was applied to produce time-series inundation maps (WFFI products) for the five annual flood seasons over the period from 2000 to 2004. The WFFI products were validated via comparisons with Landsat-derived results and inundation maps based on RADARSAT images, hydrological data, and digital elevation model data. Compared with the RADARSAT-derived inundation maps at the province level, the obtained RMSE range from 364 to 443 km² and the determination coefficients [R²] range from 0.89 to 0.92. Compared with Landsat-derived results at the 10-km grid level, the obtained RMSE range from 6.8 to 15.2 km² and the determination coefficients [R²] range from 0.77 to 0.97. The inundated area of flooded forests/marsh to the northeast of Tonle Sap Lake were underestimated, probably because of extensive vegetation cover in this area. The spatial characteristics of the estimated start dates, end dates, and duration of inundation cycles were also determined for the period from 2000 to 2004. There are clear contrasts in the distribution of the estimated end dates and duration of inundation cycles between large-scale floods (2000-2002) and medium- and small-scale floods (2003 and 2004). At the regional scale, the estimated start dates for the southern part of An Giang Province during 2003 and 2004 was distinctly later than that for surrounding areas. The results indicate that these triple-cropping areas enclosed by dikes increased in extent from 2003 to 2004. In contrast, the estimated end dates of inundation at the Co Do and Song Hau State Farms were clearly earlier than those for surrounding areas, although the estimated start dates were similar. Temporal changes in the inundation area of Flood pixels in the Dong Thap and Long An Provinces are in excellent agreement with daily water-level data recorded at Tan Chau Station. The estimated area of Long-term water body increased in size from 2000 to 2004, especially in coastal areas of the Ca Mau and Bac Lieu Provinces. Statistical data for Vietnam indicate that this trend may reflect the expansion of shrimp-farming areas. The WFFI products enable an understanding of seasonal and annual changes in the water distribution and environment of the Cambodia and the VMD from a global viewpoint.     

Mapping urban land cover based on spatial-spectral classification of hyperspectral remote-sensing data

In this article, an innovative classification framework for hyperspectral image data, based on both spectral and spatial information, is proposed. The main objective of this method is to improve the accuracy and efficiency of high-resolution land-cover mapping in urban areas. The spatial information is obtained by an enhanced marker-based minimum spanning forest (MMSF) algorithm. A pixel-based support vector machine (SVM) algorithm is first used to classify the hyperspectral image data, then the enhanced MMSF algorithm is applied in order to increase the accuracy of less accurately classified land-cover types. The enhanced MMSF algorithm is used as a binary classifier. These two classes are the low-accuracy class and remaining classes. Finally, the SVM algorithm is trained for classes with acceptable accuracy. In the proposed approach, namely MSF-SVM, the markers are extracted from the classification maps obtained by both SVM and watershed segmentation algorithms, and are then used to build the MSF. Three benchmark hyperspectral data sets are used for the assessment: Berlin, Washington DC Mall, and Quebec City. Experimental results demonstrate the superiority of the proposed approach compared with SVM and the original MMSF algorithms. It achieves approximately 5, 6, and 7% higher rates in kappa coefficients of agreement in comparison with the original MMSF algorithm for the Berlin, Washington DC Mall, and Quebec City data sets, respectively.     

An image fusion method based on object-oriented classification

In image fusion of different spatial resolution multispectral (MS) and panchromatic (PAN) images, a spectrally mixed MS pixel superimposes multiple mixed PAN pixels and multiple pure PAN pixels. This verifies that with increased spatial resolution in imaging, a low spatial resolution spectrally mixed subpixel may be unmixed to be a pure pixel. However, spectral unmixing of mixed MS subpixels is rarely considered in current remote-sensing image fusion methods, resulting in blurred fused images. In the image fusion method proposed in this article, such spectral unmixing is realized. In this method, the MS and PAN images are jointly segmented into image objects, image objects are classified to obtain a classification map of the PAN image and each MS subpixel is fused to be a pixel matching the class of the corresponding PAN pixel. Tested on spatially degraded IKONOS MS and PAN images with a significant spatial resolution ratio of 8:1, the fusion method offered fused images with high spectral quality and deblurred visualization.     

Object-based image analysis for operational fine-scale regional mapping of land cover within river corridors from multispectral imagery and thematic data

Accurate mapping of land-cover diversity within riparian areas at a regional scale is a major challenge for better understanding the influence of riparian landscapes and related natural and anthropogenic pressures on river ecological status. As the structure (composition and spatial organization) of riparian area land cover (RALC) is generally not accessible using moderate-scale satellite imagery, finer spatial resolution imagery and specific mapping techniques are needed. For this purpose, we developed a classification procedure based on a specific multiscale object-based image analysis (OBIA) scheme dedicated to producing fine-scale and reliable RALC maps in different geographical contexts (relief, climate and geology). This OBIA scheme combines information from very high spatial resolution multispectral imagery (satellite or airborne) and available spatial thematic data using fuzzy expert knowledge classification rules. It was tested over the Hérault River watershed (southern France), which presents contrasting landscapes and a total stream length of 1150 km, using the combination of SPOT (Système Probatoire d'Observation de la Terre) 5 XS imagery (10 m pixels), aerial photography (0.5 m pixels) and several national spatial thematic data. A RALC map was produced (22 classes) with an overall accuracy of 89% and a kappa index of 83%, according to a targeted land-cover pressures typology (six categories of pressures). The results of this experimentation demonstrate that the application of OBIA to multisource spatial data provides an efficient approach for the mapping and monitoring of RALC that can be implemented operationally at a regional or national scale. We further analysed the influence of map resolution on the quantification of riparian spatial indicators to highlight the importance of such data for studying the influence of landscapes on river ecological status at the riparian scale.     

Classification of Landsat images based on spectral and topographic variables for land-cover change detection in Zagros forests

Detection of land-cover changes through time can be complicated because of sensor-specific differences in spatial and spectral resolutions; classified land-cover changes can be due to either real changes on the ground or a switch in sensors used to collect data. This study focused on two objectives: (1) selecting the best predictor variables for the classification of semi-arid Zagros forests given the characteristics of the study area and available data sets and (2) evaluating the application of the random forest (RF) algorithm as a unified technique for the classification of data sets acquired from different sensors. Three images of the same study area were acquired from the Landsat-5 Thematic Mapper (TM) sensor in 2009, the Landsat-7 Enhanced Thematic Mapper (ETM+) sensor with Scan Line Corrector (SLC) in 1999 and the Landsat-2 Multispectral Scanner (MSS) sensor in 1975. Following image preprocessing, the RF algorithm was applied for variable selection and classification. A test of equivalence was used to compare the overall accuracy of the classified maps from the three sensors. Slope, normalized difference vegetation index (NDVI) and elevation were determined to be the most important predictor variables for all three images. High overall classification accuracies were achieved for all three images (97.90% for MSS, 95.43% for TM and 95.29% for ETM). The ETM- and TM-derived maps had equivalent overall accuracy and even significantly higher overall accuracy was obtained for the MSS-derived map. The post-classification comparison showed an increase in agriculture and a decrease in forest cover. The selected predictor variables were consistent with ecological reality and showed more details on the changes of the land-cover classes across biophysical variables of the study area through time.     

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.