Large Area Land Cover Mapping Research Articles

Classifying raw irregular time series (CRIT) for large area land cover mapping by adapting transformer model

For Landsat land cover classification, the time series observations are typically irregular in the number of observations in a period (e.g., a year) and acquisition dates due to cloud cover variations over large areas and acquisition plan variations over long periods. Compositing or temporal percentile calculation are usually used to transform the irregular time series to regular temporal variables so that the machine and deep learning classifiers can be applied. Recognizing that the composite and percentile calculations have information loss, this study presents a method directly Classifying the Raw Irregular Time series (CRIT) (‘raw’ means irregular good-quality surface reflectance time series without any composite or temporal percentile derivation) by adapting Transformer. CRIT uses the acquisition day of year as classification input to align time series and also takes the Landsat satellite platform (Landsat 5, 7 and 8) as input to address the inter-sensor reflectance differences.The CRIT was demonstrated by classifying Landsat analysis ready data (ARD) surface reflectance time series acquired across one year for three years (1985, 2006 and 2018) over the Conterminous United States (CONUS) with both spatial and temporal variations in Landsat availability. 20,047 training and 4949 evaluation 30-m pixel were used where each pixel was annotated as one of seven land cover classes for each year. The CRIT was compared with classifying 16-day composite time series and temporal percentiles and compared with a 1D convolution neural network (CNN) method. Results showed that the CRIT trained with three years of samples had 1.4–1.5% higher overall accuracies with less computation time than classifying 16-day composites and 2.3–2.4% higher than classifying temporal percentiles. The CRIT advantages over 16-day composites were pronounced for developed (0.05 F1-score) and cropland (0.02 F1-score) classes and for mixed or boundary pixels. This was reasonable as the 16-day composites had only on average 7.02, 16.49 and 15.78 good quality observations for the three years, respectively, in contrast to 7.89, 27.72, and 26.60 for the raw irregular time series. The CNN was not as good as CRIT in classifying the raw irregular time series as CNN simply filling temporal positions with no observations as zeros while the CRIT used a masking mechanism to rule out their contribution. The CRIT can also take the pixel coordinates and DEM variables as input which further increased the overall accuracies by 1.1–2.6% and achieved 84.33%, 87.54% and 87.01% overall accuracies for the 1985, 2006 and 2018 classifications, respectively. The CRIT land cover maps were shown consistent with the USGS Land Change Monitoring, Assessment, and Projection (LCMAP) maps. The developed codes, training data and maps were made publicly available.

A High-Performance Automated Large-Area Land Cover Mapping Framework

Land cover mapping plays a pivotal role in global resource monitoring, sustainable development research, and effective management. However, the complexity of the mapping process, coupled with significant computational and data storage requirements, often leads to delays between data processing and product publication, thereby bringing challenges to creating multi-timesteps large-area products for monitoring dynamic land cover. Therefore, improving the efficiency of each stage in land cover mapping and automating the mapping process is currently an urgent issue to be addressed. This study proposes a high-performance automated large-area land cover mapping framework (HALF). By leveraging Docker and workflow technologies, the HALF effectively tackles model heterogeneity in complex land cover mapping processes, thereby simplifying model deployment and achieving a high degree of decoupling between production models. It optimizes key processes by incorporating high-performance computing techniques. To validate these methods, this study utilized Landsat imagery data and extracted samples using GLC_FCS and FROM_GLC, all of which were acquired at a spatial resolution of 30 m. Several 10° × 10° regions were chosen globally to illustrate the viability of generating large-area land cover using the HALF. In the sample collection phase, the HALF introduced an automated method for generating samples, which overlayed multiple prior products to generate a substantial number of samples, thus saving valuable manpower resources. Additionally, the HALF utilized high-performance computing technology to enhance the efficiency of the sample–image matching phase, thereby achieving a speed that was ten times faster than traditional matching methods. In the mapping stage, the HALF employed adaptive classification models to train the data in each region separately. Moreover, to address the challenge of handling a large number of classification results in a large area, the HALF utilized a parallel mosaicking method for classification results based on the concept of grid division, and the average processing time for a single image was approximately 6.5 s.

Improving 3-m Resolution Land Cover Mapping through Efficient Learning from an Imperfect 10-m Resolution Map

Substantial progress has been made in the field of large-area land cover mapping as the spatial resolution of remotely sensed data increases. However, a significant amount of human power is still required to label images for training and testing purposes, especially in high-resolution (e.g., 3-m) land cover mapping. In this research, we propose a solution that can produce 3-m resolution land cover maps on a national scale without human efforts being involved. First, using the public 10-m resolution land cover maps as an imperfect training dataset, we propose a deep learning based approach that can effectively transfer the existing knowledge. Then, we improve the efficiency of our method through a network pruning process for national-scale land cover mapping. Our proposed method can take the state-of-the-art 10-m resolution land cover maps (with an accuracy of 81.24% for China) as the training data, enable a transferred learning process that can produce 3-m resolution land cover maps, and further improve the overall accuracy (OA) to 86.34% for China. We present detailed results obtained over three mega cities in China, to demonstrate the effectiveness of our proposed approach for 3-m resolution large-area land cover mapping.

Discriminating treed and non-treed wetlands in boreal ecosystems using time series Sentinel-1 data

Wetlands are recognized for their importance to a range of ecosystem goods and services; however, detailed information on wetland presence, type, extent, and persistence is challenging to attain over large areas and/or long time periods due to the spatial complexity and temporal dynamism of wetlands. In this study we explored the potential for within-year time series of C-band Synthetic Aperture Radar (SAR) observations from the free and open Sentinel-1 data archive to improve discrimination of treed and non-treed wetlands and non-wetlands in a boreal forest environment. Through a set of 3843 classification experiments for the year 2017, we tested the influence of three factors on classification accuracy: (i) input features (two backscatter coefficients in VV and VH polarization (σVV and σVH) and four quantitative measures derived from the Stokes vector); (ii) the temporal form of features (i.e. using all within-year observations versus generalized measures such as monthly/seasonal means or annualized statistics); and (iii) missing observations in Sentinel-1 time series due to varying observation availability across space. Among the tested features, we found the greatest utility in σVV and σVH. Directly using all within-year observations yielded higher accuracy than using generalized temporal forms. Moreover, the temporal form of the features had a greater impact on classification accuracy than the features themselves. The highest overall accuracy (0.860 ± 0.002) was achieved using σVV and σVH from all within-year observations. The majority of class confusion occurred between treed wetlands and non-wetlands. We found no significant reduction in the overall accuracy by simulated missing observations in time series when using all within-year observations. With the increasing availability of free and open data from the Sentinel-1 archive, new opportunities are emerging to readily integrate within-year time series into large-area land cover mapping, particularly if analysis-ready SAR data products further reduce preprocessing requirements for end users.

Utilizing Sentinel-1A Radar Images for Large-Area Land Cover Mapping with Machine-learning Methods

Land use and land cover maps are vital sources of information for many applications. Recently, using high-resolution and open-access satellite images has become a preferred method for mapping land cover, especially over large areas. This study was designed to map the land cover and agricultural fields of a large area using Sentinel-1A synthetic aperture radar (SAR) images. Seven machine-learning methods were employed for image analyses. The Random Forest classifier algorithm outperformed the other machine-learning methods in the training step; thus, we selected it for further use and tuned its parameters. After several image processing steps, we classified the final image into 23 land cover classes and achieved an overall accuracy of 42% for all classes, and 57% for agricultural fields. This research note highlights some characteristics and advantages of using Sentinel-1A images and provides novel methods for nation-wide large-area mapping applications.

Spectral-spatial multi-feature classification of remote sensing big data based on a random forest classifier for land cover mapping

Supplementary information, such as multi-temporal spectral data and textural features, has the potential to improve land cover classification accuracy. However, given the larger volumes of remote sensing data, it is difficult to utilize all the features of remote sensing big data having different times and spatial resolutions. Inefficiency is also a large problem when dealing with large area land cover mapping. In this study, a new mode of incorporating spatial and temporal dependencies in a complex region employing the random forests (RFs) classifier was utilized. To map land covers, spring and autumn spectral images and their spectral indexes, textural features obtained from Landsat 5 were selected, and an importance measure variable was used to reduce the data’s dimension. In addition to randomly selecting the variable, we used random sampling to furthest decrease the generalization error in RF. The results showed that utilizing random sampling, multi-temporal spectral image and texture features, the classification of the Wuhan urban agglomeration, China, using RF performed well. The RF algorithm yielded an overall accuracy of 89.2% and a Kappa statistic of 0.8522, indicating high model performance. In addition, the variable importance measures demonstrated that the type of textural features was extremely important for intra-class separability. The RF model has transitivity. The algorithm can be extended by choosing a set of appropriate features for signature extension over large areas or in time-series of Landsat imagery. Land cover mapping might be more economical and efficient if no-cost imagery is used.

Exploring diversity in ensemble classification: Applications in large area land cover mapping

Ensemble classifiers, such as random forests, are now commonly applied in the field of remote sensing, and have been shown to perform better than single classifier systems, resulting in reduced generalisation error. Diversity across the members of ensemble classifiers is known to have a strong influence on classification performance - whereby classifier errors are uncorrelated and more uniformly distributed across ensemble members. The relationship between ensemble diversity and classification performance has not yet been fully explored in the fields of information science and machine learning and has never been examined in the field of remote sensing. This study is a novel exploration of ensemble diversity and its link to classification performance, applied to a multi-class canopy cover classification problem using random forests and multisource remote sensing and ancillary GIS data, across seven million hectares of diverse dry-sclerophyll dominated public forests in Victoria Australia. A particular emphasis is placed on analysing the relationship between ensemble diversity and ensemble margin - two key concepts in ensemble learning. The main novelty of our work is on boosting diversity by emphasizing the contribution of lower margin instances used in the learning process. Exploring the influence of tree pruning on diversity is also a new empirical analysis that contributes to a better understanding of ensemble performance. Results reveal insights into the trade-off between ensemble classification accuracy and diversity, and through the ensemble margin, demonstrate how inducing diversity by targeting lower margin training samples is a means of achieving better classifier performance for more difficult or rarer classes and reducing information redundancy in classification problems. Our findings inform strategies for collecting training data and designing and parameterising ensemble classifiers, such as random forests. This is particularly important in large area remote sensing applications, for which training data is costly and resource intensive to collect.

A web-based system for supporting global land cover data production

Global land cover (GLC) data production and verification process is very complicated, time consuming and labor intensive, requiring huge amount of imagery data and ancillary data and involving many people, often from different geographic locations. The efficient integration of various kinds of ancillary data and effective collaborative classification in large area land cover mapping requires advanced supporting tools. This paper presents the design and development of a web-based system for supporting 30-m resolution GLC data production by combining geo-spatial web-service and Computer Support Collaborative Work (CSCW) technology. Based on the analysis of the functional and non-functional requirements from GLC mapping, a three tiers system model is proposed with four major parts, i.e., multisource data resources, data and function services, interactive mapping and production management. The prototyping and implementation of the system have been realised by a combination of Open Source Software (OSS) and commercially available off-the-shelf system.This web-based system not only facilitates the integration of heterogeneous data and services required by GLC data production, but also provides online access, visualization and analysis of the images, ancillary data and interim 30m global land-cover maps. The system further supports online collaborative quality check and verification workflows. It has been successfully applied to China’s 30-m resolution GLC mapping project, and has improved significantly the efficiency of GLC data production and verification. The concepts developed through this study should also benefit other GLC or regional land-cover data production efforts.

A Pixel-Based Landsat Compositing Algorithm for Large Area Land Cover Mapping

Information on the changing land surface is required at high spatial resolutions as many processes cannot be resolved using coarse resolution data. Deriving such information over large areas for Landsat data, however, still faces numerous challenges. Image compositing offers great potential to circumvent such shortcomings. We here present a compositing algorithm that facilitates creating cloud free, seasonally and radiometrically consistent datasets from the Landsat archive. A parametric weighting scheme allows for flexibly utilizing different pixel characteristics for optimized compositing. We describe in detail the development of three parameter decision functions: acquisition year, day of year and distance to clouds. Our test site covers 42 Landsat footprints in Eastern Europe and we produced three annual composites. We evaluated seasonal and annual consistency and compared our composites to BRDF normalized MODIS reflectance products. Finally, we also evaluated how well the composites work for land cover mapping. Results prove that our algorithm allows for creating seasonally consistent large area composites. Radiometric correspondence to MODIS was high (up to R <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> > 0.8), but varied with land cover configuration and selected image acquisition dates. Land cover mapping yielded promising results (overall accuracy 72%). Class delineations were regionally consistent with minimal effort for training data. Class specific accuracies increased considerably (~10%) when spectral metrics were incorporated. Our study highlights the value of compositing in general and for Landsat data in particular, allowing for regional to global LULCC mapping at high spatial resolutions.

Integrating geographical data and phenological characteristics derived from MODIS data for improving land cover mapping

The study developed a feasible method for large-area land cover mapping with combination of geographical data and phenological characteristics, taking Northeast China (NEC) as the study area. First, with the monthly average of precipitation and temperature datasets, the spatial clustering method was used to divide the NEC into four ecoclimate regions. For each ecoclimate region, geographical variables (annual mean precipitation and temperature, elevation, slope and aspect) were combined with phenological variables derived from the moderate resolution imaging spectroradiometer (MODIS) data (enhanced vegetation index (EVI) and land surface water index (LSWI)), which were taken as input variables of land cover classification. Decision Tree (DT) classifiers were then performed to produce land cover maps for each region. Finally, four resultant land cover maps were mosaicked for the entire NEC (NEC_MODIS), and the land use and land cover data of NEC (NEC_LULC) interpreted from Landsat-TM images was used to evaluate the NEC_MODIS and MODIS land cover product (MODIS_IGBP) in terms of areal and spatial agreement. The results showed that the phenological information derived from EVI and LSWI time series well discriminated land cover classes in NEC, and the overall accuracy was significantly improved by 5.29% with addition of geographical variables. Compared with NEC_LULC for seven aggregation classes, the area errors of NEC_MODIS were much smaller and more stable than that of MODIS_IGBP for most of classes, and the wall-to-wall spatial comparisons at pixel level indicated that NEC_MODIS agreed with NEC_LULC for 71.26% of the NEC, whereas only 62.16% for MODIS_IGBP. The good performance of NEC_MODIS demonstrates that the methodology developed in the study has great potential for timely and detailed land cover mapping in temperate and boreal regions.

Developing Consistent Landsat Data Sets for Large Area Applications: The MRLC 2001 Protocol

One of the major efforts in large area land cover mapping over the last two decades was the completion of two U.S. National Land Cover Data sets (NLCD), developed with nominal 1992 and 2001 Landsat imagery under the auspices of the MultiResolution Land Characteristics (MRLC) Consortium. Following the successful generation of NLCD 1992, a second generation MRLC initiative was launched with two primary goals: (1) to develop a consistent Landsat imagery data set for the U.S. and (2) to develop a second generation National Land Cover Database (NLCD 2001). One of the key enhancements was the formulation of an image preprocessing protocol and implementation of a consistent image processing method. The core data set of the NLCD 2001 database consists of Landsat 7 Enhanced Thematic Mapper Plus (ETM+) images. This letter details the procedures for processing the original ETM+ images and more recent scenes added to the database. NLCD 2001 products include Anderson Level II land cover classes, percent tree canopy, and percent urban imperviousness at 30-m resolution derived from Landsat imagery. The products are freely available for download to the general public from the MRLC Consortium Web site at http://www.mrlc.gov.

Landsat continuity: Issues and opportunities for land cover monitoring

Initiated in 1972, the Landsat program has provided a continuous record of earth observation for 35 years. The assemblage of Landsat spatial, spectral, and temporal resolutions, over a reasonably sized image extent, results in imagery that can be processed to represent land cover over large areas with an amount of spatial detail that is absolutely unique and indispensable for monitoring, management, and scientific activities. Recent technical problems with the two existing Landsat satellites, and delays in the development and launch of a successor, increase the likelihood that a gap in Landsat continuity may occur. In this communication, we identify the key features of the Landsat program that have resulted in the extensive use of Landsat data for large area land cover mapping and monitoring. We then augment this list of key features by examining the data needs of existing large area land cover monitoring programs. Subsequently, we use this list as a basis for reviewing the current constellation of earth observation satellites to identify potential alternative data sources for large area land cover applications. Notions of a virtual constellation of satellites to meet large area land cover mapping and monitoring needs are also presented. Finally, research priorities that would facilitate the integration of these alternative data sources into existing large area land cover monitoring programs are identified. Continuity of the Landsat program and the measurements provided are critical for scientific, environmental, economic, and social purposes. It is difficult to overstate the importance of Landsat; there are no other systems in orbit, or planned for launch in the short-term, that can duplicate or approach replication, of the measurements and information conferred by Landsat. While technical and political options are being pursued, there is no satellite image data stream poised to enter the National Satellite Land Remote Sensing Data Archive should system failures occur to Landsat-5 and -7.

Sampling methods for validation of large area land cover mapping

Sampling methods for validation of large area land cover mapping

An approach for edge matching large-area satellite image classifications

Large-area land cover mapping based on remotely sensed data often requires combining individual or large groups of classified images to produce final map products. Operational and logistical considerations are typically confronted when classifying medium spatial resolution satellite imagery (i.e., Landsat), with the mapping partitioned by spectral, ecological, or political considerations, or combinations thereof. Visual discontinuities can emerge at the locations where logistically based production zones join. Transparent and systematic approaches for addressing discontinuities are desired for the Earth Observation for Sustainable Development of Forests (EOSD) project. This large-area land cover mapping project is producing map products for Canada's forested ecozones. A distributed implementation plan, largely based on grouping provincial and territorial political units, was followed for production. Scene-to-scene discontinuities are rare within each production zone and are primarily related to image acquisition date and phenological state. In contrast, discontinuities can emerge at the production zone boundaries because of differences in support data available or more commonly because of differences in the attribution of density classes. Of the over 475 scenes classified for the EOSD project, it is estimated that fewer than 30 (about 6.3% of total) will require processing to minimize the cross-boundary discontinuity. Options for mitigating the discontinuities are described and demonstrated in the context of different scenarios of overlap found along the EOSD production zone boundaries (complete overlap, partial overlap, and no overlap) using two subsets of a Landsat scene along the shared provincial border between British Columbia and Alberta, Canada. Analysis of image gradients provides a quantitative basis for identification of discontinuities and also relates the results of the likelihood-based relabelling process. Through this process, only density descriptors of cover types are altered, largely maintaining classification integrity. The process as presented is generic and is suitable for addressing edge discontinuities that can emerge when undertaking a large-area land cover classification project.

Inclusion of topographic variables in an unsupervised classification of satellite imagery

Unsupervised classification has emerged as a common method for mapping the land cover of large areas with satellite data. Typically, clusters generated by an unsupervised algorithm, such as the K-means, are merged and labelled using a combination of manual and automated methods. Topographic shadows found in areas of high relief, particularly in areas with low sun angles, increase the complexity of land cover classification, as a single land cover class may have very different spectral responses in shadowed and nonshadowed areas of the image. Methods to increase classification accuracies in areas with severe topographic shadows are required for Canadian large area land cover mapping projects. In a standard supervised classification, increases in land cover map classification accuracy have been obtained by including topographic attributes as inputs to the classification algorithm. In this study we investigate the potential of such data as a means to increase the accuracy of an unsupervised land cover classification in a high-relief area in central British Columbia, Canada. Separate datasets were used in cluster labelling and accuracy assessment. Four classification trials were completed: (i) using a standard approach without the addition of topographic attributes, (ii) using elevation data as an additional input to (i), (iii) prestratifying the image into shadow and nonshadow areas prior to clustering, and (iv) combining the methods used in (ii) and (iii). The latter provided the highest level of overall attribute accuracy at 80.1%, with a 95% confidence interval of 73.6%–88.6%. This is an improvement over the standard approach, which produced an overall attribute accuracy of 68.7%, with a 95% confidence interval of 59.8%–77.6%. We concluded that the prestratification of an image into areas of shadow and nonshadow prior to clustering in conjunction with the use of elevation data as an input to the clustering process is a practical method to increase classification accuracy in areas of high relief where topographic shadow is problematic.

Operational mapping of the land cover of the forested area of Canada with Landsat data: EOSD land cover program

A priority of the Canadian Forest Service and Canadian Space Agency joint project, Earth Observation for Sustainable Development of Forests (EOSD), is the production of a land cover map of the forested area of Canada based upon Landsat data. The land cover will be produced through a partnership of federal, provincial and territorial governments, universities, and industry. The short-term goal of EOSD is to complete a land cover map representing year 2000 forested area conditions by early 2006. Over the longer term, EOSD will aim to produce land cover products to capture changes in forest conditions over time to support national and international reporting requirements. The forested area of Canada represents approximately half of Canada's landmass, requiring over 450 scenes for complete coverage (with overlap minimized). EOSD is working with provincial and territorial mapping agencies that have on-going land cover mapping programs to optimize production capacity. It is envisioned that the combined output of EOSD and provincial and territorial land cover mapping programs will be integrated with maps developed by other sectors and agencies (such as agriculture) to produce a complete representation of the land cover of Canada. Large-area land cover mapping using remote sensing is a relatively new phenomenon. Advances in data storage capabilities, computing power, and increases in the affordability of data have allowed for large-area projects to be undertaken in ways previously not possible. The manner in which a large-area mapping project is approached is related to a number of factors including the spatial extent of the area of interest, the spatial resolution of the selected sensor, and the products which are to be generated. In this communication we report on the strategy, methods, and status of the EOSD land cover mapping program of the forested area of Canada. Key words: Canada, land cover, forest inventory, EOSD, Landsat, unsupervised classification, NFI

Contextual correction: techniques for improving land cover mapping from remotely sensed images

Large area land cover mapping is an important application of remote sensing. A digital land cover map of Great Britain has recently been compiled by supervised classification of Landsat Thematic Mapper data. The work has involved development of a range of post classification procedures to correct contextual errors associated with the use of spectral classification algorithms. This paper describes these procedures and examines their effects upon the map product including a comparison with field survey data.