Land cover and land use classification performance of machine learning algorithms in a boreal landscape using Sentinel-2 data

ABSTRACT

In recent years, the data science and remote sensing communities have started to align due to user-friendly programming tools, access to high-end consumer computing power, and the availability of free satellite data. In particular, publicly available data from the European Space Agency’s Sentinel missions have been used in various remote sensing applications. However, there is a lack of studies that utilize these data to assess the performance of machine learning algorithms in complex boreal landscapes. In this article, I compare the classification performance of four non-parametric algorithms: support vector machines (SVM), random forests (RF), extreme gradient boosting (Xgboost), and deep learning (DL). The study area chosen is a complex mixed-use landscape in south-central Sweden with eight land-cover and land-use (LCLU) classes. The satellite imagery used for the classification were multi-temporal scenes from Sentinel-2 covering spring, summer, autumn and winter conditions. Using stratified random sampling, each LCLU class was allocated 1477 samples, which were divided into training (70%) and evaluation (30%) subsets. Accuracy was assessed through metrics derived from an error matrix, but primarily overall accuracy was used in allocating algorithm hierarchy. A two-proportion Z-test was used to compare the proportions of correctly classified pixels of the algorithms and a McNemar’s chi-square test was used to compare class-wise predictions. The results show that the highest overall accuracy was produced by support vector machines (0.758 ± 0.017), closely followed by extreme gradient boosting (0.751 ± 0.017), random forests (0.739 ± 0.018), and finally deep learning (0.733 ± 0.0023). The Z-test comparison of classifiers showed that a third of algorithm pairings were statistically different. On a class-wise basis, McNemar’s test results showed that 62% of class-wise predictions were significant from one another at the 5% level or less. Variable importance metrics show that nearly half of the top twenty Sentinel-2 bands belonged to the red edge (25%) and shortwave infrared (23%) portions of the electromagnetic spectrum, and were dominated by scenes from spring (38%) and summer (40%). The results are discussed within the scope of recent studies involving machine learning and Sentinel-2 data and key knowledge gaps identified. The article concludes with recommendations for future research.     

The analysis of SPOT-5 characteristics on land cover Classification

<正>KnowledgeaboutlandcoverandlandusehasbecomeincreasinglyimportantastheNationplanstoovercometheproblemsofuncontrolleddevelopment,deterioratingenvironmentalquality,lossofprimeagriculturallandsetc.Landuseandlandcoverdataareneededintheanalysisofenvironmentalprocessesandproblemstoknowiflivingconditionsandstandardsaretobeimprovedormaintainedatcurrentlevels.     

面向对象的沿海地区土地利用/覆被信息提取研究

土地利用/覆被(LUC)可为土地资源领域相关研究提供基础数据.本文构建了面向对象的LUC分类方法,并以沿海特殊土地类型区连云港市为例,应用Landsat 8影像开展了实证研究。结果表明:①总体分类精度达到85.06%,总体Kappa系数为0.83,超过了0.7的最低允许判别精度;②该方法可以有效地减少研究区因南北部区域耕地植被覆盖度不同导致的错分现象,并可以用于盐田与滩涂信息的提取工作;③该方法既可为研究区土地利用相关研究提供符合精度要求的数据.也可为其他沿海地区进行土地利用/覆被信息提取工作提供参考和借鉴。     

Development of a modified neural network-based land cover classification system using automated data selector and multiresolution remotely sensed data

Integrating multiple images with artificial neural networks (ANN) improves classification accuracy. ANN performance is sensitive to training datasets. Complexity and errors compound when merging multiple data, pointing to needs for new techniques. Kohonen's self-organizing mapping (KSOM) neural network was adapted as an automated data selector (ADS) to replace manual training data processes. The multilayer perceptron (MLP) network was then trained using automatically extracted datasets and used for classification. Two hypotheses were tested: ADS adapted from the KSOM network provides adequate and reliable training datasets, improving MLP classification performance; and fusion of Landsat thematic mapper (TM) and SPOT images using the modified ANN approach increases accuracy. ADS adapted from the KSOM network improved training data quality and increased classification accuracy and efficiency. Fusion of compatible multiple data can improve performance if appropriate training datasets are collected. This proved to be a viable classification scheme particularly where acquiring sufficient and reliable training datasets is difficult.     

Tree cover mapping based on Sentinel-2 images demonstrate high thematic accuracy in Europe

The spatial and temporal distribution of trees has a large impact on human health and the environment through contributions to important climate mechanisms as well as commercial, recreational and social activities in society. A range of tree mapping methodologies has been presented in the literature, but tree cover estimates still differ widely between the individual datasets, and comparisons of the thematic accuracy of the resulting tree maps are rather scarce. The Copernicus Sentinel-2 satellites, which were launched in 2015 and 2017, have a combination of high spatial and temporal resolution. Given that this is a new satellite, a substantial amount of research on development of tree mapping algorithms as well as accuracy assessment of said algorithms have to be done in the years to come. To contribute to this process, a tree map produced through unsupervised classification was created for six Sentinel-2 tiles. The agreement between the tree map and the corresponding national forest inventory, as a function of the band combination chosen, was analysed and the thematic accuracy was assessed for two out of the six tiles. The results show that the highest agreement between the present tree map and the national forest inventory was found for bands 2, 3, 6 and 12. The present tree map has a relative difference in tree cover between 8% and 79% compared to previous estimates, but results are characterised by large scatter. Lastly, it is shown that the overall thematic accuracy of the present map is up to 90%, with the user’s accuracy ranging from 34.85% to 92.10%, and the producer’s accuracy ranging from 23.80% to 97.60% for the various thematic classes. This demonstrates that tree maps with high thematic accuracy can be produced from Sentinel-2. In the future the thematic accuracy can be increased even more through the use of temporal averaging in the mapping procedure, which will enable an accurate estimate of the European tree cover.     

Fusion of PolSAR and PolInSAR data for land cover classification

The main research goal of this study is to investigate the complementarity and fusion of different frequencies (L- and P-band), polarimetric SAR (PolSAR) and polarimetric interferometric (PolInSAR) data for land cover classification. A large feature set was derived from each of these four modalities and a two-level fusion method was developed: Logistic regression (LR) as ‘feature-level fusion’ and the neural-network (NN) method for higher level fusion. For comparison, a support vector machine (SVM) was also applied. NN and SVM were applied on various combinations of the feature sets.     

Changing landscape in the Three Gorges Reservoir Area of Yangtze River from 1977 to 2005: Land use/land cover,vegetation cover changes estimated using multi-source satellite data

The eco-environment in the Three Gorges Reservoir Area (TGRA) in China has received much attention due to the construction of the Three Gorges Hydropower Station. Land use/land cover changes (LUCC) are a major cause of ecological environmental changes. In this paper, the spatial landscape dynamics from 1978 to 2005 in this area are monitored and recent changes are analyzed, using the Landsat TM (MSS) images of 1978, 1988, 1995, 2000 and 2005. Vegetation cover fractions for a vegetation cover analysis are retrieved from MODIS/Terra imagery from 2000 to 2006, being the period before and after the rising water level of the reservoir. Several analytical indices have been used to analyze spatial and temporal changes. Results indicate that cropland, woodland, and grassland areas reduced continuously over the past 30 years, while river and built-up area increased by 2.79% and 4.45% from 2000 to 2005, respectively. The built-up area increased at the cost of decreased cropland, woodland and grassland. The vegetation cover fraction increased slightly. We conclude that significant changes in land use/land cover have occurred in the Three Gorges Reservoir Area. The main cause is a continuous economic and urban/rural development, followed by environmental management policies after construction of the Three Gorges Dam.     

Selection of LiDAR geometric features with adaptive neighborhood size for urban land cover classification

LiDAR has been an effective technology for acquiring urban land cover data in recent decades. Previous studies indicate that geometric features have a strong impact on land cover classification. Here, we analyzed an urban LiDAR dataset to explore the optimal feature subset from 25 geometric features incorporating 25 scales under 6 definitions for urban land cover classification. We performed a feature selection strategy to remove irrelevant or redundant features based on the correlation coefficient between features and classification accuracy of each features. The neighborhood scales were divided into small (0.5–1.5 m), medium (1.5–6 m) and large (>6 m) scale. Combining features with lower correlation coefficient and better classification performance would improve classification accuracy. The feature depicting homogeneity or heterogeneity of points would be calculated at a small scale, and the features to smooth points at a medium scale and the features of height different at large scale. As to the neighborhood definition, cuboid and cylinder were recommended. This study can guide the selection of optimal geometric features with adaptive neighborhood scale for urban land cover classification.     

Assessing the suitability of data from Sentinel-1A and 2A for crop classification

Sentinel-1A C-SAR and Sentinel-2A MultiSpectral Instrument (MSI) provide data applicable to the remote identification of crop type. In this study, six crop types (beans, beetroot, grass, maize, potato, and winter wheat) were identified using five C-SAR images and one MSI image acquired during the 2016 growing season. To assess the potential for accurate crop classification with existing supervised learning models, the four different approaches namely kernel-based extreme learning machine (KELM), multilayer feedforward neural networks, random forests, and support vector machine were compared. Algorithm hyperparameters were tuned using Bayesian optimization. Overall, KELM yielded the highest performance, achieving an overall classification accuracy of 96.8%. Evaluation of the sensitivity of classification models and relative importance of data types using data-based sensitivity analysis showed that the set of VV polarization data acquired on 24 July (Sentinel-1A) and band 4 data (Sentinel-2A) had the greatest potential for use in crop classification.     

Forest cover classification using Landsat ETM+ data and time series MODIS NDVI data

Forest cover plays a key role in climate change by influencing the carbon stocks, the hydrological cycle and the energy balance. Forest cover information can be determined from fine-resolution data, such as Landsat Enhanced Thematic Mapper Plus (ETM+). However, forest cover classification with fine-resolution data usually uses only one temporal data because successive data acquirement is difficult. It may achieve mis-classification result without involving vegetation growth information, because different vegetation types may have the similar spectral features in the fine-resolution data. To overcome these issues, a forest cover classification method using Landsat ETM+ data appending with time series Moderate-resolution Imaging Spectroradiometer (MODIS) Normalized Difference Vegetation Index (NDVI) data was proposed. The objective was to investigate the potential of temporal features extracted from coarse-resolution time series vegetation index data on improving the forest cover classification accuracy using fine-resolution remote sensing data. This method firstly fused Landsat ETM+ NDVI and MODIS NDVI data to obtain time series fine-resolution NDVI data, and then the temporal features were extracted from the fused NDVI data. Finally, temporal features combined with Landsat ETM+ spectral data was used to improve forest cover classification accuracy using supervised classifier. The study in North China region confirmed that time series NDVI features had significant effects on improving forest cover classification accuracy of fine resolution remote sensing data. The NDVI features extracted from time series fused NDVI data could improve the overall classification accuracy approximately 5% from 88.99% to 93.88% compared to only using single Landsat ETM+ data.     

Enhanced land use/cover classification of heterogeneous tropical landscapes using support vector machines and textural homogeneity

Land use/cover classification is a key research field in remote sensing and land change science as thematic maps derived from remotely sensed data have become the basis for analyzing many socio-ecological issues. However, land use/cover classification remains a difficult task and it is especially challenging in heterogeneous tropical landscapes where nonetheless such maps are of great importance. The present study aims at establishing an efficient classification approach to accurately map all broad land use/cover classes in a large, heterogeneous tropical area, as a basis for further studies (e.g., land use/cover change, deforestation and forest degradation). Specifically, we first compare the performance of parametric (maximum likelihood), non-parametric (k-nearest neighbor and four different support vector machines – SVM), and hybrid (unsupervised–supervised) classifiers, using hard and soft (fuzzy) accuracy assessments. We then assess, using the maximum likelihood algorithm, what textural indices from the gray-level co-occurrence matrix lead to greater classification improvements at the spatial resolution of Landsat imagery (30 m), and rank them accordingly. Finally, we use the textural index that provides the most accurate classification results to evaluate whether its usefulness varies significantly with the classifier used. We classified imagery corresponding to dry and wet seasons and found that SVM classifiers outperformed all the rest. We also found that the use of some textural indices, but particularly homogeneity and entropy, can significantly improve classifications. We focused on the use of the homogeneity index, which has so far been neglected in land use/cover classification efforts, and found that this index along with reflectance bands significantly increased the overall accuracy of all the classifiers, but particularly of SVM. We observed that improvements in producer's and user's accuracies through the inclusion of homogeneity were different depending on land use/cover classes. Early-growth/degraded forests, pastures, grasslands and savanna were the classes most improved, especially with the SVM radial basis function and SVM sigmoid classifiers, though with both classifiers all land use/cover classes were mapped with producer's and user's accuracies of ∼90%. Our classification approach seems very well suited to accurately map land use/cover of heterogeneous landscapes, thus having great potential to contribute to climate change mitigation schemes, conservation initiatives, and the design of management plans and rural development policies.     

Fusion of HJ1B and ALOS PALSAR data for land cover classification using machine learning methods

Image classification from remote sensing is becoming increasingly urgent for monitoring environmental changes. Exploring effective algorithms to increase classification accuracy is critical. This paper explores the use of multispectral HJ1B and ALOS (Advanced Land Observing Satellite) PALSAR L-band (Phased Array type L-band Synthetic Aperture Radar) for land cover classification using learning-based algorithms. Pixel-based and object-based image analysis approaches for classifying HJ1B data and the HJ1B and ALOS/PALSAR fused-images were compared using two machine learning algorithms, support vector machine (SVM) and random forest (RF), to test which algorithm can achieve the best classification accuracy in arid and semiarid regions. The overall accuracies of the pixel-based (Fused data: 79.0%; HJ1B data: 81.46%) and object-based classifications (Fused data: 80.0%; HJ1B data: 76.9%) were relatively close when using the SVM classifier. The pixel-based classification achieved a high overall accuracy (85.5%) using the RF algorithm for classifying the fused data, whereas the RF classifier using the object-based image analysis produced a lower overall accuracy (70.2%). The study demonstrates that the pixel-based classification utilized fewer variables and performed relatively better than the object-based classification using HJ1B imagery and the fused data. Generally, the integration of the HJ1B and ALOS/PALSAR imagery can improve the overall accuracy of 5.7% using the pixel-based image analysis and RF classifier.     

Reducing landscape heterogeneity for improved land use and land cover (LULC) classification across the large and complex Ethiopian highlands

This paper presents a land use and land cover (LULC) classification approach that accounts landscape heterogeneity. We addressed this challenge by subdividing the study area into more homogeneous segments using several biophysical and socio-economic factors as well as spectral information. This was followed by unsupervised clustering within each homogeneous segment and supervised class assignment. Two classification schemes differing in their level of detail were successfully applied to four landscape types of distinct LULC composition. The resulting LULC map fulfills two major requirements: (1) differentiation and identification of several LULC classes that are of interest at the local, regional, and national scales, and (2) high accuracy of classification. The approach overcomes commonly encountered difficulties of classifying second-level classes in large and heterogeneous landscapes. The output of the study responds to the need for comprehensive LULC data to support ecosystem assessment, policy formulation, and decision-making towards sustainable land resources management.     

利用Sentinel-2A多光谱成像仪与Landsat8陆地成像仪影像进行普陀山岛植被分类效果比较

卫星遥感技术可用于海岛资源调查。Sentinel-2A与Landsat 8两颗卫星都可免费提供空间分辨率较高的多光谱遥感影像,在海岛调查中的应用潜力较大。本文以浙江舟山普陀山岛为例开展了针对这两种影像在海岛植被分类中的应用效果的研究,分别利用Sentinel-2A多光谱成像仪（MSI）和Landsat 8陆地成像仪（OLI）影像基于最大似然法分类获得了该岛阔叶林、针阔混交林、针叶林、灌丛、草丛等植被及其他地物的分布情况,并进行了精度检验,结果表明MSI的总体分类精度略高于OLI。     

Land use land cover classification of Orissa using multi-temporal IRS-P6 awifs data: A decision tree approach

An attempt has been made to understand the potential of temporal Advanced Wide Field Sensor (AWiFS) data aboard IRS-P6 (Resourcesat) to generate the land use land cover information along with the net sown area. The temporal data sets were georeferenced, converted to top of atmosphere reflectance and classified using decision tree classifier, See5. Results indicate that the temporal data set could give a better definition of training sites thereby resulting in good overall kappa (kappa = 0.8651) as well as individual classification accuracies. However, co-registration of temporal datasets accuracies also has got a significant influence on the classification accuracy. Temporal variation in cloud infestation and availability of appropriate data sets within the season (before harvest of the crop) has also affected the classification accuracy.     

Utility of image point cloud data towards generating enhanced multitemporal multisensor land cover maps

Multitemporal land cover classification over urban areas is challenging, especially when using heterogeneous data sources with variable quality attributes. A prominent challenge is that classes with similar spectral signatures (such as trees and grass) tend to be confused with one another. In this paper, we evaluate the efficacy of image point cloud (IPC) data combined with suitable Bayesian analysis based time-series rectification techniques to improve the classification accuracy in a multitemporal context. The proposed method uses hidden Markov models (HMMs) to rectify land covers that are initially classified by a random forest (RF) algorithm. This land cover classification method is tested using time series of remote sensing data from a heterogeneous and rapidly changing urban landscape (Kuopio city, Finland) observed from 2006 to 2014. The data consisted of aerial images (5 years), Landsat data (all 9 years) and airborne laser scanning data (1 year). The results of the study demonstrate that the addition of three-dimensional image point cloud data derived from aerial stereo images as predictor variables improved overall classification accuracy, around three percentage points. Additionally, HMM-based post processing reduces significantly the number of spurious year-to-year changes. Using a set of 240 validation points, we estimated that this step improved overall classification accuracy by around 3.0 percentage points, and up to 6 to 10 percentage points for some classes. The overall accuracy of the final product was 91% (kappa = 0.88). Our analysis shows that around 1.9% of the area around Kuopio city, representing a total area of approximately 0.61 km², experienced changes in land cover over the nine years considered.     

Temporal optimisation of image acquisition for land cover classification with Random Forest and MODIS time-series

The analysis and classification of land cover is one of the principal applications in terrestrial remote sensing. Due to the seasonal variability of different vegetation types and land surface characteristics, the ability to discriminate land cover types changes over time. Multi-temporal classification can help to improve the classification accuracies, but different constraints, such as financial restrictions or atmospheric conditions, may impede their application. The optimisation of image acquisition timing and frequencies can help to increase the effectiveness of the classification process. For this purpose, the Feature Importance (FI) measure of the state-of-the art machine learning method Random Forest was used to determine the optimal image acquisition periods for a general (Grassland, Forest, Water, Settlement, Peatland) and Grassland specific (Improved Grassland, Semi-Improved Grassland) land cover classification in central Ireland based on a 9-year time-series of MODIS Terra 16 day composite data (MOD13Q1). Feature Importances for each acquisition period of the Enhanced Vegetation Index (EVI) and Normalised Difference Vegetation Index (NDVI) were calculated for both classification scenarios. In the general land cover classification, the months December and January showed the highest, and July and August the lowest separability for both VIs over the entire nine-year period. This temporal separability was reflected in the classification accuracies, where the optimal choice of image dates outperformed the worst image date by 13% using NDVI and 5% using EVI on a mono-temporal analysis. With the addition of the next best image periods to the data input the classification accuracies converged quickly to their limit at around 8–10 images. The binary classification schemes, using two classes only, showed a stronger seasonal dependency with a higher intra-annual, but lower inter-annual variation. Nonetheless anomalous weather conditions, such as the cold winter of 2009/2010 can alter the temporal separability pattern significantly. Due to the extensive use of the NDVI for land cover discrimination, the findings of this study should be transferrable to data from other optical sensors with a higher spatial resolution. However, the high impact of outliers from the general climatic pattern highlights the limitation of spatial transferability to locations with different climatic and land cover conditions. The use of high-temporal, moderate resolution data such as MODIS in conjunction with machine-learning techniques proved to be a good base for the prediction of image acquisition timing for optimal land cover classification results.     

Potential of Landsat-8 and Sentinel-2A composite for land use land cover analysis

Abstract

This study proposes the development of a multi-sensor, multi-spectral composite from Landsat-8 and Sentinel-2A imagery referred to as ‘LSC’ for land use land cover (LULC) characterisation and compared with respect to the hyperspectral imagery of the EO1: Hyperion sensor. A three-stage evaluation was implemented based on the similarity observed in the spectral response, supervised classification results and endmember abundance information obtained using linear spectral unmixing. The study was conducted for two areas located around Dhundi and Rohtak in Himachal Pradesh and Haryana, respectively. According to the analysis of the spectral reflectance curves, the spectral response of the LSC is capable of identifying major LULC classes. The kappa accuracy of 0.85 and 0.66 was observed for the classification results from LSC and Hyperion data for Dhundi and Rohtak datasets, respectively. The coefficient of determination was found to be above 0.9 for the LULC classes in both the datasets as compared to Hyperion, indicating a good agreement. Thus, these three-stage results indicated the significant potential of a composite derived from freely available multi-sensor multi-spectral imagery as an alternative to hyperspectral imagery for LULC studies.     

Land use and land cover classification over a large area in Iran based on single date analysis of satellite imagery

Accelerated soil erosion, high sediment yields, floods and debris flow are serious problems in many areas of Iran, and in particular in the Golestan dam watershed, which is the area that was investigated in this study. Accurate land use and land cover (LULC) maps can be effective tools to help soil erosion control efforts. The principal objective of this research was to propose a new protocol for LULC classification for large areas based on readily available ancillary information and analysis of three single date Landsat ETM+ images, and to demonstrate that successful mapping depends on more than just analysis of reflectance values. In this research, it was found that incorporating climatic and topographic conditions helped delineate what was otherwise overlapping information. This study determined that a late summer Landsat ETM+ image yields the best results with an overall accuracy of 95%, while a spring image yields the poorest accuracy (82%). A summer image yields an intermediate accuracy of 92%. In future studies where funding is limited to obtaining one image, late summer images would be most suitable for LULC mapping. The analysis as presented in this paper could also be done with satellite images taken at different times of the season. It may be, particularly for other climatic zones, that there is a better time of season for image acquisition that would present more information.     

Object-based urban land cover classification using rule inheritance over very high-resolution multisensor and multitemporal data

Very high spatial and temporal resolution remote sensing data facilitate mapping highly complex and diverse urban environments. This study analyzed and demonstrated the usefulness of combined high-resolution aerial digital images and elevation data, and its processing using object-based image analysis for mapping urban land covers and quantifying buildings. It is observed that mapping heterogeneous features across large urban areas is time consuming and challenging. This study presents and demonstrates an approach for formulating an optimal land cover classification rule set over small representative training urban area image, and its subsequent transfer to the multisensor, multitemporal images. The classification results over the training area showed an overall accuracy of 96%, and the application of rule set to different sensor images of other test areas resulted in reduced accuracies of 91% for the same sensor, 90% and 86% for the different sensors temporal data. The comparison of reference and classified buildings showed ±4% detection errors. Classification through a transferred rule set reduced the classification accuracy by about 5%–10%. However, the trade-off for this accuracy drop was about a 75% reduction in processing time for performing classification in the training area. The factors influencing the classification accuracies were mainly the shadow and temporal changes in the class characteristics.