Very High Resolution Images Research Articles

Adaptive Effective Receptive Field Convolution for Semantic Segmentation of VHR Remote Sensing Images

Convolutional neural networks (CNNs) have facilitated impressive improvements in the semantic segmentation of very high-resolution (VHR) remote sensing images. The success of semantic segmentation depends on an effective receptive field (RF) large enough to cover the entire object. Popular methods to enlarge the effective RF include dilated filters, subsampling operations, and stacking layers. Unfortunately, the methods are inefficient or able to cause grid artifacts. Moreover, although the object sizes vary greatly in remote sensing images, the size of the RF cannot reach a compromise between small and large objects. To tackle these problems, we propose adaptive effective receptive convolution (AERFC) for VHR remote sensing images. AERFC adaptively controls the sampling location of convolution and automatically adjusts the effective RF without significantly increasing the parameter number and computational cost. Thus, AERFC reduces the training difficulty, decreases overfitting risk, and reserves details in VHR images. AERFC is also integrated with spatial pyramid pooling (SPP) to aggregate diverse multiscale features for exploring contextual information. Experimental results of the quantitative and qualitative evaluation over four benchmark data sets show that AERFC outperforms state-of-the-art methods.

Detection, Classification and Boundary Regularization of Buildings in Satellite Imagery Using Faster Edge Region Convolutional Neural Networks

With the development of effective deep learning algorithms, it became possible to achieve high accuracy when conducting remote sensing analyses on very high-resolution images (VHRS), especially in the context of building detection and classification. In this article, in order to improve the accuracy of building detection and classification, we propose a Faster Edge Region Convolutional Neural Networks (FER-CNN) algorithm. This proposed algorithm is trained and evaluated on different datasets. In addition, we propose a new method to improve the detection of the boundaries of detected buildings. The results of our algorithm are compared with those of other methods, such as classical Faster Region Convolution Neural Network (Faster R-CNN) with the original VGG16 and the Single-Shot Multibox Detector (SSD). The experimental results show that our methods make it possible to obtain an average detection accuracy of 97.5% with a false positive classification rate of 8.4%. An additional advantage of our method is better resistance to shadows, which is a very common issue for satellite images of urban areas. Future research will include designing and training the neural network to detect small buildings, as well as irregularly shaped buildings that are partially obscured by shadows or other occlusions.

A Bayesian characterization of urban land use configurations from VHR remote sensing images

The composition and arrangement of spatial entities, i.e., land cover objects, play a key role in distinguishing land use types from very high resolution (VHR) remote sensing images, in particular in urban environments. This paper presents a new method to characterize the spatial arrangement for urban land use extraction using VHR images. We derive an adjacency unit matrix to represent the spatial arrangement of land cover objects obtained from a VHR image, and use a graph convolutional network to quantify the spatial arrangement by extracting hidden features from adjacency unit matrices. The distribution of the spatial arrangement variables, i.e., hidden features, and the spatial composition variables, i.e., widely used land use indicators, are then estimated. We use a Bayesian method to integrate the variables of spatial arrangement and composition for urban land use extraction. Experiments were conducted using three VHR images acquired in two urban areas: a Pleiades image in Wuhan in 2013, a Superview image in Wuhan in 2019, and a GeoEye image in Oklahoma City in 2012. Our results show that the proposed method provides an effective means to characterize the spatial arrangement of land cover objects, and produces urban land use extractions with overall accuracies (i.e., 86% and 93%) higher than existing methods (i.e., 83% and 88%) that use spatial arrangement information based on building types on the Pleiades and GeoEye datasets. Moreover, it is unnecessary to further categorize the dominant land cover type into finer types for the characterization of spatial arrangement. We conclude that the proposed method has a high potential for the characterization of urban structure using different VHR images, and for the extraction of urban land use in different urban areas.

Open-source data-driven urban land-use mapping integrating point-line-polygon semantic objects: A case study of Chinese cities

Reliable urban land-use maps are essential for urban analysis because the spatial distribution of land use reflects the complex environment of cities under the combined effects of nature and socio-economics. In recent years, very high resolution (VHR) remote sensing imagery interpretation has resolved the “semantic gap” between the low-level data and the high-level semantic scenes, and has been used to map urban land use. Nevertheless, the existing frameworks cannot easily be applied to practical urban analysis, which can be attributed to three main reasons: 1) the indistinguishable socio-economic attributes of the same ground object layouts; 2) the weak transferability of the supervised frameworks and the time-consuming training sample annotation; and 3) the category system inconsistency between the data source and the urban land-use application. In this paper, to achieve an “application gap” breakthrough for urban land-use mapping, a data-driven point, line, and polygon semantic object mapping (PLPSOM) framework is proposed, which makes full use of open-source VHR images and multi-source geospatial data. In the PLPSOM framework, point, line, and polygon semantic objects are represented by the points of interest (POIs), OpenStreetMap (OSM) data, and VHR images corresponding to the scenes in the land-use mapping units, respectively. OSM line semantic objects are utilized to supply the boundaries of the land-use mapping units for the POIs and VHR images, forming urban land parcels (street blocks). To reduce the cost of the data annotation, the training dataset is constructed using multiple open-source data sources. An enhanced deep adaptation network (EDAN) is then proposed to acquire the categories of the VHR scene images in the case of partial transfer learning. Finally, in order to meet the actual needs, a rule-based category mapping (RCM) model is applied to integrate the categories of the POIs and VHR images into the urban land-use category system, allowing us to acquire the land-use maps of the cities. The effectiveness of the proposed method was tested in four cities of China, including six specific areas: Beijing and Wuhan city centers; the Hanyang District of Wuhan; the Hannan District of Wuhan; Macao; and the Wan Chai area of Hong Kong, achieving a high classification accuracy. The “urban image” analysis confirmed the practicality of the obtained urban land-use maps.

Automatic building detection from very high-resolution images using multiscale morphological attribute profiles

ABSTRACT Morphological building indexes (MBI) have proven to be effective tools for automated building spatial-feature-extraction tasks in images from urban areas. However, owing to the intrinsic shortcomings of MBI, commission and omission errors occur in regions with spectral properties similar to those of buildings and dark heterogeneous roofs, respectively. Some targets (such as bright bare land or roads) can cause substantial interference, which poses an even greater challenge in performing accurate building detection from images of complex environments. In this study, a new automated building detection approach based on a morphological attribute profile is presented with the goal of reducing commission and omission errors. As the first step, corners are detected in very high-resolution (VHR) remote sensing images through an automatic optimization procedure, and weighted spatial voting is performed to predict the presence of built-up areas. Then, by investigating the properties between the attribute filters and buildings, a novel morphological attribute building index is constructed by considering the extracted built-up area as an input image. To validate the detection performance, the approach was tested using VHR images with 1-m spatial resolution. The quantitative assessment indicates that the proposed approach improves the building detection accuracy in images of complex environments.

An Improved Boundary-Aware Perceptual Loss for Building Extraction from VHR Images

With the development of deep learning technology, an enormous number of convolutional neural network (CNN) models have been proposed to address the challenging building extraction task from very high-resolution (VHR) remote sensing images. However, searching for better CNN architectures is time-consuming, and the robustness of a new CNN model cannot be guaranteed. In this paper, an improved boundary-aware perceptual (BP) loss is proposed to enhance the building extraction ability of CNN models. The proposed BP loss consists of a loss network and transfer loss functions. The usage of the boundary-aware perceptual loss has two stages. In the training stage, the loss network learns the structural information from circularly transferring between the building mask and the corresponding building boundary. In the refining stage, the learned structural information is embedded into the building extraction models via the transfer loss functions without additional parameters or postprocessing. We verify the effectiveness and efficiency of the proposed BP loss both on the challenging WHU aerial dataset and the INRIA dataset. Substantial performance improvements are observed within two representative CNN architectures: PSPNet and UNet, which are widely used on pixel-wise labelling tasks. With BP loss, UNet with ResNet101 achieves 90.78% and 76.62% on IoU (intersection over union) scores on the WHU aerial dataset and the INRIA dataset, respectively, which are 1.47% and 1.04% higher than those simply trained with the cross-entropy loss function. Additionally, similar improvements (0.64% on the WHU aerial dataset and 1.69% on the INRIA dataset) are also observed on PSPNet, which strongly supports the robustness of the proposed BP loss.

Object-Oriented Key Point Vector Distance for Binary Land Cover Change Detection Using VHR Remote Sensing Images

Very high-resolution (VHR) remote sensing images can geometrically depict ground targets in detail but are usually insufficient in the spectral domain. This characteristic leads to a considerable amount of noise and pseudo change in the produced binary change detection maps (BCDMs) when VHR remote sensing images are used for change detection. Here, to solve the aforementioned problem, an object-oriented key point vector distance (KPVD) is proposed to measure the change magnitude between bitemporal VHR images when land cover changes are detected. The proposed KPVD-based change detection approach comprises the following major steps. First, multiscale objects based on a postevent image are extracted by the fractional net evaluation segmentation approach, and then, the segments are taken as the unit for measuring the change magnitude between bitemporal images. Second, key points and the corresponding vector are defined to describe the object feature instead of using the total pixels within the object. Finally, KPVD is proposed to measure the change magnitude between the local areas referenced to the object in the bitemporal images. The change magnitude image (CMI) between the bitemporal images is generated while the entire images are scanned and processed object by object. A well-known automatic binary method, the Otsu approach, is employed in this article to divide CMI into a BCDM. Experimental results conducted on four real data sets demonstrate the feasibility and outperformance of the proposed KPVD-based change detection approach compared with five state-of-the-art methods in terms of visual performance and quantitative measurements.

Object-Based Change Detection of Very High Resolution Images by Fusing Pixel-Based Change Detection Results Using Weighted Dempster–Shafer Theory

Change detection (CD), one of the primary applications of multi-temporal satellite images, is the process of identifying changes in the Earth’s surface occurring over a period of time using images of the same geographic area on different dates. CD is divided into pixel-based change detection (PBCD) and object-based change detection (OBCD). Although PBCD is more popular due to its simple algorithms and relatively easy quantitative analysis, applying this method in very high resolution (VHR) images often results in misdetection or noise. Because of this, researchers have focused on extending the PBCD results to the OBCD map in VHR images. In this paper, we present a proposed weighted Dempster-Shafer theory (wDST) fusion method to generate the OBCD by combining multiple PBCD results. The proposed wDST approach automatically calculates and assigns a certainty weight for each object of the PBCD result while considering the stability of the object. Moreover, the proposed wDST method can minimize the tendency of the number of changed objects to decrease or increase based on the ratio of changed pixels to the total pixels in the image when the PBCD result is extended to the OBCD result. First, we performed co-registration between the VHR multitemporal images to minimize the geometric dissimilarity. Then, we conducted the image segmentation of the co-registered pair of multitemporal VHR imagery. Three change intensity images were generated using change vector analysis (CVA), iteratively reweighted-multivariate alteration detection (IRMAD), and principal component analysis (PCA). These three intensity images were exploited to generate different binary PBCD maps, after which the maps were fused with the segmented image using the wDST to generate the OBCD map. Finally, the accuracy of the proposed CD technique was assessed by using a manually digitized map. Two VHR multitemporal datasets were used to test the proposed approach. Experimental results confirmed the superiority of the proposed method by comparing the existing PBCD methods and the OBCD method using the majority voting technique.

A Multi-Scale Superpixel-Guided Filter Feature Extraction and Selection Approach for Classification of Very-High-Resolution Remotely Sensed Imagery

In this article, a novel feature selection-based multi-scale superpixel-based guided filter (FS-MSGF) method for classification of very-high-resolution (VHR) remotely sensed imagery is proposed. Improved from the original guided filter (GF) algorithm used in the classification, the guidance image in the proposed approach is constructed based on the superpixel-level segmentation. By taking into account the object boundaries and the inner-homogeneity, the superpixel-level guidance image leads to the geometrical information of land-cover objects in VHR images being better depicted. High-dimensional multi-scale guided filter (MSGF) features are then generated, where the multi-scale information of those land-cover classes is better modelled. In addition, for improving the computational efficiency without the loss of accuracy, a subset of those MSGF features is then automatically selected by using an unsupervised feature selection method, which contains the most distinctive information in all constructed MSGF features. Quantitative and qualitative classification results obtained on two QuickBird remotely sensed imagery datasets covering the Zurich urban scene are provided and analyzed, which demonstrate that the proposed methods outperform the state-of-the-art reference techniques in terms of higher classification accuracies and higher computational efficiency.

Domain adaptation for unsupervised change detection of multisensor multitemporal remote-sensing images

ABSTRACTWith the advancement of the space-based imaging technology, large amounts of images of different modality and spatial, spectral, and temporal resolutions are available. However, it is clearly demonstrated that combining images from different sources improves results in remote-sensing applications, it still remains a challenge to make full use of their benefits. These challenges become complicate, extending to multitemporal analysis and very high resolution (VHR) images, which includes different acquisition conditions, and consequently radiometric and geometric differences between images. To design methods that are robust to data set shifts, recent remote-sensing literature has considered solutions based on domain adaptation (DA) approaches. In this research, we aim to propose a method based on autoencoder which is a DA method to achieve fused features of synthetic aperture radar (SAR) and optical to benefit from their complementary information, and simultaneously aligning multitemporal features by reduction in spectral and radiometric differences and make multitemporal features more similar for better accuracy in change detection (CD). In continuation, to cope with the problem of detecting changes using VHR images, we introduce a framework to obtain the change map, in which, the differences caused by the registration error as well as the geometric differences of VHR images will be minimized, by maximum use of spatial information of images. For these purposes, a multitemporal pair of co-registered images, each composed of optical multispectral channels and a SAR amplitude channel acquired over the same urban area before and after changes, is assumed. Finally, the strategy of detecting changes resulted in improvement compared to common and simpler methods and also individual sensor CD method, which demonstrate the effectiveness of the proposed approach.

Heuristic sample learning for complex urban scenes: Application to urban functional-zone mapping with VHR images and POI data

Urban functional zones are basic units of urban planning and resource allocation, and contribute to a wide range of urban studies and investigations. Existing studies on functional-zone mapping with very-high-resolution (VHR) satellite images focused much on feature representations and classification techniques, but ignored zone sampling which however was fundamental to automatic zone classifications. Functional-zone sampling is much complicated and can hardly be resolved by classical sampling methods, as functional zones are complex urban scenes which consist of heterogeneous land covers and have highly abstract categories. To resolve the issue, this study presents a novel sampling paradigm, i.e., heuristic sample learning (HSL). It first proposes a sparse topic model to select representative functional zones, then uses deep forest to select confusing zones, and finally embraces Chinese restaurant process to label these selected zones. The presented method collects both representative and confusing zone samples and identifies their categories accurately, which makes the functional-zone classification process robust and the classification results accurate. Experiments conducted in Beijing indicate that HSL is effective and efficient for functional-zone sampling and classifications. Compared to traditional manual sampling, HSL reduces the time cost by 55% and improves the classification accuracy by 11.3% on average; furthermore, HSL can reduce the variation in sampling and classification results caused by different proficiency of operators. Accordingly, HSL significantly contributes to functional-zone mapping and plays an important role in urban studies.

Change Detection in Multisource VHR Images via Deep Siamese Convolutional Multiple-Layers Recurrent Neural Network

With the rapid development of Earth observation technology, very-high-resolution (VHR) images from various satellite sensors are more available, which greatly enrich the data source of change detection (CD). Multisource multitemporal images can provide abundant information on observed landscapes with various physical and material views, and it is exigent to develop efficient techniques to utilize these multisource data for CD. In this article, we propose a novel and general deep siamese convolutional multiple-layers recurrent neural network (RNN) (SiamCRNN) for CD in multitemporal VHR images. Superior to most VHR image CD methods, SiamCRNN can be used for both homogeneous and heterogeneous images. Integrating the merits of both convolutional neural network (CNN) and RNN, SiamCRNN consists of three subnetworks: deep siamese convolutional neural network (DSCNN), multiple-layers RNN (MRNN), and fully connected (FC) layers. The DSCNN has a flexible structure for multisource image and is able to extract spatial-spectral features from homogeneous or heterogeneous VHR image patches. The MRNN stacked by long-short term memory (LSTM) units is responsible for mapping the spatial-spectral features extracted by DSCNN into a new latent feature space and mining the change information between them. In addition, FC, the last part of SiamCRNN, is adopted to predict change probability. The experimental results in two homogeneous data sets and one challenging heterogeneous VHR images data set demonstrate that the promising performances of the proposed network outperform several state-of-the-art approaches.

Semantic segmentation of very high-resolution remote sensing image based on multiple band combinations and patchwise scene analysis

Large intraclass variance and low interclass variance are among the most challenging problems in very high-resolution (VHR) image classification. Semantic segmentation constructed in a deep convolution neural network is used as a classification algorithm conducted via end-to-end training, which combines spectral–spatial features and context information. However, large-scale remote sensing images cannot be directly processed because they are limited by GPU memory and segmentation algorithm. At the same time, classification using single band combinations is also unsatisfactory due to the extraordinary complex features of VHR images. Therefore, a method is proposed based on multiple band combinations and patchwise scene analysis. A complex remote sensing image can be considered as into a combination of simple scenes from multiple patchwise images. And optimal band combinations of each patchwise image are selected according to their scene. The segmentation results of each patchwise image are merged to get the desired results according to geographical coordinates. Our method is validated on the ISPRS 2-D Semantic Labeling dataset of Potsdam, on which results competitive with the state-of-the-art are obtained. The proposed scheme has strong universality and can be used for large-scale high-resolution remote sensing image classification.

BMF-CNN: an object detection method based on multi-scale feature fusion in VHR remote sensing images

ABSTRACTObject detection in very-high-resolution (VHR) remote sensing images is one of the important technical means in many fields. In recent years, conventional object detection methods have been completely replaced by Convolutional Neural Network (CNN)-based methods, which are more accurate and efficient. However, most current CNN-based methods applied in VHR image sets have certain defects: (1) Scale preference is common in the framework designs, and the representation ability of feature maps for large and small objects is quite different, so accuracy promotion can hardly be made comprehensively in the detection of different objects. (2) The scale difference of the objects leads to training difficulties. (3) Some high-precision methods require high hardware costs, and the overall frameworks lack practicality. To address such problems, we propose a new object detection method in this paper, namely Balanced Multi-Scale Fusion-based CNN (BMF-CNN). It is a redesigned two-stage detection framework according to the region-based object detection methods, which enabled the detection accuracy of both large and small objects to reach a high level. Through the evaluation in the open VHR remote sensing image sets, we found that BMF-CNN showed a better integrative performance than the current mainstream detection methods.

Automatic Extraction of Built-Up Areas from Very High-Resolution Satellite Imagery Using Patch-Level Spatial Features and Gestalt Laws of Perceptual Grouping

Automatic extraction of built-up areas from very high-resolution (VHR) satellite images has received increasing attention in recent years. However, due to the complexity of spectral and spatial characteristics of built-up areas, it is still a challenging task to obtain their precise location and extent. In this study, a patch-based framework was proposed for unsupervised extraction of built-up areas from VHR imagery. First, a group of corner-constrained overlapping patches were defined to locate the candidate built-up areas. Second, for each patch, its salient textures and structural characteristics were represented as a feature vector using integrated high-frequency wavelet coefficients. Then, inspired by visual perception, a patch-level saliency model of built-up areas was constructed by incorporating Gestalt laws of proximity and similarity, which can effectively describe the spatial relationships between patches. Finally, built-up areas were extracted through thresholding and their boundaries were refined by morphological operations. The performance of the proposed method was evaluated on two VHR image datasets. The resulting average F-measure values were 0.8613 for the Google Earth dataset and 0.88 for the WorldView-2 dataset, respectively. Compared with existing models, the proposed method obtains better extraction results, which show more precise boundaries and preserve better shape integrity.

Extraction of urban building damage using spectral, height and corner information from VHR satellite images and airborne LiDAR data

Earth observation-based damage assessment has been widely studied in recent years. Considering that the height and spatial variability of buildings change significantly in earthquake-devastated areas, a novel multi-stage urban building damage extraction method that uses bi-temporal spectral, height and corner information is proposed in this study. The post-event height features were directly derived from airborne light detection and ranging (LiDAR) data, whereas pre-event height features were generated using pre-event stereo-paired images from different satellites. The spatial features were quantified using the density of corner points (DCP) in spectral images. The proposed method of urban building damage extraction is summarised as follows. Bi-temporal height and corner features were first generated from bi-temporal very high resolution (VHR) satellite data and post-event airborne LiDAR data. Then, vegetation, bare land (pavement and soil) and shadow were extracted from post-event VHR image and height data, and masked out. Finally, building damage was extracted from the remaining areas using the height difference and DCP difference between pre- and post-event images. A post-processing procedure was used to further refine the initial extraction results. The proposed method was evaluated using bi-temporal VHR images and post-event LiDAR data collected in Port au Prince, Haiti, which was heavily hit by an earthquake in January 2010. The results showed that the proposed method significantly outperformed the two comparative methods in the extraction of urban building damage.

Novel Adaptive Histogram Trend Similarity Approach for Land Cover Change Detection by Using Bitemporal Very-High-Resolution Remote Sensing Images

Detecting land cover change through very-high-resolution (VHR) remote sensing images is helpful in supporting urban sustainable development, natural disaster evaluation, and environmental assessment. However, the intraclass spectral variance in VHR remote sensing images is usually larger than that of median-low remote sensing images. Furthermore, the bitemporal images are usually acquired under different atmospheric conditions, sun height, soil moisture, and other factors. Consequently, in practical applications, many pseudo changes are presented in the detected map. In this paper, an adaptive histogram trend (AHT) similarity approach is promoted to quantitatively measure the magnitude between the corresponding pixels in bitemporal images in terms of change semantic. In the proposed approach, to reduce the phenological effect on the bitemporal images of land cover change detection (LCCD), we first define the quantitative description of AHT. Second, the change magnitudes between pairwise pixels are quantitatively measured by an improved bin-to-bin (B2B) distance between the corresponding AHTs. Then, the change magnitudes between two entire bitemporal images are measured AHT-by-AHT. Finally, binary threshold methods, such as the Otsu method or the double-window flexible pace search (DFPS) method, are used to divide the change magnitude image into binary change detection maps and obtain the final change detection map. The performance of the AHT-based LCCD approach is verified by four pairs of VHR remote-sensing images that correspond to two types of real land cover change cases. The detected results based on the four pairs of bitemporal VHR images outperformed the compared state-of-the-art LCCD methods.

Some Thoughts on Measuring Earthquake Deformation Using Optical Imagery

Optical imagery has been proven to be an effective tool for measuring earthquake deformation in continental regions since its first application in the 1999 Izmit earthquake. In this article, we compile and analyze all the earthquakes that have been investigated with optical image matching by 2019, based on which we comment on various issues regarding measuring earthquake deformation with optical imagery. New generations of very high-resolution (VHR) data are effective for earthquake studies, but orthorectification of the VHR images is the major source of error, which is often ignored. We found that the displacements derived from the WorldView images strongly correlate with the errors in the Shuttle Radar Topography Mission (SRTM) digital elevation model (DEM) that was used in orthorectification. Based on the observed correlation between displacements and topography, we propose a new DEM-based method using the Advanced Land Observing Satellite (ALOS) World 3-D DEM to reduce the orthorectification errors. Combining the published optical data of earthquake deformation, we re-analyze the coseismic slip distribution and shallow slip deficit (SSD). The SSD model states that the coseismic slip in many strike-slip earthquakes decreases in magnitude toward the surface, but this model remains arguable because the interferometric synthetic aperture radar (InSAR)-derived slip is usually not well-constrained at shallow depths due to decorrelation. Because optical matching directly measures the surface slip, we re-examine the slip distribution of 11 strike-slip earthquakes and find that the SSD model may primarily be artifacts in the InSAR measurements. It is therefore of great importance to include the optical data in earthquake studies to constrain coseismic slip inversions.

Road Extraction from Very High Resolution Images Using Weakly labeled OpenStreetMap Centerline

Road networks play a significant role in modern city management. It is necessary to continually extract current road structure, as it changes rapidly with the development of the city. Due to the success of semantic segmentation based on deep learning in the application of computer vision, extracting road networks from VHR (Very High Resolution) imagery becomes a method of updating geographic databases. The major shortcoming of deep learning methods for road networks extraction is that they need a massive amount of high quality pixel-wise training datasets, which is hard to obtain. Meanwhile, a large amount of different types of VGI (volunteer geographic information) data including road centerline has been accumulated in the past few decades. However, most road centerlines in VGI data lack precise width information and, therefore, cannot be directly applied to conventional supervised deep learning models. In this paper, we propose a novel weakly supervised method to extract road networks from VHR images using only the OSM (OpenStreetMap) road centerline as training data instead of high quality pixel-wise road width label. Large amounts of paired Google Earth images and OSM data are used to validate the approach. The results show that the proposed method can extract road networks from the VHR images both accurately and effectively without using pixel-wise road training data.

From woody cover to woody canopies: How Sentinel-1 and Sentinel-2 data advance the mapping of woody plants in savannas

Woody vegetation is a central component of savanna ecosystems providing ecosystem services for local livelihoods. Accurate monitoring of woody vegetation in savannas is therefore desirable, yet large scale mapping approaches rely on relatively coarse spatial resolution satellite data, which cannot directly capture the scattered nature of savanna trees. Studies at regional scale thus estimate the fractional cover of woody plants for a given area, whereas binary tree/no-tree estimates are restricted to the use of very high-resolution (VHR) images at local scales. With the launch of the Sentinel satellite systems (Sentinel-1 and Sentinel-2), the spatial resolution of images approaches the size of medium/large tree crowns, providing the opportunity to map the presence/absence of tree canopies, rather than the fraction of woody cover or forested areas. Here, we used a support vector machine (SVM) to classify the presence/absence of woody canopies from Sentinel-1 and Sentinel-2 data at a 10-m spatial resolution for the entire African Sahel. Training samples for the SVM classifier were collected from VHR images provided by Google Earth and Sentinel satellite data were processed in Google Earth Engine. Accuracy assessment was performed based on independent VHR images, showing an overall accuracy of 93% (71% and 98% for producer’s accuracy of woody and non-woody pixels, 91% and 93% for user’s accuracy of woody and non-woody pixels) when combining Sentinel-1 and Sentinel-2 data (overall accuracy of 89% using Sentinel-1 only and 91% using Sentinel-2 only). The combined use proved to perform significantly better (p < 0.05) than the single use of any of the two. A comparison with existing tree cover maps (by aggregating presence/absence of tree canopies into fractional cover) showed noticeable differences, reflecting the need for new woody cover products adapted to the nature of savanna ecosystems. The Sentinel woody canopy map was able to reproduce the general pattern of scattered woody canopies, but generally overestimated the woody coverage (11.37 ± 26.13% (mean ± sd) when aggregating to 250 m resolution) due to the 10 × 10-m spatial resolution which exceeds the crown size of a typical savanna tree. The cloud-based Sentinel-1 and Sentinel-2 analysis presented is a step towards large scale mapping of woody canopy (tree/no-tree) in savannas. Ultimately, such direct assessment of woody canopy areas will allow monitoring of temporal dynamics of woody canopies in future studies as Sentinel time-series expands to multiple years.