The building footprint extraction (BFE) from aerial images is important for the creation and continuous monitoring of building inventories useful for urban planning, among others. Existing methods frequently extract roofs of buildings from aerial images assuming that they overlap with the footprint. This assumption does not hold in the case of off-nadir images. This letter proposes a novel multilabel learning of oblique building features—footprint, roof, and shape—with a Vision Transformer (ViT) for accurate BFE from off-nadir aerial images. A shape calculation algorithm is developed to derive shape polygons from the existing footprint and roof polygons. The method is compared with several convolutional neural networks (CNNs) and ViTs, and a postprocessing algorithm is further devised to achieve regular building footprint polygons. The proposed method outperforms existing scores of BFE on the BONAI dataset (0.727 versus 0.643F1), and our shape calculation algorithm provides labels as accurate as Segment Anything 2 without the need for a GPU. The results conclude that models trained with shapes in addition to the footprint and roof provide consecutively higher scores (F1 score: 0.747 w/ shape versus 0.727 w/o shape versus 0.682 w/ only footprint) and substantially improve the BFE on off-nadir images. The codes and datasets are available at:https://github.com/bipulneupane/Multilabel-BONAI/.

City planning and optimal land use design are becoming challenges considering environmental sustainability, urban growth, and citizen science. Earth observation data sources contribute significantly to preparing accurate man-made infrastructure like buildings and their varieties in addressing such challenges. However, despite the efforts in designing and managing urban buildings from Earth Observation datasets, there is a clear gap in knowledge in developing the spatial ecosystem of existing urban building management. This paper draws a perspective by developing analytics on representative urban building footprint datasets with a focus on data quality issues. Issues such as the off-nadir imagery, omission/commission errors, and positional accuracy are considered. These datasets are generated by the community for the public good and by the industry sectors for commercial benefits. This perspective and the proposed spatial ecosystem ultimately help in developing an operational framework for urban infrastructure in sustainable city planning.

Hyperspectral remote sensing of advanced plant traits and physiological status are explored for leaf nitrogen (N) and phosphorus (P) monitoring in the context of sustainable forestry. Previous studies on agricultural species have demonstrated that plant biochemical and biophysical constituents derived via radiative transfer models (RTMs), and other parameters, such as solar-induced chlorophyll fluorescence (SIF), provided an improved prediction of leaf N compared to traditional methods based on chlorophyll indices. In this study, we assessed the transferability of such methods to assess needle N and P in coniferous canopies, where highly heterogeneous tree-crown structures dominate. Our study across three years showed that RTM-based functional traits (chlorophyll a+b (Ca+b), carotenoids (Car), anthocyanins (Anth), and Leaf Area Index (LAI)) along with SIF could provide moderate prediction accuracy for N (Ca+b, Car, Anth, SIF: R2=0.4 to 0.72) and P (Ca+b, LAI, SIF: R2=0.4 to 0.73). Furthermore, this work highlighted that SIF contributed to needle P and N assessment to different degrees.

Building footprints are extracted using convolutional neural networks (CNNs) on airborne image data. However, the method suffers from the need for large accurately labelled training data and computational cost. We develop a novel open mutual learning (OML), an extension of deep mutual learning (DML), to leverage noise-tolerant lightweight networks for precise building extraction from open datasets. OML distils n lightweight Student networks from n datasets and one Teacher network while penalising each Student by the incompleteness of the dataset used to distil them. Three new Australian building footprint datasets are developed with labels from − OpenStreetMap (OSM), Microsoft’s Building Footprints (MBF), and Geoscape Buildings (GBs) − and images from Google, Bing, and ESRI. OML is evaluated with lightweight CNNs from Apple and Google. Compared to DML, OML distils lightweight networks of upto 96.4% fewer network parameters with performance gains of upto 13.5% IoU and 12.5% F1 in the GBs dataset.

Accurate extraction of urban features such as buildings and roads lays the foundation for the current trends of digital twins of urban systems to support planning, monitoring, navigation, and decision processes. The process of such extraction involves training convolutional neural networks (CNNs) on high-resolution earth observation (EO) images. The spatial resolution of images has increased to a centimetre level and the CNNs are fast evolving in computer vision. The last 10 years of this development have resulted in both high-performance and computationally efficient CNNs, but they are merely benchmarked under a uniform setting. We present a survey-driven benchmark of CNNs starting with a systematic survey of 165 research articles to understand the state-of-the-art of urban feature extraction. The survey looks for the most prominent urban feature, EO source, benchmark dataset, CNN-based deep learning configuration, and hyperparameters. Further, more CNNs are searched in the computer vision domain. Identified from the survey and search, 65 CNNs are trained and evaluated in an encoder–decoder configuration using a benchmark dataset under uniform settings. Extensive hyperparameter tuning of the best-performing CNN is performed with six optimisers and nine loss functions. The tuned CNN is then tested as an encoder in other state-of-the-art encoder–decoder networks. The CNNs and network configurations with the highest scores are further benchmarked on the Massachusetts Building and WHU Building datasets. The findings from this survey-driven benchmark of CNNs will be useful for both academia and industry involved in the science of earth observation and computer vision.

Vegetation is crucial for sustainable and resilient cities providing various ecosystem services and well-being of humans. However, vegetation is under critical stress with rapid urbanization and expanding infrastructure footprints. Consequently, mapping of this vegetation is essential in the urban environment. Recently, deep learning (DL) for point cloud semantic segmentation has shown significant progress. Advanced models attempt to obtain state-of-the-art performance on benchmark datasets, comprising multiple classes and representing real-world scenarios. However, class-specific segmentation with respect to vegetation points has not been explored. Therefore, selection of a DL model for vegetation points segmentation is ambiguous. To address this problem, we provide a comprehensive assessment of point-based DL models for semantic segmentation of vegetation class. We have selected seven representative-point-based models, namely, PointCNN, KPConv (omni-supervised), RandLANet, SCFNet, PointNeXt, SPoTr, and PointMetaBase. These models are investigated on three different datasets, specifically Chandigarh, Toronto3D, and Kerala, which are characterized by diverse nature of vegetation and varying scene complexity combined with changing per-point features and classwise composition. PointMetaBase and KPConv (omni-supervised) achieve the highest mIoU on the Chandigarh (95.24%) and Toronto3D datasets (91.26%), respectively while PointCNN provides the highest mIoU on the Kerala dataset (85.68%). The article develops a deeper insight, hitherto not reported, into the working of these models for vegetation segmentation and outlines the ingredients that should be included in a model specifically for vegetation segmentation. This article is a step toward the development of a novel architecture for vegetation points segmentation.

Deep learning (DL) models combined with mobile laser scanning (MLS) datasets have demonstrated immense potential for vegetation segmentation. However, restricted performance and inconsistent behavior across datasets by generic DL models offer notable concerns. Furthermore, to capture the characteristic distribution of vegetation points toward effective segregation, a dedicated model for vegetation segmentation is essential. In addition, with curated class-specific DL models being conceptualized, the same is indispensable for vegetation. To address this problem, we propose a novel DL architecture, green segmentation network (GreenSegNet), tailored for vegetation segmentation from MLS point cloud data. Toward a comprehensive assessment, GreenSegNet has been investigated on MLS datasets from three study sites, Chandigarh, Toronto3D, and Kerala. GreenSegNet has illustrated state of the art (SOTA) as well as consistent segmentation performance across all the datasets. GreenSegNet has achieved mean intersection over union (mIoU) as follows: Chandigarh 96.43%, Toronto3D 92.70%, and Kerala 90.16%. In addition, with less than one million parameters, the architecture is the most efficient with respect to the number of parameters among the representative DL models. The associated ablation studies conform to the effectiveness of GreenSegNet. Unlike other SOTA models, GreenSegNet is found robust across different datasets and terrains.

Agricultural field boundary information is an essential input for precision agriculture. This paper proposes a Multi-scale Multi-task Boundary Detection Deep Learning (DL) Network (MMBDNet) based on spatial attention mechanisms to delineate agricultural fields using high-resolution optical satellite imagery. The designed DL architecture simultaneously learns three tasks - a major task for field prediction and two auxiliary tasks for boundary prediction and distance estimation. We experimented with the agricultural landscape of Île-de-France, France, using the cloud-free time-series images from PlanetScope satellite that capture key phenological stages of crops. The segmentation results from different months are combined and post-processed using hierarchical watershed segmentation to extract field instances. We compared the MMBDNet with the baseline single-task U-Net and multitask BsiNet models at pixel- and object-level. Our results show that the MMBDNet has the highest pixel-level (above 85%) and object-level (above 70%) accuracy compared to U-Net and BsiNet.

Convolutional neural networks (CNNs) have achieved high precision in the extraction of low-rise buildings from very high-resolution (VHR) remote sensing images. However, there are two major challenges with high-rise buildings: i) off-nadir aerial images causing the misalignment in the high-risers and their labels, and (ii) finding the CNN of the right size fit for the available data. To address these challenges, we develop a workflow for two knowledge transfer techniques namely, knowledge distillation and supervised domain adaptation with three new multi-resolution urban building datasets: (i) teacher data from off-nadir images, (ii) student data from orthorectified images, and (iii) evaluation data with off-nadir images. The student and evaluation data include complex high-rise and skyscrapers buildings. The results show that knowledge distillation and supervised domain adaptation can minimise the effects of misaligned labels and possible domain shifts while also compressing network parameters by at least 74.99% and 99.5%.

This study evaluate the performance of a Wheat Canopy Scattering Model (WCSM) at L-band, which was initially developed to simulate the backscatter of C-band Synthetic Aperture Radar (SAR), using the L-band UAVSAR data and ground-based measurements of soil moisture, soil surface roughness and crop parameters collected from wheat fields during the SMAPVEX12 field campaign. Results show that WCSM is capable of estimating HH-pol backscatter with an error less than 2.48 dB. On the other hand, relatively large RMSE of 4.38 dB and 4.34 dB were observed for VV and VH backscatter coefficients, respectively. Furthermore, it was observed that model tends to overestimate VV backscatter. It is also observed that co-pol total backscatter from a wheat canopy is sensitive to incidence angle followed by root mean square (RMS) height and soil moisture while cross-pol backscatter showing high sensitivity to wheat crop biophysical parameters.

Classification of very high-resolution (VHR) aerial remote sensing (RS) images is a well-established research area in the RS community as it provides valuable spatial information for decision-making. Existing works on VHR aerial RS image classification produce an excellent classification performance; nevertheless, they have a limited capability to well-represent VHR RS images having complex and small objects, thereby leading to performance instability. As such, we propose a novel plug-and-play multiscale attention feature extraction block (MSAFEB) based on multiscale (MS) convolution at two levels with skip connection, producing discriminative/salient information at a deeper/finer level. The experimental study on two benchmark VHR aerial RS image datasets (AID and NWPU) demonstrates that our proposal achieves a stable/consistent performance (minimum standard deviation (SD) of 0.002) and competent overall classification performance (AID: 95.85% and NWPU: 94.09%).

Beetle infestations have caused significant damage to the pine forest in North America. Early detection of beetle infestation in near real time is crucial, in order to take appropriate steps to control the damage. In this letter, we consider near-real-time detection of beetle infestation in North American pine forests using high temporal resolution and coarse spatial resolution MODIS (eight-day 500-m) satellite data. We show that the parameter sequence of a stationary vegetation index time series, which is derived by fitting an underlying triply modulated cosine model over a sliding window using nonlinear least squares, resembles a martingale sequence. The advantage of such properties of the parameter sequence is that standard martingale central limit theorem and well-known Gaussian distribution statistics can be effectively used to detect any nonstationarity in the vegetation index time series with high accuracy. The proposed method exploits these properties of the parameter time series and, hence, does not require threshold tuning. The threshold is selected based on a well-documented procedure of z-value selection from the table of Gaussian distribution, depending upon the percentage of the distribution considered as outlier. The proposed framework is tested on different vegetation index data sets derived from MODIS eight-day 500-m image time series of beetle infestations in North America. The results show that the proposed framework can detect nonstationarities in the vegetation index time series accurately and performs the best on red-green index.

The paper considers the detection of beetle infestations in North American pine forests using high temporal resolution, coarse spatial resolution MODIS remotely sensed satellite images. Two methods are proposed to detect beetle infestation, both applying a triply modulated cosine model. The first method uses an Extended Kalman Filter (EKF) for estimating model parameters, and the second a Least Squares estimator. When beetles infest a forest, the changes in the affect large geographical area. Therefore, the change detection metrics are based on the time series of each pixel, and do not utilize information from neighboring pixels. Using data from the Rocky Mountain region of the United States and of British Columbia in Canada, we show that our methods are highly effective at detecting beetle infestations.

With the global availability and accessibility of environmental data sources it is possible to address the water related problems. Locally, in the Australian context, the water industry is in a unique position due to the extremes with a vast experience of drought and flood conditions. Water in Australia is a national priority and there is a need to develop an accurate and timely decision support system regarding efficient and optimal water usage. To address this issue, in this paper, we proposed an integrated environmental knowledge recommendation system based on large scale dynamic web data mining and contextual knowledge integration to provide an expert water resource management solution. We integrated five different environmental data sources namely SILO, AWAP, ASRIS, CosmOz, and MODIS imagery to develop and test the proposed knowledge recommendation framework called intelligent Environmental knowledgebase (i-Ekbase). The developed system was tested for its robustness and applicability.

Thematic maps prepared from remotely sensed images require a statistical accuracy assessment. For this purpose, the /spl kappa/-statistic is often used. This statistic does not distinguish between whether one unit is classified as another, or vice versa. In this paper, the Bradley-Terry (BT) model is applied for accuracy assessment. This model compares categories pairwise. The probability of one class over another class is estimated as well as the expected values of class pixels. The study is illustrated with an Advanced Spaceborne Thermal Emission and Reflection Radiometer image from the Netherlands, to which a maximum-likelihood classification with the Euclidean distance is applied. An error matrix is generated using an IKONOS image from the same area as ground truth. It is shown to which degree the BT model extends the /spl kappa/-statistic. A comparison with the Mahalanobis distance is made. Standardization is carried out to overcome problems emerging from the fact that a common BT model does not include the number of correctly classified pixels. The study shows how the BT model serves as an alternative to the usual /spl kappa/-statistic.