High Spatial Resolution Aerial Imagery Research Articles

Comparison of Deep and Machine Learning Approaches for Quebec Tree Species Classification Using a Combination of Multispectral and LiDAR Data

Accessing tree species information is required for making appropriate decisions in forest management. 3D photo-interpretation using high spatial resolution aerial imagery is used to provide information on tree species in the province of Quebec. However, the shortage of qualified interpreters and the increasing costs of 3D photo-interpretation have affected the production of the forest inventory. In this study, we employed deep and machine learning models to classify nine tree species (i.e., paper birch, yellow birch, red maple, poplar, black spruce, white spruce, tamarack, jack pine, and balsam fir). We used a combination of spectral and vertical structural information extracted from 30-cm aerial imagery and airborne light detection and ranging (LiDAR) data in a 9 , 100 km 2 forested area in Quebec, Canada. The results indicated that Dense Convolution Network (DenseNet) achieved the best overall accuracy of 78%, outperforming machine learning methods by 5%. In addition, the models’ performance was independently assessed using permanent and temporary ground sample plots, acknowledging the superiority of DenseNet in terms of overall accuracy (73%) in predicting the dominant species. Our results suggest that the combination of aerial imagery and airborne LiDAR data, using deep learning approaches, can be applied to accurately map tree species.

Assessing beech bark-diseased forest canopies over landscapes using high resolution open-source imagery in an ecological framework

Locating and tracking outbreaks of pests and pathogens enables forest ecologists and managers to assess impact or develop interventions. Beech bark disease (BBD) is actively affecting eastern US forests, including within the upper Great Lakes region. Our goal was to develop and apply an open-source framework for identification of BBD-affected tree crowns and for assessment of their spatial–temporal distribution over heterogeneous forested landscapes. We used the freely available USDA NAIP high spatial resolution aerial imagery, developed the identification algorithm in the open-source R environment, and integrated these within several open-data ecological classifications. The study landscape was the uplands of the ∼ 4200-ha University of Michigan Biological Station (UMBS) in northern Lower Michigan. Our objectives were to: 1) collect field data to characterize BBD on the landscape and to train and test remote sensing methods; 2) compile mapped ecological classifications; 3) explore NAIP spectral separability of BBD-affected crowns from other types of interest; 4) develop and apply a mostly-automated segmentation algorithm for identification of BBD-affected tree crowns; 5) assess mapped results over the entire landscape and within ecological classification strata; and 6) suggest an ecologically based framework for extensibility beyond the study landscape. Of a sample of visually interpreted yellowed tree crowns on NAIP at Training Sites, 99% were field-verified as beech with Neonectria. Spectral explorations confirmed affected (yellow) beech crowns had consistently higher reflection in NAIP red and green bands and were distinct from healthy beech and aspen undergoing age-related mortality; there was some spectral confusion with yellowed herbaceous areas. The segmentation algorithm was accurate to 82 %, and over the 2012–2018 biennial time series mapped > 9,700 BBD-affected canopies (single or interspersed beech tree crowns). Optimized Getis-Ord Gi* showed significant clustered distributions of BBD hotspots and coldspots on the landscape at each date. Ecological classifications presented strata statistically associated with the greatest density of mapped BBD-affected canopies and provide a framework for extensibility. This integrated approach may provide an accessible option for forest managers to map and assess hotspots of canopy-level BBD symptoms over heterogeneous landscapes and over time.

INTEROPERABILITY OF VARIOUS DATA STREAMS WITHIN GUYANA’s MRV SYSTEM

Abstract. In 2010 Guyana started work on the development of a national Monitoring Reporting and Verification System (MRVs) to quantify and measure the changes in the country’s forest cover carbon and carbon emissions. A necessary part of this process involved the identification of reliable Earth Observation data of sufficient resolution to detect and quantify land use change in Guyana's forests. Over the past 10 years the MRVs has evaluated and integrated many data streams used to; map national-scale forest change, support the analysis, and importantly datasets suitable to determine the accuracy of the change area mapped. Guyana’s approach has evolved over time, to accommodate new technologies, but at its core the MRVs recognises the importance of local management, existing datasets and linking these elements to appropriate EO data such as Landsat, RapidEye, Sentinel, Planet Scope and very high spatial resolution aerial imagery. From the outset the MRVs development was divided into phases. This approach recognises that not all MRVs reporting functions can be satisfied immediately. For Phase 1 (Years 2010 to 2014) of the MRVS, historical change analysis was conducted using Landsat 30 m resolution imagery. Being a persistently cloudy country alternative EO data sources were included, with Landsat and DMC imagery largely superseded by 5 m resolution RapidEye imagery. After five years of monitoring the forest change baselines, methods, reporting processes and standard operating procedures had been well established and able to provide the required performance-based indicators. The focus of MRVs phase 2 (Years 2015 to 2019) was to retain the reporting standards already achieved while also streamlining processes, improving functionality and reducing operational costs (i.e. the reliance on commercial image data) post-2019. Process improvements and operational research targeted two areas; the feasibility of using freely available Landsat and 10 metre resolution Sentinel data to map countrywide deforestation, and development of a sample-based approach to estimate degradation from aerial imagery and Planet Scope (3–5 metres resolution). Guyana’s move to integrate multiple data streams has been driven by the need for higher temporal resolution, repeated monitoring, and the creation of a data agnostic system to supports multiple decision-making processes on forest management. While the originally the MRVs was intended to support REDD+ reporting, the flexibility of the system design has meant it is proving to be invaluable tool for natural resources management in Guyana.

Deep Convolutional Neural Networks for Automated Characterization of Arctic Ice-Wedge Polygons in Very High Spatial Resolution Aerial Imagery

The microtopography associated with ice-wedge polygons governs many aspects of Arctic ecosystem, permafrost, and hydrologic dynamics from local to regional scales owing to the linkages between microtopography and the flow and storage of water, vegetation succession, and permafrost dynamics. Wide-spread ice-wedge degradation is transforming low-centered polygons into high-centered polygons at an alarming rate. Accurate data on spatial distribution of ice-wedge polygons at a pan-Arctic scale are not yet available, despite the availability of sub-meter-scale remote sensing imagery. This is because the necessary spatial detail quickly produces data volumes that hamper both manual and semi-automated mapping approaches across large geographical extents. Accordingly, transforming big imagery into ‘science-ready’ insightful analytics demands novel image-to-assessment pipelines that are fueled by advanced machine learning techniques and high-performance computational resources. In this exploratory study, we tasked a deep-learning driven object instance segmentation method (i.e., the Mask R-CNN) with delineating and classifying ice-wedge polygons in very high spatial resolution aerial orthoimagery. We conducted a systematic experiment to gauge the performances and interoperability of the Mask R-CNN across spatial resolutions (0.15 m to 1 m) and image scene contents (a total of 134 km2) near Nuiqsut, Northern Alaska. The trained Mask R-CNN reported mean average precisions of 0.70 and 0.60 at thresholds of 0.50 and 0.75, respectively. Manual validations showed that approximately 95% of individual ice-wedge polygons were correctly delineated and classified, with an overall classification accuracy of 79%. Our findings show that the Mask R-CNN is a robust method to automatically identify ice-wedge polygons from fine-resolution optical imagery. Overall, this automated imagery-enabled intense mapping approach can provide a foundational framework that may propel future pan-Arctic studies of permafrost thaw, tundra landscape evolution, and the role of high latitudes in the global climate system.

MAPPING THE RISK OF FOREST WIND DAMAGE USING AIRBORNE SCANNING LiDAR

Abstract. Wind damage is known for causing threats to sustainable forest management and yield value in boreal forests. Information about wind damage risk can aid forest managers in understanding and possibly mitigating damage impacts. The objective of this research was to better understand and quantify drivers of wind damage, and to map the probability of wind damage. To accomplish this, we used open-access airborne scanning light detection and ranging (LiDAR) data. The probability of wind-induced forest damage (PDAM) in southern Finland (61°N, 23°E) was modelled for a 173 km2 study area of mainly managed boreal forests (dominated by Norway spruce and Scots pine) and agricultural fields. Wind damage occurred in the study area in December 2011. LiDAR data were acquired prior to the damage in 2008. High spatial resolution aerial imagery, acquired after the damage event (January, 2012) provided a source of model calibration via expert interpretation. A systematic grid (16 m x 16 m) was established and 430 sample grid cells were identified systematically and classified as damaged or undamaged based on visual interpretation using the aerial images. Potential drivers associated with PDAM were examined using a multivariate logistic regression model. Risk model predictors were extracted from the LiDAR-derived surface models. Geographic information systems (GIS) supported spatial mapping and identification of areas of high PDAM across the study area. The risk model based on LiDAR data provided good agreement with detected risk areas (73 % with kappa-value 0,47). The strongest predictors in the risk model were mean canopy height and mean elevation. Our results indicate that open-access LiDAR data sets can be used to map the probability of wind damage risk without field data, providing valuable information for forest management planning.

Parameter optimization of image classification techniques to delineate crowns of coppice trees on UltraCam-D aerial imagery in woodlands

Estimating the optimal parameters of some classification techniques becomes their negative aspect as it affects their performance for a given dataset and reduces classification accuracy. It was aimed to optimize the combination of effective parameters of support vector machine (SVM), artificial neural network (ANN), and object-based image analysis (OBIA) classification techniques by the Taguchi method. The optimized techniques were applied to delineate crowns of Persian oak coppice trees on UltraCam-D very high spatial resolution aerial imagery in Zagros semiarid woodlands, Iran. The imagery was classified and the maps were assessed by receiver operating characteristic curve and other performance metrics. The results showed that Taguchi is a robust approach to optimize the combination of effective parameters in these image classification techniques. The area under curve (AUC) showed that the optimized OBIA could well discriminate tree crowns on the imagery (AUC=0.897), while SVM and ANN yielded slightly less AUC performances of 0.819 and 0.850, respectively. The indices of accuracy (0.999) and precision (0.999) and performance metrics of specificity (0.999) and sensitivity (0.999) in the optimized OBIA were higher than with other techniques. The optimization of effective parameters of image classification techniques by the Taguchi method, thus, provided encouraging results to discriminate the crowns of Persian oak coppice trees on UltraCam-D aerial imagery in Zagros semiarid woodlands.

Tree crown delineation on UltraCam-D aerial imagery with SVM classification technique optimised by Taguchi method in Zagros woodlands

Support vector machine (SVM) is a superior machine learning methodology with great results in classification of remotely sensed data sets. Determination of optimal parameters applied in SVM classification technique is still vague to some scientists. This study was aimed to detect tree crowns on UltraCam-D (UCD), very high spatial resolution aerial imagery in Zagros woodlands by SVM optimised by Taguchi method. A 500 × 600 m2 plot covered with Persian oak (Quercus brantii var. persica) coppice trees was selected in Zagros woodlands, Iran. The UCD aerial imagery of the plot (0.06-m spatial resolution) was obtained to extract crowns of the trees in this study. The SVM classification technique parameters were optimised by Taguchi method, and thereafter, the imagery was classified with optimal parameters. The results showed that Taguchi method is a robust approach to optimise the combination of parameters of SVM classification technique. It was also concluded that the technique could detect the tree crowns with a KHAT coefficient of 0.961, which showed a great agreement with the observed samples, and overall accuracy of 97.7% that showed the accuracy of the final map. Finally, the authors suggest applying this method to optimise the parameters of classifiers like SVM classification technique.

Early detection of emerald ash borer infestation using multisourced data: a case study in the town of Oakville, Ontario, Canada

The emerald ash borer (EAB) poses a significant economic and environmental threat to ash trees in southern Ontario, Canada, and the northern states of the USA. It is critical that effective technologies are urgently developed to detect, monitor, and control the spread of EAB. This paper presents a methodology using multisourced data to predict potential infestations of EAB in the town of Oakville, Ontario, Canada. The information combined in this study includes remotely sensed data, such as high spatial resolution aerial imagery, commercial ground and airborne hyperspectral data, and Google Earth imagery, in addition to nonremotely sensed data, such as archived paper maps and documents. This wide range of data provides extensive information that can be used for early detection of EAB, yet their effective employment and use remain a significant challenge. A prediction function was developed to estimate the EAB infestation states of individual ash trees using three major attributes: leaf chlorophyll content, tree crown spatial pattern, and prior knowledge. Comparison between these predicted values and a ground-based survey demonstrated an overall accuracy of 62.5%, with 22.5% omission and 18.5% commission errors.

Automated tree crown delineation from imagery based on morphological techniques

In current tree crown delineation from imagery, treetops and three dimensional (3D) radiometric shapes of tree crowns are frequently extracted from a spectral band or a brightness component of the image and taken as references to localize and delineate tree crowns. However, color components of the image are rarely used together with the brightness component of the image to facilitate localizing and delineating crowns. The 3D radiometric shape of a crown can be derived from a brightness or color component and may be taken as a half-ellipsoid. From top to bottom of such a half-ellipsoid, multiple horizontal slices can be drawn, contain the treetop, and indicate both the location and the horizontal extent of the crown. Based on such a concept of horizontal slices of crowns, a novel multi-scale method for individual tree crown delineation from imagery was proposed in this study. In this method, the brightness and color components of the image are morphologically opened within the scale range of target crowns, horizontal slices of target crowns are extracted from the resulting opened images and integrated together to localize crowns, and one component is segmented using the watershed approach with reference to the integrated slices. In an experiment on high spatial resolution aerial imagery over natural closed canopy forests, the proposed method correctly delineated approximately 74% of mixedwood tree crowns and 59% of deciduous crowns in the natural forests.

CANOPY DENSITY MAPPING ON ULTRACAM-D AERIAL IMAGERY IN ZAGROS WOODLANDS, IRAN

Abstract. Canopy density maps express different characteristics of forest stands, especially in woodlands. Obtaining such maps by field measurements is so expensive and time-consuming. It seems necessary to find suitable techniques to produce these maps to be used in sustainable management of woodland ecosystems. In this research, a robust procedure was suggested to obtain these maps by very high spatial resolution aerial imagery. It was aimed to produce canopy density maps by UltraCam-D aerial imagery, newly taken in Zagros woodlands by Iran National Geographic Organization (NGO), in this study. A 30 ha plot of Persian oak (Quercus persica) coppice trees was selected in Zagros woodlands, Iran. The very high spatial resolution aerial imagery of the plot purchased from NGO, was classified by kNN technique and the tree crowns were extracted precisely. The canopy density was determined in each cell of different meshes with different sizes overlaid on the study area map. The accuracy of the final maps was investigated by the ground truth obtained by complete field measurements. The results showed that the proposed method of obtaining canopy density maps was efficient enough in the study area. The final canopy density map obtained by a mesh with 30 Ar (3000 m2) cell size had 80% overall accuracy and 0.61 KHAT coefficient of agreement which shows a great agreement with the observed samples. This method can also be tested in other case studies to reveal its capability in canopy density map production in woodlands.

Digital high spatial resolution aerial imagery to support forest health monitoring: the mountain pine beetle context

We summarize the capacity of high spatial resolution (< 1m ) digital aerial imagery to support forest health monitoring. We review the current use of digital aerial imagery in the con- text of the recent mountain pine beetle epidemic in western Canada. Supported by this review, we posit that high spatial resolution digital aerial imagery can play at least two critical roles in forest health monitoring. First, the capacity to characterize damage at the individual tree level directly supports a broad range offorest health information needs (e.g., tree-level attributes for estimating the population at risk and for inputs to models, estimates of mortality, rates of population growth). Second, the level of detail afforded by the digital high spatial resolution aerial imagery provides critical calibration and validation data for lower spatial resolution remotely sensed ima- gery (e.g., QuickBird, Landsat) for large-area detection and mapping of forest damage and can be used in a double sampling scheme as a bridge between detailed field measures and landscape- level estimates of mortality. In an era with increasing numbers of commercially deployed sensors capable of acquiring high spatial resolution satellite imagery, the flexibility and cost- effectiveness of aerial image options should not be disregarded. Moreover, experiences with airborne imagery can continue to inform applications using high spatial resolution satellite ima- gery for forest health information needs. © 2012 Society of Photo-Optical Instrumentation Engineers (SPIE). (DOI: 10.1117/1.JRS.6.062527)

Monitoring the impacts of mountain pine beetle mitigation

Since 1999, the mountain pine beetle ( Dendroctonus ponderosae Hopk. [Coleoptera: Scolytidae]) has impacted over 13 million hectares of pine forests in western Canada. Successful mitigation of the beetle depends on the accurate and timely identification of currently infested trees and on sustained control activities over several successive years. We monitored the success of mitigation activities in reducing damage caused by mountain pine beetle at two sites (A and B) on the leading edge of the current beetle epidemic in western Canada. Using three years of digital high spatial resolution aerial imagery (2006–2008) and one season of field measurements (2008), we estimated retrospective ratios of trees attacked by beetle in the current year (green attack) to trees attacked in the previous year (red attack), hereafter referred to as G:R. Our results indicate that mitigation activities slowed the rate of population growth, with G:R found to be decreasing or stable over sites A and B while mitigation was ongoing in 2005 and 2006 (site A 1.06:1; site B 0.32:1). When mitigation was discontinued over site A in 2007, the G:R increased markedly (1.94:1), while continued mitigation at site B in 2007 further reduced the G:R (0.22:1). Despite the cost associated with mitigation, its efficacy is rarely assessed and even more rarely documented. The approach presented herein enables a sample-based appraisal of mitigation efforts.

Urban cover mapping using digital, high-spatial resolution aerial imagery

High-spatial resolution digital color-infrared aerial imagery of Syracuse, NY was analyzed to test methods for developing land cover classifications for an urban area. Five cover types were mapped: tree/shrub, grass/herbaceous, bare soil, water and impervious surface. Challenges in high-spatial resolution imagery such as shadow effect and similarity in spectral response between classes were found. Classification confusion among objects with similar spectral responses occurred between water and dark impervious surfaces, concrete and bare- soil, and grass/herbaceous and trees/shrub. Methods of incorporating texture, band ratios, masking of water objects, sieve functions, and majority filters were evaluated for their potential to improve the classification accuracy. After combining these various techniques, overall cover accuracy for the study area was 81.75%. Highest accura- cies occurred for water (100%), tree/shrub (86.2%) and impervious surfaces (82.6%); lowest accuracy were for grass/herbaceous (69.3%) and bare soil (40.0%). Methods of improving cover map accuracy are discussed.