Species classification of individually segmented tree crowns in high-resolution aerial images using radiometric and morphologic image measures

doi:10.1016/j.rse.2004.04.006

Remote Sensing of Environment

Volume 91, Issues 3–4, 30 June 2004, Pages 469-477

https://doi.org/10.1016/j.rse.2004.04.006 Get rights and content

Abstract

This paper presents a method to automatically classify segmented tree crowns from high spatial resolution colour infrared aerial images as one of the four most common tree species in Sweden. The species are Norway spruce (Picea abies Karst.), Scots pine (Pinus sylvestris L.), birch (Betula pubescens Ehrh.), and aspen (Populus tremula L.). The proposed method uses four different image measures, one measure for each species. The measures are based on colour information as well as the shape of the segmented tree crowns. A segment is examined by the measures one by one and if one measure becomes true, the segment is interpreted as that species. The analysis continues with the next segment. The method is evaluated on two sets of images. The first set consists of 14 images of naturally regenerated forest with pixel size corresponding to 3 cm. These images contain approximately 50 visible tree crowns each; a total of 791 crown segments are used. The overall classification result for these images is 77%. If only the distinction between conifers and deciduous is made, the result is 91%. The second set consists of two images with a pixel size of 10 cm. Here, the overall classification result is 71%.

Introduction

In forest inventory and management planning, a combination of field measurements and interpretation of aerial photographs is often used to find the resource data of interest. In Sweden, according to forest owners with more than 5000 ha of forest, the four most important parameters for forest stands are stem volume, age, tree species composition, and ecological values such as key biotopes and habitats (Walter, 1998).

To estimate the stem volume, segmentation of the individual tree crowns must be done, as well as identification of the species. Segmentation of individual tree crowns in high spatial resolution images has been an ongoing research field for several years. The techniques employ template matching, valley following, local maximum filtering, edge detection, spatial clustering. Many algorithms utilize combinations of these Brandtberg & Walter, 1998, Culvenor, 2002, Erikson, 2003a, Erikson, 2003b, Gougeon, 1995, Pinz, 1989, Pollock, 1996, Pouliot et al., 2002.

The results from many of these methods are rather good, although more research can probably improve the result. When it comes to classification of the tree crowns into species, less research has been done. Brandtberg (2002) developed a method for classification of the tree crowns using fuzzy sets and Gougeon et al. (1998) a method using spectral signatures. Leckie et al. (2003) have developed a method for classification at stand level.

The method presented in this paper classifies a segmented region, representing a tree crown, into one of the four most common species in Sweden: Norway spruce (Picea abies Karst.), Scots pine (Pinus sylvestris L.), birch (Betula pubescens Ehrh.), and aspen (Populus tremula L.). In order to perform the classification, four different measures are used, where each measure is used to detect a specific species. The measures are used sequentially on each tree crown and some extra rules are used if the tree remains unspecified after applying all of them.

Section snippets

Material

A data set of 50 colour infrared aerial photographs was captured on a sunny day, 10th August 1995, between 13:00 and 14:30 local time, from an aircraft at 600 m above ground level. The photographs were captured over an area of 5×5 km at Huljen outside Sundsvall in Central Sweden (62°27′N, 16°55′E) with Kodak Aerochrome Infrared Film 2443. During the acquisition, the solar zenith angle was between 46.9° and 49.7° and the azimuth angle between 180.7° and 209.6°. The wavelengths of the three

Methods

The classification is based on four different measures constructed to find the characteristics of each species. The first measure (M_b) finds birch, the second (M_a) aspen, the third (M_s) spruce, and the last (M_p) finds pine.

The outline of the algorithm using these four measures is shown in Algorithm 1. The algorithm stops analysing a segment and continues with the next as soon as one measure becomes true. If all four measures are false for a segment, it is stored in a list of unclassified

Results

One image example from the first data set is shown in Fig. 6 and one from the second in Fig. 7. The left image shows the original image and the right the result after classification.

The author evaluated the accuracy by photo-interpreting the original scanned images to decide the delineation as well as the species. This may introduce bias in the accuracy, but no other evaluation technique was available. If a segment is found completely wrong by the method and includes more than one tree crown,

Discussion

The evaluation is made by the author and can therefore introduce bias. Thus, field reference data must be collected and evaluated against the result of the method in order to get proper accuracy. This is a subject for future work.

The order of the measures is important and the only reordering that can be made is between M_b and M_a and between M_s and M_p. However, since the forest used here consists of more birch than aspen it is much more suitable to have them in the presented order. The same

Conclusions

The method presented in this paper shows that there is a potential for classifying individual tree crowns into species for resolution corresponding up to 10 cm by using radiometric and morphologic image measures if the tree crowns are seen straight from above. It also shows that many problems remain to be solved, including better segmentation methods, larger viewing angles, and the problem with different illuminations.

Acknowledgements

The author would like to thank his supervisor Prof. Gunilla Borgefors as well as Dr. Petra Philipson.

References (16)

G. Borgefors
Distance transformations in digital images
Computer Vision, Graphics, and Image Processing
(1986)
G. Borgefors et al.
Analyzing nonconvex 2D and 3D patterns
Computer Vision and Image Understanding
(1996)
T. Brandtberg
Individual tree-based species classification in high spatial resolution aerial images of forests using fuzzy sets
Fuzzy Sets and Systems
(2002)
D.S. Culvenor
TIDA: An algorithm for delineation of tree crowns in high spatial resolution remotely sensed imagery
Computers & Geosciences
(2002)
D.G Leckie et al.
Stand delineation and composition estimation using semi-automated individual tree crown analysis
Remote Sensing of Environment
(2003)
D.A. Pouliot et al.
Automated tree crown detection and delineation in high-resolution digital camera imagery of coniferous forest regeneration
Remote Sensing of Environment
(2002)
T. Brandtberg et al.
Automatic delineation of individual tree crowns in high spatial resolution aerial images by multiple-scale analysis
Machine Vision and Applications
(1998)
M. Erikson
Segmentation of individual tree crowns in colour aerial photographs using region growing supported by fuzzy rules
Canadian Journal of Forest Research
(2003)

There are more references available in the full text version of this article.

Cited by (83)

Detecting and mapping tree crowns based on convolutional neural network and Google Earth images
2022, International Journal of Applied Earth Observation and Geoinformation
Mapping tree crown is critical for estimating the functional and spatial distribution of ecosystem services. However, accurate and up-to-date urban crown mapping remains a challenge due to the time-consuming nature of field sampling and spatial heterogeneity. Another challenge is the data cost, which is always a concern for low-cost processing of forest maps on large scales. Here, we developed a novel working framework by integrating an advanced deep learning technology, the Mask Region-based Convolutional Neural Network (Mask R-CNN) model with Google Earth images to detect tree crown cover in New York’s Central Park, which is a typical testbed for an urban forest area with highly heterogeneous tree crown cover. The results indicated that the tree number detection rate estimated by the Mask R-CNN crown detection model was 82.8% and the crown area detection rate was 81.8% for the entire study area. The model detected isolated trees and closed forest trees areas with a recall of 87.5% and 81.6% of the tree numbers, respectively. The analysis indicates that the tree crown detection model could accurately detect tree crowns under highly complex environments and demonstrates great potential to map urban tree crown covers.
Estimating tree-related power outages for regional utility network using airborne LiDAR data and spatial statistics
2021, International Journal of Applied Earth Observation and Geoinformation
Trees play an integral role in the “green” framework of an urban ecosystem. However, just as they are beneficial to the environment, they can pose a significant risk to utility infrastructure networks, particularly in severe weather events. The objectives of this research were to explore the effect of scale and spatial variation on the relationships between trees and utility assets for vegetation-related power outages through the incorporation of remote sensing and geographic information system (GIS) analysis. Tree location and structural metrics derived from airborne Light Detection and Ranging (LiDAR) data were combined with regional utility network GIS data to test the prediction analysis capabilities of global and local statistics at multiple scales. Pearson’s correlation was carried out to examine the relationships between tree structure and utility asset variables to vegetation-related power outages, including the effect of the resolution, or grid-cell size, on those relationships. To test the performance of global and local regression modeling on outage prediction, ordinary least square (OLS) and geographically weighted regression (GWR) models were evaluated using four explanatory variables (utility wire length, utility pole count, tree canopy area, maximum tree height) at four different grid cell scales (50 m, 500 m, 1 km, 2 km). In general, Pearson’s correlation demonstrated the strongest positive relationship between explanatory variables and power outages when only aggregating 50-m grid cells exhibiting co-location of trees and utility assets to 2-km grid cells. Local regression models performed better than global models at all scales, with GWR producing the highest adjusted R² and lowest Akaike information criterion (AIC) values of 0.955 and 3213, respectively. Additionally, the performance of OLS and GWR models increased with scale as both models produced the highest adjusted R² at 2-km grid-cell scale. GWR model outputs demonstrated unique spatial patterning across the study area. This research demonstrated the effect of scale and spatial variation on regression analysis for the estimation of tree-related power outages.
A keystone species, European aspen (Populus tremula L.), in boreal forests: Ecological role, knowledge needs and mapping using remote sensing
2020, Forest Ecology and Management
European aspen (Populus tremula L.) is a keystone species in boreal forests that are dominated by coniferous tree species. Both living and dead aspen trees contribute significantly to the species diversity of forest landscapes. Thus, spatial and temporal continuity of aspen is a prerequisite for the long-term persistence of viable populations of numerous aspen-associated species. In this review, we collate existing knowledge on the ecological role of European aspen, assess the knowledge needs for aspen occurrence patterns and dynamics in boreal forests and discuss the potential of different remote sensing techniques in mapping aspen at various spatiotemporal scales. The role of aspen as a key ecological feature has received significant attention, and studies have recognised the negative effects of modern forest management methods and heavy browsing on aspen occurrence and regeneration. However, the spatial knowledge of occurrence, abundance and temporal dynamics of aspen is scarce and incomprehensive. The remote sensing studies reviewed here highlight particularly the potential of three-dimensional data derived from airborne laser scanning or photogrammetric point clouds and airborne imaging spectroscopy in mapping European aspen, quaking aspen (Populus tremuloides Michx.) and other Populus species. In addition to tree species discrimination, these methods can provide information on biophysical, biochemical properties and even genetic diversity of aspen trees. Major obstacles in aspen detection using remote sensing are the low proportion and scattered occurrence of European aspen in boreal forests and the overlap of spectral and/or structural properties of European aspen and quaking aspen with some other tree species. Furthermore, the suitability of remote sensing data for aspen mapping and monitoring depends on the geographical coverage of data, the availability of multitemporal data and the costs of data acquisition. Our review highlights that integration of ecological knowledge with spatiotemporal information acquired by remote sensing is key to understanding the current and future distribution patterns of aspen-related biodiversity.
A methodology based on GEOBIA and WorldView-3 imagery to derive vegetation indices at tree crown detail in olive orchards
2019, International Journal of Applied Earth Observation and Geoinformation
Precision agriculture (PA) is becoming an essential practice for the implementation of sustainable agriculture that encompasses the efficient use of resources and a systematic crops monitoring. The increasing temporal and spatial resolution of satellite imagery, coupled with their availability and decreasing costs, create new possibilities for generating accurate datasets on different crops variables, more frequently available as ready-to-use data. The availability of very high-resolution (VHR) satellite imagery, such as the WorldView-3 (WV-3), leads to the advanced potential of satellite Remote Sensing (RS), becoming in the last decade one of the main data source in precision agriculture (PA). In the broad overview of these procedures, geographic object-based image classification (GEOBIA) techniques, gained broad interest as methods to produce geographic information in GIS-ready format. In this paper, methodologies for a semiautomatic process workflow is presented, providing olive tree crown detection in two different olive orchards in Calabria (Italy), collected by means of GEOBIA procedures, in order to investigate olive tree spectral behavior and the reliability of WW-3 derived vegetation indices (VIs). The semi-automated classification method, accomplished by imagery pre-processing steps, may constitute an operational processing chain for mapping and monitoring olive orchards at tree scale detail. Five VIs were investigated: Normalized Difference Vegetation Index (NDVI), Modified Soil Adjusted Vegetation Index 2 (MSAVI 2), Normalized Difference Red Edge Vegetation Index (NDRE), Modified Chlorophyll Absorption Ratio Index Improved (MCARI2), and NDVI2. The obtained results were statistically tested and their accuracy assessed. Thematic accuracy ranges from 95.33% to 96% in both study areas with an overall tree detection rate of 96.8%. Statistical analysis showed that the major differences in spectral behavior, over different plots of the investigated olive orchards, are mainly due to the component of the red-infrared regions of the electromagnetic spectrum (EM), where the red-edge becomes important in assessing the state of general vigor. Moreover, the proposed methodology increases the possibility of detecting tree stress at earlier stages and the benefits of using satellite-based approaches in terms of: larger area coverage, less processing and operator interaction coupled with more spectral information, thus reducing the need to collect costly reference data sampling.
Automatic citrus tree extraction from UAV images and digital surface models using circular Hough transform
2018, Computers and Electronics in Agriculture
Tree counts and sizes are important information to apply to crop yield estimation and agricultural planning. Therefore, obtaining automatic extraction of trees, their locations, diameters, and counts from remotely sensed data is a challenging task. In this study, a novel approach is proposed for the automatic extraction of citrus trees using unmanned aerial vehicle (UAV) multispectral images (MSIs) and digital surface models (DSMs). The tree boundaries were extracted by using sequential thresholding, Canny edge detection and circular Hough transform algorithms. The performance of the developed approach was assessed on three test areas that include different characteristics with regard to tree counts, diameters, densities and background covers. The proposed tree extraction procedure was applied to DSM that were generated from UAV images (Data Set 1), UAV MSIs (Data Set 2) and both of them together (Data Set 3). The accuracies of the obtained results were assessed using three different techniques that evaluate the tree extraction results according to the counts, areas and locations. The obtained results indicate the success of the developed approach with delineation accuracies that exceeded 80% for each test area using each data set. The most accurate results were obtained when Data Set 1 was used. Although Data Set 2 provides the lowest accuracies when compared with other data sets, the delineation accuracies are still high and can be used especially for counting trees and detecting tree locations.
Individual tree crown detection in sub-meter satellite imagery using Marked Point Processes and a geometrical-optical model
2018, Remote Sensing of Environment
This article describes a new algorithm for the detection and delineation of tree crowns using optical sub-meter resolution satellite images. The algorithm focuses on detecting individual semi-isolated trees in a variety of environments defined as trees outside forests (TOF). The concept of Marked Point Processes (MPPs), which alternates phases of “birth” and “death” iterations to satisfy a density factor was used as a theoretical basis. The “mark” in the MPP represents the object being sought. Unlike most applications of MPP to object recognition, the mark used in our algorithm is computed from a 3D geometrical optical model artificially lit using the same illumination parameters as the image itself. Because trees differ in size, the process also incorporates a tree crown radius variable. The algorithm is tested on four sub-meter satellite images, each in a different environment. Validation was performed on both detection and delineation. The detection was based on tree crown counting and yielded an accuracy ranging from 0.81 to 0.95 for the four images. The delineation accuracy was estimated based on the crown pixel count and yielded an accuracy of ≈ 0.63 (0.57–0.72).

View all citing articles on Scopus

View full text

Species classification of individually segmented tree crowns in high-resolution aerial images using radiometric and morphologic image measures

Abstract

Introduction

Section snippets

Material

Methods

Results

Discussion

Conclusions

Acknowledgements

Computer Vision, Graphics, and Image Processing

Computer Vision and Image Understanding

Fuzzy Sets and Systems

Computers & Geosciences

Remote Sensing of Environment

Remote Sensing of Environment

Automatic delineation of individual tree crowns in high spatial resolution aerial images by multiple-scale analysis

Machine Vision and Applications

Segmentation of individual tree crowns in colour aerial photographs using region growing supported by fuzzy rules

Canadian Journal of Forest Research