Estimating leaf area index in tree species using the PocketLAI smart app
Abstract
Aim
To evaluate the PocketLAI® smart app for estimating leaf area index (LAI) in woody canopies.
Location
Northern Italy.
Methods
PocketLAI – a smartphone application for LAI estimates based on gap fraction derived from the real-time processing of images acquired at 57° below the canopy – was tested on continuous forest stands, plantations, spotted shrub-lands and spotted tree-lands. LAI data from hemispherical photography (images post-processed with Can-eye software) were taken as reference values. Plants were clustered on the basis of leaf type and canopy structure.
Results
In general, PocketLAI showed satisfactory performances in the case of broad-leaf plants (R2 = 0.78, P < 0.001) for all shrub and tree clusters. On the other hand, poor results were obtained for conifers (R2 = 0.16), likely because of the unfavourable leaf area to perimeter ratio. Best performances were observed for dense broad-leaf canopies characterized by a regular arrangement of crowns (R2 = 0.95 for row-planted trees, R2 = 0.87 for tall forest trees), although satisfying results were achieved also in the case of canopies made irregular and non-homogeneous by pruning (R2 = 0.73 for small fruit trees). Concerning shrubs, the agreement between PocketLAI and hemispherical photography was higher for species with big leaves (R2 = 0.72).
Conclusions
These results suggest that PocketLAI can be an alternative to other methods in case of broad-leaf woody species, especially in contexts where resources and portability are key issues, whereas further improvements are required for conifers.
Nomenclature
-
- USDA Plants Database (http://plants.usda.gov/java/)
-
-
- Euro+Med PlantBase (http://www.emplantbase.org/home.html)
-
Introduction
The availability of reliable leaf area index (LAI: total one-sided area of leaf tissue per unit ground surface; m2 leaf area·m−2 soil) data is crucial for a variety of agro-environmental studies, since this variable represents the main interface between plants and atmosphere (e.g. Whitford et al. 1995). However, direct LAI determination (planimetric method) is time-consuming and, although sometimes performed for herbaceous species, it is practically unfeasible in most of the contexts where tree canopies are involved (Jonckheere et al. 2004).
For forest stands, the need for LAI information on large areas is increasingly leading to estimated LAI using overhead remote sensing techniques (Tang et al. 2014). However, despite the great potential of this technique, images partly lose their usefulness in case of spotted vegetation, mainly due to the presence of mixed pixels and to noise in the spectral response due to the soil background (Huete & Tucker 1991). This is why indirect methods for LAI estimates (like those implemented in LAI-2000) are often used in sparse vegetation or for row-structure tree canopies (Khabba et al. 2009). Among indirect methods, hemispherical photography through fish-eye lens is one of the most popular techniques for woody plants. It is based on permanent image recording, thus providing unlimited possibilities of ex-post image treatment and reprocessing (Jonckheere et al. 2004), as well as the possibility of analysing canopy architecture and identifying errors during acquisition. Moreover, the methodology does not require dedicated and costly instruments.
The effectiveness and reliability of fish-eye images for LAI estimates have been demonstrated in a variety of studies performed on different types of woody canopy, ranging from apple and orange orchards (Khabba et al. 2009; Liu et al. 2013) to conifer and deciduous forests (e.g. Macfarlane et al. 2000, 2007). Despite the availability of dedicated software tools (e.g. Can-eye, Weiss et al. 2004; Winscanopy, Rich et al. 1993), the main problem in using hemispherical photography for LAI estimate is the time required for image post-processing and the subjectivity of the user in setting the thresholds used by the segmentation algorithm (Jonckheere et al. 2004).
Confalonieri et al. (2013) proposed an application for estimating LAI using a smartphone. The app (PocketLAI) accuracy was evaluated in a comparative study with other commercial instruments for paddy rice using an adaptation of the ISO protocol for method validation (Confalonieri et al. 2014). Despite the low cost, PocketLAI performances were similar to those achieved by LAI-2000 and AccuPAR ceptometer. Despite the relative simplicity of the approach, it was shown to be effective also for canopies markedly deviating from the assumption (random distribution of infinitely small leaves) behind the simplified transmittance model implemented (Francone et al. 2014).
The aims of this study were to (1) assess PocketLAI reliability for tree canopies and (2) develop protocols for PocketLAI application to different canopy types: continuous forest stands, plantations, spotted shrub-lands and spotted tree-lands. LAI data obtained from hemispherical images – acquired with a focus-free camera with a CMOS image sensor and processed with the CAN-EYE software (Weiss et al. 2004) – were taken as reference values.
Methods
The PocketLAI mobile app
PocketLAI uses the device accelerometer and camera to automatically take images from below the canopy at a view angle of 57.5° while the user is rotating the device along its main axes (Fig. 1a). This is made possible by a digital inclinometer implemented in the app, which provides real-time view angle information based on the data provided by the accelerometer. This particular angle allows the light transmittance model adopted to be independent of leaf angle distribution (Warren-Wilson 1963; Baret et al. 2010). The device camera is used in live-preview mode to allow taking of images at 57.5° from a ~25 frames s−1 stream. Invoking a full-frame image capturing event would instead lead to a time-lag between the moment when the inclinometer detects the target angle and the moment when the image is actually acquired. Since information is acquired while the user is rotating the device, the time lag would be translated in a shift from the target angle and, in turn, in errors in LAI estimates (Baret et al. 2010). Images are then processed using an automatic segmentation algorithm to derive the gap fraction and, in turn, LAI by inverting the Warren-Wilson (1963) light transmittance model.
Leaf area index estimates
The LAI estimates were collected on woody canopies differing in height, shape, fractal dimension of the crown, age and growth stage (Table 1). Measurements were taken in continuous canopy biomes, row-structured plantations and sparse tree- and shrub-lands in Northern Italy (i.e. Piedmont and Lombardy Regions) during 2014.
| Species | Number of Observations (cfr. Fig. 1) | Use |
|---|---|---|
| Acer platanoides | 3 | Forestry |
| Aesculus hippocastanum L. | 4 | Forestry |
| Betula pendula, Roth. | 1 | Forestry |
| Carpinus betulus, L. | 2 | Ornamental |
| Catalpa bignonioides, Walter | 2 | Ornamental |
| Cedrus libani, A.Rich. | 5 | Forestry |
| Cercis siliquastrum, L. | 2 | Ornamental |
| Chaenomeles japonica, Lindl. | 5 | Ornamental |
| Cornus mas, L. | 3 | Ornamental |
| Corylus avellana, L. | 36 | Fruiting |
| Cryptomeria japonica, (L. f.) D. Don | 1 | Forestry |
| Diospyros kaki, L. f. | 2 | Fruiting |
| Fagus sylvatica, L. | 4 | Forestry |
| Ficus carica, L. | 2 | Fruiting |
| Forsythia viridissima, Lindley | 3 | Ornamental |
| Ilex aquifolium, L. | 4 | Ornamental |
| Liriodendron tulipifera, L. | 1 | Ornamental |
| Magnolia grandiflora, L. | 3 | Ornamental |
| Malus domestica, Borkh. | 2 | Fruiting |
| Olea europaea, L. | 13 | Fruiting |
| Picea pungens | 2 | Ornamental |
| Pinus, spp. | 3 | Forestry |
| Populus spp. | 12 | Forestry |
| Prunus armeniaca, L. | 2 | Fruiting |
| Prunus avium, L. | 4 | Fruiting |
| Prunus laurocerasus, L. | 3 | Ornamental |
| Prunus persica, L. | 4 | Fruiting |
| Pyracantha coccinea, M. Roem. | 1 | Ornamental |
| Pyrus communis | 2 | Fruiting |
| Quercus robur, L. | 1 | Forestry |
| Robinia pseudoacacia, L. | 2 | Forestry |
| Salix nigra | 1 | Forestry |
| Sequoia sempervirens, (D. Don) Endlicher | 1 | Forestry |
| Sorbus domestica, L. | 4 | Forestry |
| Taxus baccata, L. | 6 | Ornamental |
| Ulmus × hollandica, Mill. | 1 | Forestry |
For the acquisition of LAI data with PocketLAI, three protocols were used. For isolated plants, data were collected – rotating the smartphone toward the trunk – from eight points below the canopy, moving around the trunk and acquiring measures approximately each 45° (Fig. 1b). For continuous canopies, eight data points were acquired turning around a point and taking measures approximately each 45° (Fig. 1c) but with the device oriented outwards. For plants arranged in rows, the eight measures were acquired while moving in parallel to the row (Fig. 1d). For isolated and row-arranged plants, the device was positioned below the canopy at a distance from the trunk of approximately half crown radius. Mean LAI values were calculated after discarding outliers from the eight replicates using the Grubbs’ test (Grubbs 1950).
For this study, PocketLAI – currently available for Android devices at [email protected] – was installed on a Samsung GT-i9105 Galaxy S II Plus.
The LAI data from hemispherical photography were used as reference values for the PocketLAI evaluation. Fish eye images were collected using a protocol similar to that used for PocketLAI (Fig. 1b, c, d), although ten replicates were taken instead of eight. Hemispherical photographs were acquired below the canopy with the camera oriented upward, turning around the trunk (Fig. 1b) and around a point (Fig. 1c) for isolated plants and continuous canopies, respectively. For plants arranged in rows, the ten replicates were taken as shown in Fig. 1d.
The Can-Eye software (v 6.314; www.avignon.inra.fr/can_eye; Weiss et al. 2004) was used for processing fish-eye images, since it was shown to be the best package in a comparative study performed by Liu et al. (2013). Before the actual image processing, the following steps were performed: (1) eight photographs (Weiss et al. 2004) were selected among the ten acquired for each set of replicates by manually discarding images with non-uniform lighting conditions (Garrigues et al. 2008); (2) undesired objects (e.g. operator, buildings) were removed using the dedicated masking function; (3) on the basis of the k-means clustering method, the total number of distinctive colours was reduced to 324 (Spath 1985).
Evaluation of PocketLAI for different types of tree canopy
Tree and shrub canopies (Table 1) were clustered by identifying a number of classification keys for leaf type and canopy structure and binary attributes for each key (Table 2). The main distinction was between broad-leaf and conifer plants, whereas canopy architecture (i.e. scaffold branch height and plant arrangement) and leaf type (i.e. shape and size) were secondary attributes used to further classify trees and shrubs. Classification keys and attributes were defined in light of their role in radiation interception and transmittance, and on a priori knowledge of the PocketLAI functioning for woody canopies. As an example, leaf type is among the binary attributes for shrubs but not for tree since pre-tests showed the risk – especially with dense, short shrubs – of acquiring measures affected by a single leaf and not by the canopy (leaf too close to the device lens). This possibility could not be verified for trees, thus the attribute was used only for shrubs.
| Classification Key | Attribute | |
|---|---|---|
| Tree | Division | Broad-leaf |
| Conifer | ||
| Canopy architecture | Low scaffold branches (height from ground <0.5 m) | |
| Medium scaffold branches (height from ground between 0.5 and 1.0 m) | ||
| High scaffold branches (height from ground >2 m) | ||
| Arrangement random | ||
| Arrangement in row | ||
| Shrub | Division | Broad-leaf |
| Conifer | ||
| Leaf type | Big wide leaf (>25 cm2) | |
| Small wide leaf (<25 cm2) | ||
| Slim or needle-like |
Polythetic divisive hierarchical clustering (Cormark 1971) was then carried out separately for trees and shrubs, using the Jaccard's binary similarity measure coefficient to assess the similarity between objects on the basis of their attributes (Cheetham & Hazel 1969; Sneath & Sokal 1973).
The agreement between LAI values estimated using hemispherical photography and PocketLAI was quantified using the following metrics: relative root mean square error (RRMSE; 0 to +∞, optimum 0; Jørgensen et al. 1986), mean absolute error (MAE; 0 to +∞, optimum +∞; Schaeffer 1980), modelling efficiency (EF; −∞ to +1, optimum +1; Nash & Sutcliffe 1970), coefficient of residual mass (CRM; from −∞ to +∞, optimum 0; if positive means underestimation and vice versa; Loague & Green 1991). Parameters of the linear regression equation between values estimated with hemispherical photography and PocketLAI were also calculated.
The comparison was performed for each class of canopy type, as well as for the whole data set.
Results
The clustering procedure led to identification of five classes of tree and three classes of shrub (Table 3). Results of the comparison between LAI values estimated with PocketLAI and hemispherical photography are shown in Table 4 and Fig. 2.
| Group | Description | Number of Data | |
|---|---|---|---|
| Tree | 1 | Broad-leaf tree in sparse canopy, with medium scaffold branches | 49 |
| 2 | Conifer tree in sparse canopy, with medium scaffold branches and slim or needle-like leaves | 7 | |
| 3 | Broad-leaf tree in plantation row, with low scaffold branches | 10 | |
| 4 | Broad-leaf tree in sparse or continuous canopy, with high scaffold branches | 14 | |
| 5 | Conifer tree in sparse or continuous canopy, with high scaffold branches and slim or needle-like leaves | 8 | |
| Shrub | 6 | Broad-leaf shrub in sparse canopy, with small wide leaves | 13 |
| 7 | Conifer shrub in sparse canopy, with slim or needle-like leaves | 4 | |
| 8 | Broad-leaf shrub in sparse canopy, with big wide leaves | 42 |
| Group | MAE (m2·m−2) | RRMSE (%) | EF | CRM | R 2 | |
|---|---|---|---|---|---|---|
| Trees | 1 | 0.66 | 25.87 | 0.73 | 0.00 | 0.73*** |
| 2 | 0.45 | 21.56 | 0.41 | 0.09 | 0.61* | |
| 3 | 0.27 | 14.06 | 0.89 | −0.01 | 0.95*** | |
| 4 | 0.41 | 18.20 | 0.86 | 0.06 | 0.87*** | |
| 5 | 1.72 | 48.10 | −22.77 | 0.48 | 0.56* | |
| Shrubs | 6 | 1.64 | 37.00 | −0.83 | 0.33 | 0.61*** |
| 7 | 2.25 | 51.08 | −6.04 | 0.45 | 0.07 | |
| 8 | 0.70 | 14.94 | 0.49 | 0.09 | 0.72*** | |
| Broad-Leaf | 0.71 | 23.06 | 0.74 | 0.09 | 0.78*** | |
| Conifers | 1.36 | 47.31 | −1.18 | 0.36 | 0.16 | |
| All Data | 0.80 | 26.61 | 0.64 | 0.12 | 0.72*** |
- *P < 0.05; **P < 0.01; ***P < 0.001.
In general, PocketLAI showed satisfactory performances for broad-leaf canopies, with LAI values very close to those estimated using hemispherical photography (R2 = 0.78; Table 4), with highly significant correlation (P < 0.001) for all the groups of broad-leaf shrubs (groups 6 and 8; Table 4) and trees (groups 1, 3 and 4; Table 4).
The best agreement between PocketLAI and hemispherical photography was obtained for row-planted broad-leaf trees with low scaffold branches (Fig. 2c; group 3 in Table 3), for which the lowest RRMSE (14.06%) and the highest R2 (0.95) were achieved, without systematic under- or over-estimation (CRM = −0.01; Table 4). For this group, absolute difference between LAI values estimated with PocketLAI and hemispherical photography exceeded 0.4 m2·m−2 only in three cases, with a MAE of 0.27 m2·m−2. Very good values for all the agreement metrics were also achieved for broad-leaf trees with high scaffold branches in sparse or continuous canopies (Fig. 2d; group 4). For this group, MAE was 0.41 m2·m−2 and absolute differences between LAI values achieved with the two methods were lower than 0.35 m2·m−2 for 65% of the observations. Metric values were slightly less satisfactory (Table 4) for sparse broad-leaf trees with medium scaffold branches (group 1). For this canopy type, indeed, the overall accuracy was lower than for the other two groups of broad-leaf canopies (RRMSE = 25.87%, EF = 0.73), and an average 16% underestimation of LAI values in the range 3.5–5.0 m2·m−2 was observed (Fig. 2a).
Concerning broad-leaf shrubs, PocketLAI performances were fully satisfactory for species with leaf area >~25 cm2 (Fig. 2h; group 8). Metric values were in this case always good (RRMSE = 14.94%, R2 = 0.72, EF largely positive), although a small tendency of PocketLAI to underestimate LAI values was observed. However, MAE was in this case 0.70 m2·m−2 and absolute difference between LAI values estimated with the two methods exceeded 1 m2·m−2 in 22% of the cases. Poor results were instead achieved for broad-leaf shrubs with small leaves (Fig. 2f), for which negative EF and MAE > 1.6 m2·m−2 were achieved. For this group, indeed, the absolute difference between LAI values estimated with PocketLAI and hemispherical photography was <1 m2·m−2 only in 31% of cases.
Results were instead generally unsatisfying for conifers species (MAE = 1.36 m2·m−2; Table 4), with the poorest results obtained for shrubs (Fig. 2g; group 7). The only exception was for sparse conifer trees with medium scaffold branches and slim or needle-like leaves (Fig. 2b; group 2) for which – despite the app often underestimated (CRM = 0.09) LAI values in the explored range – an overall good agreement was achieved (RRMSE = 1.56; R2 = 0.61, P < 0.05; EF largely positive). For this group, absolute difference between LAI values estimated with the two methods exceeded 0.4 m2·m−2 only in three cases.
Discussion
The agreement between estimates provided by PocketLAI and hemispherical photography is comparable with the agreement obtained by other authors while comparing other indirect methods. Hyer & Goetz (2004) estimated correlation coefficients (r) between LAI-2000 and AccuPAR ceptometer of 0.85 and 0.56 for poplar and spruce stands, respectively, suggesting that a lower level of sophistication in the method can affect estimates, especially with conifers. However, in that study, the ceptometer consistently underestimated LAI values measured by LAI-2000. Gower & Norman (1991) highlighted that also LAI-2000 underestimated direct LAI values from dimension analysis, and that the underestimation was more relevant for conifers.
Our results suggest that – in case of woody plants – the best performance can be achieved by PocketLAI for dense canopies, made homogenous by the regular arrangement of the crown either because of species characteristics or because forced by pruning activities and row structure. These considerations explain the better agreement between estimates provided by the app and by hemispherical photography for group 4 (r2 = 0.87; Fig. 2d) and group 3 (r2 = 0.95; Fig. 2c) compared to group 1 (r2 = 0.73; Fig. 2a). In group 3, plants of Robinia pseudoacacia L. and Populus spp. of different ages were grown in row-structured plantation; group 4 included tall trees, characterized by wide crowns and homogenous canopy density, and for the most part forest species, like Fagus sylvatica L., Sorbus domestica L., Populus spp., Salix spp., Aesculus hippocastanum L. Group 1, instead, included mainly small fruit trees, like Ficus carica L., Malus domestica Borkh, Prunus avium L., Olea europaea L., Prunus persica L. These entities were characterized by irregular and uneven crown shapes and non-homogeneous canopy densities due to pruning, which reduces the crown wideness and changes the natural distribution of foliage for specific production purposes. The crown size explains a big part of the lower performances of PocketLAI for group 1 (RRMSE = 25.87%, EF = 0.73) compared to group 3 (RRMSE = 14.06; EF = 0.89) and 4 (RRMSE = 18.20%, EF = 0.86). For group 1, indeed, the analysis of the frames acquired by the app at 57.5° revealed that in some cases the small size of fruit trees led the app to acquire images where the crown was surrounded by portions of sky. This, in turn, led the automatic segmentation algorithm to detect the sky pixel around the crown as part of the gap fraction, being represented by sky pixels, whereas most of the software for processing fish-eye images implements options for masking what is around the crown. Of course, this is one of the ‘cons’ of the high level of automation implemented in the app, since the gap fraction should be estimated in canopy pictures only as a proxy of the radiation transmitted through the canopy. This explains most of the measurements where PocketLAI underestimated LAI values above 3.5 m2·m−2 for canopies belonging to group 1.
In general, PocketLAI performances were unsatisfactory for conifers, with the only exception represented by group 2 (sparse conifer trees with medium scaffold branches). Problems with conifers have also been highlighted by other authors who used other indirect methods (e.g. Gower & Norman 1991; Hyer & Goetz 2004). The analysis of the frames acquired by the app led us to realize that the bad functioning with conifers is explained by the ratio of leaf perimeter to leaf area. PocketLAI acquires frames with the device camera in live preview mode to avoid deviations from the target angle that would affect estimates in the case of acquiring full-frame images. However, the resolution of live preview frames is low and – although enough for broad-leaf species – it proved partly unsuitable when needle-like leaves are analysed by segmenting pictures directly invested by light beams. This is because the low resolution leads to degradation, especially of the sharpness of borders, and this effect is particularly significant in the case of conifers, for which the ratio of leaf border to leaf area is exceptionally high. The resulting underestimation in LAI values – quantified by processing live preview and full-frame images using the same segmentation algorithm – reached 22%, thus demonstrating the partial unsuitability of the current version of PocketLAI for conifers.
The analysis of gap fraction data from images segmented with PocketLAI and Can-Eye demonstrated the higher reliability of the multi-angle approach allowed with the hemispherical photography. Indeed, LAI values underestimated by the app where related to gap fraction overestimations, and the same was observed for gap fraction values estimated by Can-Eye at 57° zenith angle.
Conclusions
In general, the agreement between LAI estimates provided by PocketLAI and hemispherical photography can be considered as satisfactory for broad-leaf species (both trees and shrubs). In contrast, problems often emerged for conifers, even if the analysis of frames allows us to conclude that, in the case of diffuse radiation, the unsuitability due to the leaf border to area ratio is reduced.
Like any other method for LAI estimation, PocketLAI has pros and cons. The two main cons that emerged from this study are due to the high level of automation during collection of measurements. The first con is related to the absence of masking utilities – available in other methods/instruments, like hemispherical photography or in the CI-100 plant canopy analyser (CID Bio-Science) – that led to overestimation of the gap fraction in the case of tall and isolated plants. In this case, indeed, portions of sky surrounding the tree canopy were sometimes included in the picture and, in turn, included in the gap fraction by the automatic segmentation algorithm. This source of uncertainty can potentially affect LAI estimates, especially in the case of sparse canopies with high scaffold branches, since for other canopy types the user has the possibility to minimize the effect by adjusting the distance between the device and the canopy.
The second con we observed is related to the level of coherence needed between the target angle and the angle at which the image is actually acquired. The requirement for developing a friendly tool led us to use the device accelerometer to automatically acquire the image at 57° while the user is rotating the device. This led to acquiring the frame while the camera is in live preview mode to avoid the time lag that occurs in the case of acquiring a full-frame image. Since the user is rotating the device while these events (angle reading and image acquisition) take place, the time lag will be translated in a deviation from the target angle, and Baret et al. (2010) demonstrated that even small deviations from the target angle can lead to relevant errors in LAI estimates with the light transmittance model used (Warren-Wilson 1963).
However, given (1) the objectives of a method for LAI estimates implemented in a smart app and (2) the research and operational activities the app was designed for, we consider the cons as fully balanced by the PocketLAI's usability, portability and the immediacy in data collection. As an example, hemispherical photography is surely more accurate and was indeed used as a reference in this study. However, the processing of an image from hemispherical photography required between 2 and 20 min according to the complexity of the image and the user's experience. This means that, according to specific research or operational contexts, the two methods can be alternatively selected: hemispherical photography will guarantee the highest accuracy, whereas PocketLAI will allow one to easily collect a huge amount of data with an accuracy level that, in case of broad-leaf canopies, can be considered as satisfactory.
Acknowledgements
We gratefully acknowledge Pier Mario Chiarabaglio and Gaetano Castro for the pleasant time spent together at CRA-PLF measuring LAI of their wonderful poplar trees. We also acknowledge Lorenzo Busetto and Francesco Nutini (CNR-IREA) for the in-depth training on how to work with Can-eye.
