Accurate inference of shoot biomass from high-throughput images of cereal plants
© Golzarian et al; licensee BioMed Central Ltd. 2011
Received: 11 October 2010
Accepted: 1 February 2011
Published: 1 February 2011
With the establishment of advanced technology facilities for high throughput plant phenotyping, the problem of estimating plant biomass of individual plants from their two dimensional images is becoming increasingly important. The approach predominantly cited in literature is to estimate the biomass of a plant as a linear function of the projected shoot area of plants in the images. However, the estimation error from this model, which is solely a function of projected shoot area, is large, prohibiting accurate estimation of the biomass of plants, particularly for the salt-stressed plants. In this paper, we propose a method based on plant specific weight for improving the accuracy of the linear model and reducing the estimation bias (the difference between actual shoot dry weight and the value of the shoot dry weight estimated with a predictive model). For the proposed method in this study, we modeled the plant shoot dry weight as a function of plant area and plant age. The data used for developing our model and comparing the results with the linear model were collected from a completely randomized block design experiment. A total of 320 plants from two bread wheat varieties were grown in a supported hydroponics system in a greenhouse. The plants were exposed to two levels of hydroponic salt treatments (NaCl at 0 and 100 mM) for 6 weeks. Five harvests were carried out. Each time 64 randomly selected plants were imaged and then harvested to measure the shoot fresh weight and shoot dry weight. The results of statistical analysis showed that with our proposed method, most of the observed variance can be explained, and moreover only a small difference between actual and estimated shoot dry weight was obtained. The low estimation bias indicates that our proposed method can be used to estimate biomass of individual plants regardless of what variety the plant is and what salt treatment has been applied. We validated this model on an independent set of barley data. The technique presented in this paper may extend to other plants and types of stresses.
Plant biomass is an important factor in the study of functional plant biology and growth analysis, and it is the basis for the calculation of net primary production and growth rate [1–4]. Depending on the available budget, accuracy required, structure and composition of the vegetation, and also different disciplines of plant biology, there are several techniques to measure plant biomass . In the study of biomass of an individual plant, shoot dry weight is one of the acceptable measures. This method is typically used to estimate a plant's yield, but it is also an accurate measure of plant biomass.
The conventional means of determining shoot dry weight (SDW) is the measurement of oven-dried samples. In this method, tissue is harvested and dried, and then shoot dry weight is measured at the end of the experiment. To investigate the biomass of a large number of plants, this method is very time consuming and labor intensive. Also, since this method is destructive, it is impossible to take several measurements on the same plant at different time points. Therefore, an imaging method has been proposed to infer plant biomass accurately as a non-destructive and fast alternative. The Plant Accelerator  and the High Resolution Plant Phenotyping Centre  in Australia, the Leibniz Institute of Plant Genetics and Crop Plant Research (IPK) in Germany , the Institute of Biological, Environmental and Rural Sciences (IBERS) in the UK , and PHENOPSIS system being built by the National Institute for Agricultural Research (INRA) in Montpellier, France , have established or are planning to establish advanced plant phenotyping facilities that each provide the capability of hundreds to thousands of plants to be automatically imaged from standard positions and then analyzed via image analysis programs every day.
Digital image analysis has been an important tool in biological research and also has been applied to satellite images, aerial photographs and macroscopic and microscopic images . A relevant application of image analysis which has been used for decades is in the area of remote sensing forestry and precision agriculture in which the area of plant species cover and the biomass of the above-ground canopy are estimated from satellite and airborne images [12–20]. These techniques have found a recent application in estimating the biomass of individual plants in a controlled environment and also in the field. There have been only a few projects on the application of image analysis techniques to estimate above-ground biomass of an individual plant. In these, the projected shoot area of the plants captured on two dimensional images was used as a parameter to predict the plant biomass [1, 15, 21–25]. Except for predicting cereal plant biomass as a linear function of plant area, however, none of the methods described in the literature was developed explicitly for high throughput phenotyping facilities. A robust and accurate method is required for high throughput phenotyping.
An additional factor to consider is the level of salinity to which the plant has been exposed. Arid and semi-arid agricultural lands such as those in Australia inevitably pose some levels of soil salinity, which is one of the major environmental stresses that significantly affects crop productivity. The crop plants are stressed when the high concentrations of salts in the soil make it harder for their roots to extract water [26, 27]. Salinity seems to have some effect on wheat growth in terms of their morphology, physiology and anatomical changes [26, 28–30]. The applied salt treatment on the plants in the simulates the effect of soil salinity on crop plants in an agricultural field. The linear model, the predominant method used to estimate plant biomass, shows biased estimation of plant biomass particularly for salt stressed plants.
The objective of the present study is to develop a generalized method to estimate the biomass of cereals from their projected shoot area on two dimensional images. We have developed a method that significantly reduces the bias in biomass estimation of stressed cereal plants, which is the main source of the estimation error. We have demonstrated that a model that uses mixed variables of plant area and plant age achieves this reduction and therefore the method we proposed can be used to compute accurately the biomass of cereal plants regardless of whether or not they are salt stressed. In order to generalize our method to cereal plants we tested our method on both wheat and barley datasets and achieved promising results.
Plant images were captured using a LemnaTec 3D Scanalyzer (LemnaTec, GmbH, Wuerselen, Germany) at Australian Centre for Plant Functional Genomics (University of Adelaide, Waite Campus, Adelaide). Comparable imaging systems are also used in other phenotyping facilities. Three 1280 × 960 resolution RGB images were taken of every plant: one top view image and two side view images at a 90° horizontal rotation. The images were stored in PNG format. In order to increase the accuracy of separating the background from the region of interest (plant region), a roughly uniform blue background was used and the plant pot was also wrapped in a blue paper tube at the time of imaging. To develop the model and ensure sufficient variation, a total of 320 wheat plants were used for this study. The plants were of two Australian bread wheat varieties, Krichauff and Berkut, grown under two salt treatments, 0 and 100 mM NaCl. The bread wheat (Triticum aestivum L.) cultivars Berkut and Krichauff are quite distinct, and have different pedigrees. Berkut comes from CIMMYT, Mexico and Krichauff is a southern Australian commercial cultivar. They have been selected as diverse parents of a mapping population, which have been identified to have significant variation in salinity tolerance traits, one of which has even been mapped in a large genetic study, published recently .
In terms of salinity, 100 mM salinity is a moderate level of salinity which reduces growth by approximately 10 to 30%. This level of salinity has been estimated to cover as much as 69% of the Australian wheat belt  and is a global problem at this and much higher levels .
Details of Shoot Dry Weight measurements (wheat dataset).
(days after planting)
Control plants (no salt)
Salt stressed plants
Two salt treatments combined
Details of Shoot Dry Weight measurements (barley dataset).
(days after transplant)
Salt stressed plants
Two salt treatments combined
Image processing algorithm
We used the LemnaTec 3D Image Analyser (LemnaTec GmbH, Wuerselen, Germany) to run image processing algorithms to extract information from the plant RGB images. The plant color images were first converted into the "Hue Saturation Intensity (HSI)" color model in order to increase the contrast between plant region and background region. A threshold was applied on the hue image in order to separate plant area from the background. The segmentation process was accomplished by selecting the pixels with values over the threshold belonging to plant region and rejecting all the other pixels to the background region. The resulting image is a binary or two-level image, using white and black to distinguish the plant and background regions, respectively. The number of pixels inside the plant region was counted in each of the three orthogonal views, converted to mm2 using the appropriate calibration factor, and then summed to give the projected shoot area. This is not the actual shoot surface area but the sum of the areas of the image projected in three planes. There are many cases when a mature plant's leaves are overlapping, appearing behind one another in side view images. In these cases, a top view image provides a means of correction of plant area for those overlapping leaves in side view images. The three orthogonal views (two side views from 90 rotational difference) and a top view correct for hidden areas in the other views and give a robust representation of plant area overall.
Cross validation technique
To measure the generalization or estimation error of a predictive model, the technique of cross validation was used. Cross validation, or rotation estimation, is a technique for assessing the prediction error. This technique estimates the generalization error, L(Y,Ŷ), where L is the distance function and Ŷ is the model applied to the independent test sample from the distribution of X and Y. Cross validation is a robust method and preferred over the R2 statistic. The main reason is that R2 inevitably increases with additional predictors, and more predictors automatically yield improved prediction within one dataset. However, the cross validation error decreases only as long as the additional predictor improves the predictive capability of the model in an independent dataset . In the cross validation technique, the observations are randomly assigned indices, integers 1 to K. In this way, the dataset is partitioned into K approximately equal-sized parts. Then, the model is fitted to K-1 parts of the dataset (with one part, say the k th part, removed), and the prediction error of the fitted model is calculated from the k th part. This procedure is repeated for k = 1, 2, .., K rounds, and estimation errors, such as the root mean square errors (RMSE values), are averaged over the rounds [37–39]. Typical choices of K are 5 or 10, and in this study fivefold (i.e, K = 5) cross validation was used. The estimation errors obtained from applying this technique were used to compare the performance of different predictive models. The cross validation analysis was performed in Matlab (Mathworks Inc., Natick, MA).
Results and Discussion
Significance of regression coefficient of different methods used to estimate plant biomass from the plant area.
As can be seen from the Table 3, among polynomial models, only the linear model is significant. The linear model may be compared with the non-linear power model by inspecting their estimation errors achieved using five fold cross validation analysis applied on the wheat dataset.
Estimation error for linear and power models used to estimate plant biomass.
Plant biomass predictive model
SDW = a0 + a1A
SDW = a0 Aa1
where n is the total number of images. The estimation error of the linear method is significantly smaller (P-value < 0.00005) than that from the power method (Table 4).
The linear model seems to be the best of those considered so far, justifying its common use in the literature.
where SDW means Shoot Dry Weight (g), Area means projected shoot area on the image plane (mm2), and the density can be estimated as a linear function of plant age as for PSW above, and a0 and a1 are the equation coefficients. It is not necessary that the coefficient a0 be zero.
where HD is plant age in days after planting.
Significance of regression coefficient of our proposed method.
Data analysis and performance comparison of the models
As the linear model proved to be better than the non-linear models we considered, we compared our proposed model with the linear model described in Table 4.
Prediction errors obtained from cross validation method for the linear model and the proposed model.
Model A: SDW = a0 + a1A
Model B: SDW = c0 + c1A+ c2AH
Mean square estimation error and mean total estimation error (bias) for two models under two salt treatment categories.
Considering two groups of salt treatments, the root mean square estimation error from Model A is about one and one-half times greater than that for Model B for the control plants and salt-stressed plants. Also, the MTE error for Model A is very high - three times higher than that for Model B for the two groups of salt treatments. According to Table 7, on average the linear model over-estimates the weight of a control plant by 35 mg and under-estimates the weight of a stressed plant by 34 mg. Considering Table 1, this estimation error is about 10% of the mean of all shoot dry weight measurements. Meanwhile, the observations indicate that there is very small bias obtained from Model B to estimate plant biomass. When using the proposed model (Model B), this bias is only 11 mg greater and less than the actual weight of the plant for control and salt stressed plant, respectively.
Validating the model using barley dataset
Prediction errors obtained from cross validation method on barley dataset for the linear and proposed models.
Model A: SDW = a0 + a1A
a0 = -0.132; a1= 0.003;
Model B: SDW = c0 + c1A+ c2AH
c0 = -0.025; c1= 5.48E-4; c2= 5.60E-5
Mean square estimation error and mean total estimation error (bias) for two models under two salt treatment categories for barley dataset.
Bias (MTE) (g)
Bias (MTE) (g)
In this study we have presented a method for accurate estimation of plant shoot dry weight from two dimensional images. Our proposed model employs information obtained from the images of plants and their age. This approach provides an accurate and practical model for the estimation of wheat and barley shoot dry weight as a substitute for conventional destructive methods of biomass measurement. We also demonstrated that, for salt stressed plants, the estimation bias between the actual and predicted shoot dry weight values can be overcome to a large extent by using plant biomass estimators with plant age as an additional input. Without this method, we cannot accurately infer the plant biomass for salt stressed plants. We tested our proposed model on wheat and barley from different contrasted varieties and under salt stress and found out that with our method the error in biomass estimation was reduced significantly. Thus, our method enables high throughput non-destructive estimation of biomass for cereal plants under salt stress and may possibly do so for other types of plants and stresses.
The authors are thankful to Mr. James Eddes for creating the image analysis routines in LemnaTec software. This research was supported by the Government of South Australia through the Premier's Science and Research Fund. We also appreciate Dr Juergen Zanghellini for his valuable comments on the final version of the manuscript.
- Tackenberg O: A new method for non-destructive measurement of biomass, growth rates, vertical biomass distribution and dry matter content based on digital image analysis. Annals of Botany. 2007, 99: 777-783. 10.1093/aob/mcm009.PubMed CentralView ArticlePubMedGoogle Scholar
- Niklas KJ, Enquist BJ: On the Vegetative Biomass Partitioning of Seed Plant Leaves, Stems, and Roots. The American Naturalist. 2002, 159: 482-497. 10.1086/339459.View ArticlePubMedGoogle Scholar
- Poorter H, Nagel O: The role of biomass allocation in the growth response of plants to different levels of light, CO2, nutrients and water: a quantitative review. Functional Plant Biology. 2000, 27: 595-607. 10.1071/PP99173.View ArticleGoogle Scholar
- Wilson PJ, Thompson K, Hodgson JG: Specific Leaf Area and Leaf Dry Matter Content as Alternative Predictors of Plant Strategies. New Phytologist. 1999, 143: 155-162. 10.1046/j.1469-8137.1999.00427.x.View ArticleGoogle Scholar
- Wheeler WRCCJ: Estimating plant biomass: A review of techniques. Austral Ecology. 1992, 17: 121-131. 10.1111/j.1442-9993.1992.tb00790.x.View ArticleGoogle Scholar
- Plant Accelerator. [http://www.plantaccelerator.org.au/]
- High Resolution Plan Phenotyping Centre. [http://www.plantphenomics.org.au/HRPPC]
- The Leibniz Institute of Plant Genetics and Crop Plant Research (IPK). [http://www.ipk-gatersleben.de]
- Institute of Biological, Environmental and Rural Sciences (IBERS). [http://www.aber.ac.uk/en/ibers/]
- French National Institute for Agricultural Research (INRA). [http://www.international.inra.fr/]
- Nilsson H: Remote Sensing and Image Analysis in Plant Pathology. Annual Review of Phytopathology. 1995, 33: 489-528. 10.1146/annurev.py.33.090195.002421.View ArticlePubMedGoogle Scholar
- Montès N, Gauquelin T, Badri W, Bertaudière V, Zaoui EH: A non-destructive method for estimating above-ground forest biomass in threatened woodlands. Forest Ecology and Management. 2000, 130: 37-46.View ArticleGoogle Scholar
- Lim KS, Treitz PM: Estimation of above ground forest biomass from airborne discrete return laser scanner data using canopy-based quantile estimators. Scandinavian Journal of Forest Research. 2004, 19: 558-570. 10.1080/02827580410019490.View ArticleGoogle Scholar
- Dietz H, Steinlein T: Determination of plant species cover by means of image analysis. Journal of Vegetation Science. 1996, 7: 131-136. 10.2307/3236426.View ArticleGoogle Scholar
- Sher-Kaul S, Oertli B, Castella E, Lachavanne JB: Relationship between biomass and surface area of six submerged aquatic plant species. Aquatic Botany. 1995, 51: 147-154. 10.1016/0304-3770(95)00460-H.View ArticleGoogle Scholar
- Rutchey K, Schall T, Sklar F: Development of Vegetation Maps for Assessing Everglades Restoration Progress. Wetlands. 2009, 28: 806-816. 10.1672/07-212.1.View ArticleGoogle Scholar
- Rutchey K, Vilchek L: Air Photointerpretation and Satellite Imagery Analysis Techniques for Mapping Cattail Coverage in a Northern Everglades Impoundment. Photogrammetric Engineering and Remote Sensing. 1999, 65: 185-191.Google Scholar
- Lu D, Mausel P, Brondizio E, Moran E: Aboveground biomass estimation of successional and mature forests using TM images in the Amazon Basin. Advances in Spatial Data Handling. Edited by: Richardson D, Oosterom Pv. 2002, New York: Springer-Verlag, 183-196.View ArticleGoogle Scholar
- Seelan SK, Laguette S, Casady GM, Seielstad GA: Remote sensing applications for precision agriculture: A learning community approach. Remote Sensing of Environment. 2003, 88: 157-169. 10.1016/j.rse.2003.04.007.View ArticleGoogle Scholar
- Lamb DW, Brown RB: Precision Agriculture: Remote-Sensing and Mapping of Weeds in Crops. Journal of Agricultural Engineering Research. 2001, 78: 117-125. 10.1006/jaer.2000.0630.View ArticleGoogle Scholar
- Paruelo JM, Lauenroth WK, Roset PA: Estimating Aboveground Plant Biomass Using a Photographic Technique. Journal of Range Management. 2000, 53: 190-193. 10.2307/4003281.View ArticleGoogle Scholar
- Mizoue N, Masutani T: Image analysis measure of crown condition, foliage biomass and stem growth relationships of Chamaecyparis obtusa. Forest Ecology and Management. 2003, 172: 79-88. 10.1016/S0378-1127(02)00281-5.View ArticleGoogle Scholar
- Lukina EV, Stone ML, Raun WR: Estimating vegetation coverage in wheat using digital images. Journal of Plant Nutrition. 1999, 22: 341-350. 10.1080/01904169909365631.View ArticleGoogle Scholar
- Smith SM, Garrett PB, Leeds JA, McCormick PV: Evaluation of digital photography for estimating live and dead aboveground biomass in monospecific macrophyte stands. Aquatic Botany. 2000, 67: 69-77. 10.1016/S0304-3770(99)00085-6.View ArticleGoogle Scholar
- Smith MAL, Spomer LA, Meyer MJ, McClelland MT: Non-invasive image analysis evaluation of growth during plant micropropagation. Plant Cell, Tissue and Organ Culture. 1989, 19: 91-102. 10.1007/BF00035809.View ArticleGoogle Scholar
- Munns R, Tester M: Mechanisms of salinity tolerance. Annual Review of Plant Biology. 2008, 59: 651-681. 10.1146/annurev.arplant.59.032607.092911.View ArticlePubMedGoogle Scholar
- Skirycz A, Inzé D: More from less: plant growth under limited water. Current Opinion in Biotechnology. 2010, 21: 197-203. 10.1016/j.copbio.2010.03.002.View ArticlePubMedGoogle Scholar
- Maas EV, Poss JA: Salt sensitivity of wheat at various growth stages. Irrigation Science. 1989, 10: 29-40.Google Scholar
- Madhava Rao KV, Raghavendra AS, Janardhan Reddy K, Springer-Verlag: Physiology and molecular biology of stress tolerance in plants. 2006, Dordrecht: SpringerView ArticleGoogle Scholar
- Fricke W, Akhiyarova G, Wei W, Alexandersson E, Miller A, Kjellbom PO, Richardson A, Wojciechowski T, Schreiber L, Veselov D: The short-term growth response to salt of the developing barley leaf. J Exp Bot. 2006, 57: 1079-1095. 10.1093/jxb/erj095.View ArticlePubMedGoogle Scholar
- Genc Y, Oldach K, Verbyla A, Lott G, Hassan M, Tester M, Wallwork H, McDonald G: Sodium exclusion QTL associated with improved seedling growth in bread wheat under salinity stress. TAG Theoretical and Applied Genetics. 2010, 121: 877-894. 10.1007/s00122-010-1357-y.View ArticlePubMedGoogle Scholar
- Rengasamy P: Transient salinity and subsoil constraints to dryland farming in Australian sodic soils: an overview. Australian Journal of Experimental Agriculture. 2002, 42: 351-361. 10.1071/EA01111.View ArticleGoogle Scholar
- Rajendran K, Tester M, Roy SJ: Quantifying the three main components of salinity tolerance in cereals. Plant, Cell & Environment. 2009, 32: 237-249.View ArticleGoogle Scholar
- Garnier E, Shipley B, Roumet C, Laurent G: A Standardized Protocol for the Determination of Specific Leaf Area and Leaf Dry Matter Content. Functional Ecology. 2001, 15: 688-695. 10.1046/j.0269-8463.2001.00563.x.View ArticleGoogle Scholar
- Cornelissen JHC, Lavorel S, Garnier E, Diaz S, Buchmann N, Gurvich DE, Reich PB, ter Steege H, Morgan HD, van der Heijden MGA: A handbook of protocols for standardised and easy measurement of plant functional traits worldwide. Australian Journal of Botany. 2003, 51: 335-380. 10.1071/BT02124.View ArticleGoogle Scholar
- Genc Y, McDonald GK, Tester M: Reassessment of tissue Na+ concentration as a criterion for salinity tolerance in bread wheat. Plant, Cell & Environment. 2007, 30: 1486-1498.View ArticleGoogle Scholar
- Hastie T, Tibshirani R, Friedman JH: The elements of statistical learning: data mining, inference, and prediction. 2001, New York: SpringerView ArticleGoogle Scholar
- Witten IH, Frank E: Data mining: practical machine learning tools and techniques. 2005, San Francisco, Calif.: Morgan Kaufman, 2Google Scholar
- Hastie T, Tibshirani R, Friedman JH: The elements of statistical learning: data mining, inference, and prediction. 2009, New York: Springer, 2View ArticleGoogle Scholar
- Leister D, Varotto C, Pesaresi P, Niwergall A, Salamini F: Large-scale evaluation of plant growth in Arabidopsis thaliana by non-invasive image analysis. Plant Physiology and Biochemistry. 1999, 37: 671-678. 10.1016/S0981-9428(00)80097-2.View ArticleGoogle Scholar
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.