Académique Documents
Professionnel Documents
Culture Documents
(Received 15 August 2002; accepted in revised form 15 May 2003; published online 21 June 2003)
The objectives of precision agriculture are profit maximisation, agricultural input rationalisation and
environmental damage reduction, by adjusting the agricultural practices to the site demands. The fall
armyworm (Spodoptera frugiperda) is one of the most important maize pests in Brazil and the use of insecticide
is the main control method. It is believed that site-specific control can be implemented by using a machine
vision system. The objective of this work was to develop and evaluate an algorithm at simplified lighting
conditions for identifying damaged maize plants by the fall armyworm using digital colour images. Images of
damaged and non-damaged maize plants were taken in eight different stages and in three different light
intensities. The proposed algorithm had two stages: the processing and the image analysis. During the first
stage, the images were processed to create binary images where the leaves were segmented from the other
pixels. At the second stage, the images were subdivided into blocks and classified as ‘damaged’ or ‘non-
damaged’ depending on the number of objects found in each block. The algorithm correctly classified 9472%
of 720 images.
# 2003 Silsoe Research Institute. All rights reserved
Published by Elsevier Science Ltd
Object identification
Binary filtered image and counting
(a) (b)
(c) (d)
Fig. 3. (a) Original image (changed to its greyscale representation for typographical reasons); (b) image processed with the
normalised excess green index; (c) binary image created with the iterative method; and (d) the selected blocks
Table 2
Classification errors using 8, 9, 10 and 11 objects as a threshold to classify the plants as damaged or non-damaged from a total
sample of 720 images
Damage threshold Misclassification of damaged plants Misclassification of undamaged plants Total plant misclassification
No. % No. % No. %
8 7 194 73 2028 80 1111
9 9 250 46 1278 55 764
10 12 333 26 722 38 528
11 37 1028 17 472 54 750
classification errors occurred for the non-damaged plants number of plants. The algorithm classified damaged
with the values of 8, 9 and 10 objects. Applying chemicals plants as non-damaged only at the first two growth
to a healthy plant is a less critical error than an application stages, when the damagesites were few and small. The
being missed. The lower is the threshold value, the lower is greatest errors occurred in the first stage (17 days after
the classification error in the damaged plants and the seedling emergence and 2 days after infestation) with
greater the classification error in the non-damaged plants. 22% error. On the other hand, on the group of non-
The best algorithm for overall performance was presented damaged plants, the greatest errors occurred in the last
when using ten objects as threshold value. stages, when the leaf veins were developed and presented
Table 3 shows the number of plants incorrectly a bright colour. For this group, the greatest error
classified using ten objects as the threshold and the per occurred at 24 days after seedling emergence with an
cent of error for damaged, non-damaged and total error of 389%.
ARTICLE IN PRESS
MAIZE PLANT IDENTIFICATION 453
Table 3
Number of plants incorrectly classified in a total sample of 720 images, per cent of error for damaged, not damaged and total number
of plants using ten objects as threshold, for different stages after seedling emergence and after insect infestation
Stage Misclassification of damaged Misclassification of undamaged Total mean
plants plants error, %
Time after seed- Time after insect No. % No. %
ling emergence, infestation, day
day
17 2 8 1778 2 444 1111
18 3 4 889 2 444 667
19 4 0 0 1 222 111
20 5 0 0 1 222 111
21 6 0 0 0 0 0
22 7 0 0 2 444 222
23 8 0 0 4 889 444
24 9 0 0 14 3111 1556
Total 12 3.33 26 722 528
Fig. 5. Correctly classified blocks (a) and (b) of fall armyworm damaged leaves and (c) and (d) undamaged leaves
Figures 4 and 5 show blocks of damaged and non- the image block of Fig. 4(d), causing the classification
damaged plants incorrectly and correctly classified, error.
respectively. The image block of Fig. 4(a) shows an
atypical injury, at the edge of the leaf, which is mixed
up with the background. Figure 4(b) shows a plant 4. Conclusions
presenting too early damage with few visible injuries.
The leaf veins show in Fig. 4(c) were not segmented as An algorithm for identifying damaged maize plants
one object, causing the classification error. The centre by the fall armyworm (Spodoptera frugiperda) using
of the plant was segmented as more than one object in digital colour images was developed and tested. The
ARTICLE IN PRESS
454 D.G. SENA JR. ET AL.
algorithm was tested using damaged and non-damaged Fancelli A L; Dourado Neto D (2000). Produ-ca* o de milho.
maize plant images that were taken under different [Maize Production]. Livraria e Editora Agropecu!aria
lighting conditions and stages of maturity. The pro- Gua!ıba
Kacira M; Ling P P; Short T H (1999). Non-contact sensing of
posed algorithm performed well, correctly classifying plant water stress by ir thermometry and image processing.
9472% of the images. ASAE Paper No. 99 5004
Midgarden D; Fleischer S J; Weisz R; Smilowitz Z (1997). Site-
specific integrated pest management impact on development
Acknowledgements of esfenvalerate resistance in Colorado potato beetle
(Coleoptera:Chrysomelidae) and on densities of natural
enemies. Journal of Economic Entomology, 4(90), 855–867
This research was sponsored by the World Bank and Omoto C; Schmidt F B; Diez-Rodriguez G I; Silva R B (2000).
Embrapa through the project PRODETAB 030-01/99. Resist#encia de Spodoptera frugiperda a inseticidas no Brasil:
The authors have been sponsored by the Brazilian situa-ca* o atual e perspectivas. [Spodoptera frugiperda chemi-
Agencies CNPq and FAPEMIG. All the mentioned cal resistance in Brazil: actual situation and prospection]. In:
XXIII congresso nacional de milho e sorgo, resumos
supports are gratefully acknowledged.
expandidos, Uberl#andia, Associa-ca* o Brasileira de Milho e
Sorgo
Pinto F A C; Sena Jr D G; Queiroz D M; Gomide R L
References (2001).Vis*ao artificial na agricultura de precis*ao. [Machine
Vision on precision agriculture]. In: Avan-cos na agricultura
Borhan M S; Panigrahi S (1999). Multi-spectral imaging de precis*ao no Brasil no per!ıodo 1999–2001 (Balastreire L A,
techniques for nitrogen determination in potato leaf. ASAE ed). ESALQ/USF, Piracicaba
Paper No. 99 5005 Steward B L; Tian L F (1998). Real-time machine vision weed-
Concei-ca* o M Z (2000). Manejo integrado em defesa vegetal. sensing. ASAE Paper No. 98–3033
[Integrated management plant protection] In: Manejo Viana P A (2000). Manejo de pragas na cultura do milho. [Pest
integrado } doen-cas, pragas e plantas daninhas (Zambolim management on maize crop]. In: XXIII congresso nacional
L, ed), pp 1–80, DFP/UFV, Vi-cosa de milho e sorgo, resumos expandidos, Uberl#andia,
Cruz I (1995). A lagarta do cartucho na cultura do milho. Associa-ca* o Brasileira de Milho e Sorgo
[The fall armyworm on maize production]. Sete Lagoas: V!ızh!any!o T; Felfoldi J (2000). Enhancing colour differences in
EMBRAPA-CNPMS, (EMBRAPA-CNPMS. Circular images of diseased mushrooms. Computers and Electronics
T!ecnica, 21) in Agriculture, 26, 187–198
Cruz I; Figueiredo M L C; Alberton M (2000). Efici#encia de Weisz R; Fleischer S; Smilowitz Z (1996). Site-specific
novos piretro! ides no controle da Spodoptera frugiperda em integrated pest management for high-value crops: impact
milho e seletividade ao predador Doru luteipes [New on potato pest management. Journal of Economic Ento-
pyrethroid efficiency for Spodoptera frugiperda controlling mology, 2(89), 501–509
in maize and Doru luteipes selectivity]. In: XXIII congresso Yang X; Beyenal H; Gary H; Lewandowski Z (2001).
nacional de milho e sorgo, resumos expandidos, Uberl#andia, Evaluation of biofilm image thresholding methods. Water
Associa-ca* o Brasileira de Milho e Sorgo. Research, 35(5), 1149–1158