Classification model for chlorophyll content using CNN and aerial images

• Published in: Computers and Electronics in Agriculture
• Main Author: Wagimin M.N.; Ismail M.H.B.; Fauzi S.S.M.; Seng C.T.; Latif Z.A.; Muharam F.M.; Zaki N.A.M.
• Format: Article
• Language: English
• Published: Elsevier B.V. 2024
• Online Access: https://www.scopus.com/inward/record.uri?eid=2-s2.0-85192500783&doi=10.1016%2fj.compag.2024.109006&partnerID=40&md5=c3752ceef6986184f1aae8975a06c23f

Chlorophyll content is usually used as a quantitative measurement of plant health. The chlorophyll content is also a continuous number of data type, leading to a regression approach when developing the deep learning model. The regression model will predict the chlorophyll content in number format, which requires experts to analyse the outcome. Nevertheless, the analysis of the outcome could possibly lead to human error in diagnosing the plant's health condition. Therefore, this study proposed a classification approach in developing a deep learning model to analyse the plant's health condition without human intervention. The classification approach requires a discrete group for dependent variables instead of continuous numbers. When forming the chlorophyll content index groups in this study, which are low, optimum and high levels, two research studies were combined to form the groups, which were (1) the product of the standard range of nitrogen value in mango plant and (2) the correlation analysis between nitrogen value and chlorophyll content index. The classification model in this study used transfer learning algorithms, which were InceptionV3, DenseNet121 and ResNet50, with the canopy-scale level of mango plant RGB images with complex leaf structures in an uncontrolled and open area. Based on the findings, the classification model could classify the chlorophyll content index levels on both mango plant images, which were infected and not infected with black sooty mould. The finding also shows that a clearer distribution pattern of spectral information extracted from the mango plant images can influence the performance result of the classification model. Besides that, the starting point of the Digitization Footprint for this study site across the development stages of the classification model was 308.5756 MB/ha. Finally, the overall accuracy performances for the classification models that used the transfer learning algorithms, which were InceptionV3, DenseNet121, and ResNet50, and trained using the images of the mango plant infected with pest were 96.49 %, 92.98 %, and 89.47 %, respectively, and for using the images of the mango plant not infected with pest were 88.10 %, 78.57 %, and 69.05 %, respectively. © 2024 Elsevier B.V.