Articles
Optimizing image classification hyperparameters in tree orchards using Sentinel-2 data
Article number
1422_63
Pages
525 – 533
Language
English
Abstract
Using optical remote sensing images, the classification of crops in agricultural areas with trees is crucial for optimal management of agricultural and forest lands.
The classification performance depends on various factors, like the size and distribution of training and validation samples, the algorithms used, the features of agricultural cover, and the specifications of the satellite (including time and space resolution). In this investigation, two different research sites were used.
Firstly, the Mejjate Plain, located 70 km west of Marrakech, is characterized by the predominance of wheat and olive crops and a dry climate.
Secondly, the Smimou mountainous area, about 200 km east of Essaouira, has a humid climate and is mostly covered with argan trees.
The objective was to identify optimal hyperparameter values in random forest (RF) and support vector machine (SVM) classifiers on two different sites to achieve the highest accuracy.
The ground truth data were collected from Mejjate and Smimou for the years 2018 and 2019, respectively.
The results illustrate the performance of the SVM in both locations.
The influence of the gamma parameter on accuracy was examined.
The use of a grid search approach allowed the identification of optimal configurations (kernel=rbf and C=10). This resulted in an accuracy of 97.17% (kappa=96.14%) in Smimou with gamma=121, and 99.36% (kappa=99.21) in Mejjat with gamma=221. These findings expand the understanding of an approach in diverse geographical and temporal contexts, underscoring the importance of carefully considering elements such as the specific study site and algorithm selection to optimize image classification performance.
The classification performance depends on various factors, like the size and distribution of training and validation samples, the algorithms used, the features of agricultural cover, and the specifications of the satellite (including time and space resolution). In this investigation, two different research sites were used.
Firstly, the Mejjate Plain, located 70 km west of Marrakech, is characterized by the predominance of wheat and olive crops and a dry climate.
Secondly, the Smimou mountainous area, about 200 km east of Essaouira, has a humid climate and is mostly covered with argan trees.
The objective was to identify optimal hyperparameter values in random forest (RF) and support vector machine (SVM) classifiers on two different sites to achieve the highest accuracy.
The ground truth data were collected from Mejjate and Smimou for the years 2018 and 2019, respectively.
The results illustrate the performance of the SVM in both locations.
The influence of the gamma parameter on accuracy was examined.
The use of a grid search approach allowed the identification of optimal configurations (kernel=rbf and C=10). This resulted in an accuracy of 97.17% (kappa=96.14%) in Smimou with gamma=121, and 99.36% (kappa=99.21) in Mejjat with gamma=221. These findings expand the understanding of an approach in diverse geographical and temporal contexts, underscoring the importance of carefully considering elements such as the specific study site and algorithm selection to optimize image classification performance.
Publication
Authors
E. El Moussaoui, A. Moumni, S. Khabba, A. Lahrouni, S. Sbahi, A. Chehbouni
Keywords
classifier optimization, grid search, remote sensing, temporal context, supervised classification
Online Articles (63)