Cascading Feature Filtering and Boosting Algorithm for Plant Type Classification Based on Image Features

Adel Bakhshipour

doi:10.1109/access.2021.3086269

Abstract

Crop and weeds identification is of important steps towards the development of efficient automotive weed control systems. The higher the accuracy of plant detection and classification, the higher the performance of the weeding machine. In this study, the capability of two popular boosting methods including Adaboost.M1 and LogitBoost algorithms was evaluated to enhance the plant classification performance of four classifiers, namely Multi-Layer Perceptron (MLP), k-Nearest Neighbors (kNN), Random Forest (RF), and Support Vector Machine (SVM). Four feature filtering techniques including Correlation-based Feature Selection (CFS), Information Gain (IG), Gain Ratio (GR), and OneR were applied to the image-extracted features and 10 of the most significant features were selected and fed into single and boosted classifiers. The RF model trained by IG selected features (IG-RF) was the most appropriate classifier among the evaluated models whether in single or boosted modes. It was also found that boosting of IG-RF by using Adaboost.M1 and LogitBoost algorithms improved the classification accuracy. Regarding the performance values, the LogitBoost-IG-RF structure, which provided a classification accuracy of 99.58%, a kappa ( k) of 0.9948, and a Root Mean Squared Error (RMSE) of 0.0688 on training dataset, was selected as the most appropriate classifier for plant discrimination in peanut fields. The accuracy, k, and RMSE criteria of this combination on test dataset were 95.00%, 0.9375, and 0.1591, respectively. It was concluded that combination of boosting algorithms and feature selection methods can promote plant type discrimination accuracy, which is a crucial factor in the development of precision weed control systems.

Highlights

Presence of weeds in fields and their competition with the main plant for water, light, nutrients, and space can cause irreparable damage to crop performance if the weeds are not appropriately treated.Yield losses from 37% to 61% were reported by Dille et al [1] in grain sorghum with weed interference in different regions of the United States
The Multi-Layer Perceptron (MLP) and Support Vector Machine (SVM) classifiers resulted in lower performances than k-Nearest Neighbors (kNN), and Random Forest (RF) classifiers when used as the weaker classifiers in the LogitBoost ensemble algorithm
The most successful LogitBoost structure for plant type classification had the base classifier of RF and trained using the Information Gain (IG) feature selection method

Summary

Introduction

Presence of weeds in fields and their competition with the main plant for water, light, nutrients, and space can cause irreparable damage to crop performance if the weeds are not appropriately treated.Yield losses from 37% to 61% were reported by Dille et al [1] in grain sorghum with weed interference in different regions of the United States. The associate editor coordinating the review of this manuscript and approving it for publication was Senthil Kumar. Precision weed control such as selective spraying or accurate mechanical removal of weeds, is a challenging task that aims to reduce the amount of herbicides without compromising the quality of crops [5]. Accurate weed detection in croplands, as a prerequisite for applying any precision weed management technology [6], is still a challenging step toward the development of site-specific weed control machines, especially when there are intra-row weeds that are highly overlapped with the main plant. Efficient weed removal weather using variable-rate sprayers, or precise mechanical, electrical, or thermal hoeing systems, preliminary requires to detect and segregate weeds from main crop [7].

Methods

Results

Discussion

Conclusion