Identification of bagworm (Metisa plana) instar stages using hyperspectral imaging and machine learning techniques

Abstract

Listen

Translate article

• Young larvae can be detected using hyperspectral imaging and machine learning. • The best model was obtained using green bands and weighted k-nearest neighbour. • Larvae activities gave a significant impact on the selection of the best dataset. • Larvae size is nearly identical, therefore morphological features underperform. A serious outbreak of leaf-eating insects namely bagworm (Lepidoptera: Psychidae), especially Metisa plana species, may cause a 43% yield loss in oil palm production due to late proper control of bagworm populations. Identification of the bagworm instar stage is important to ensure proper control measures are applied in the infested area. This study aims to distinguish the bagworm larvae from second (S2) to fifth (S5) instar stages using hyperspectral imaging and machine learning technique. The capability of spectral reflectance and morphological features namely area, perimeter, major axis length, and minor axis length to classify the instar stage were studied. A total of 2000 sample points of larva were extracted from hyperspectral images. It was then followed by the identification of sensitive wavelengths of each stage using analysis of variance (ANOVA). Results show that seven wavelengths from the blue and green band (i.e., 470 nm, 490 nm, 502 nm, 506 nm, 526 nm, 538 nm, and 554 nm) gave the most significant difference in distinguishing the larval instar stages. To provide a more economical approach, only two wavelengths were used for model development. Later, the classifications models were developed separately using five different types of datasets: (A) significant morphological feature, (B) all significant wavelengths, (C) two wavelengths from the same spectral region, (D) two wavelengths from different spectral regions, and (E) two significant wavelengths and a significant morphological feature. Results have shown the dataset which used green bands at 506 nm and 538 nm with a weighted k-nearest neighbour classifier achieved the best value of accuracy (91% – 95%), precision (0.83 – 0.87), sensitivity (0.77 – 0.99), specificity (0.94 – 0.96) and F1-score (0.81 – 0.91). It was mainly due to green pigments which strongly correlates with the chlorophyll content of the frond leaves fed by the larvae to build and enlarge the case. The capability of the model to detect the young larval instar stages (S2 - S3) where an active feeding activity takes place allows quick decisions about outbreak control measures.