An impact analysis of pre-processing techniques in spectroscopy data to classify insect-damaged in soybean plants with machine and deep learning methods

Spectroscopy is essential to understand a series of phenomena in multiple fields study. In remote sensing, vegetation analysis one the most prominent explore, aiming improve specific task. As task, modeling insect damage plants establish correct management agricultural farmlands. Hyperspectral data, which can be acquired with field spectroscopy at plant or leaf level, non-direct, rapid, and trustworthy approach indicate its health. However, spectral redundancy inherent challenge for information extraction process, making pre-processing phase an part analysis. Currently, artificial intelligence techniques, mostly based on machine deep learning methods, are standard application data processing, being techniques it. But few studies aimed measure impact such processes monitoring, specifically data. Here, we provide algorithms’ performance over said classification For this, used spectroradiometer that operates within 350–1,000 nm 1,000–2,500 ranges. The dataset was composed measurements took place different days controlled environment soybean plants. methods like baseline removal, smoothing, first second-order derivatives, normal variate (SNV), multiplicative scatter correction (MSC), principal components (PCA) were investigated. Several algorithms method applied model datasets. measured validation metrics relate accuracy. Our results indicated Extra-Tree (ExT) algorithm better, mainly when first-order derivative extracted from (accuracy equal 93.68%). A ranking contributive region situates near-infrared, between 784 911 nm. investigation also demonstrates neural network (DNN) did not return satisfactory result raw reflectance considering combination PCA 2nd it achieved similar ExT 91.95%). implications such, alongside approach, discussed this paper. We hope presented here serves as framework future research applying