MACHINE LEARNING INTEGRATION, REMOTE SENSING DATA PREPROCESSING TECHNIQUES TO MAP PESTS COTTON CROPS

Abstract

Cotton has a considerable economic impact on agribusiness. Strategies to reduce production loss due, for example, to pest attacks are increasingly required. Spodoptera frugiperda, known as fall armyworm, causes irreversible damage to cotton. In this context, a current approach is the use of hyperspectral measurements obtained by remote sensors and processed by machine learning algorithms. However, such measures generate data redundancy, making it difficult to extract information. An alternative is to apply pre-processing techniques, but little is known about the impact these generate on the learning ability of algorithms. This work evaluates the performance of machine learning algorithms in identifying cotton plants attacked by pests using pre-processed and raw hyperspectral measurements. Data are collected by EMBRAPA, and consist of hyperspectral measurements, in the range of 350-2500 nm, referring to eight days of collections in healthy cotton plants and attacked by Spodoptera frugiperda. Pre-processing techniques to try are baseline removal, smoothing, first and second order derivatives. A group of machine learning algorithms, such as Random Forest, Support Vector Machine, Extra Tree, was used to model pre-processed and non-pre-processed hyperspectral measurements. According to the proposed metric, the F-Score and the Extra Trees (ExT) algorithm performed better (0.77). So it overlapped the other results with the preprocessed dataset. In addition to obtaining the most important lengths for the algorithm to have its best performance. Concluding that machine learning with spectroscopy can help the field in a promising way. Studies in other crops and with other factors applied to the plant are recommended.