Sensor multiespectral e aprendizagem de máquina para classificação de híbridos de milho irrigado

Abstract

With the advancement of agriculture, methods for evaluating phenotypes have emerged with the aid of sensing and machine learning techniques. The objective of Chapter 1 is to classify maize hybrids in different irrigation managements through multispectral images, searching for the best machine learning algorithm for classification and input that improves the performance of the models. The objective of Chapter 2 is to find the most accurate machine learning algorithm in the classification of maize hybrids and to determine input data from the models that improve the performance of the algorithms. The experiment was implemented in the experimental area of the Federal University of Mato Grosso do Sul in the municipality of Chapadão do Sul, Brazil. The corn hybrids used in the experiment were: H1 (AS 1868), H2 (DKB 360), H3 (FS 615 PWU), H4 (K 7510 VIP3), H5 (NK 520 VIP3), H6 (P 3858 PWU), H7 (SS 182E VIP3) in two irrigation managements (Irrigated and rainfed). After 60 days of culture emergence, the ARP Sensefly eBee RTK was used to obtain the wavelengths (SB): Blue (475nm, B_475), green (550 nm, G_550), red (660 nm, R_660), edge of red (735 nm, RE_735) and NIR (790 nm, NIR_790). After obtaining the SB data, it was possible to perform calculations of vegetation indices (VIs). The data were subjected to machine learning analysis, testing six algorithms: Artificial Neural Networks (ANN), J48 Decision Trees (J48), REPTree (DT), Random Forest (RF), Support Vector Machine (SVM) and Regression logistics (RL) used as default. Three accuracy metrics were used in order to ascertain the accuracy of the algorithms in classifying corn hybrids: correct classifications (CC), Kappa coefficient and F-Score. In the classification of hybrids, the Artificial Neural Networks (ANN) algorithm was the one that presented the best result. The best input was spectral bands (SB) providing better classification accuracy. In the classification of irrigated and rainfed management, the best algorithms were ANN and RF, using IVs and SB+IVs inputs.