Associação entre variáveis climáticas e índices de vegetação na estimativa da maturação de amendoim utilizando redes neurais artificiais

Abstract

The assessment of peanut maturity (PMI) using the maturity frame is very subjective. One of the solutions to reduce these errors is the combination of artificial intelligence with vegetation indices (IVs) and climate data. For the evaluation of the NASA-POWER platform, four trade fields were used for cultivating Georgia-06G located in the state of Georgia, USA (Tifton, Dougherty, Berrien, Coffee). The PMI was evaluated into 5 seasons for Dougherty and 4 seasons to Tifton and 6 seasons of Coffe and Berrien from the Hull-scrape method. The climate variables evaluated were maximum temperature (tmax), minimum (tmin), average (tmean), solar radiation (Qg), wind speed (WS), relative humidity (UR) and surface pressure (PS). linear regression analysis was carried out taking as independent variable the values of NASA-POWER and as dependent variables the value of surface weather stations. The metrics of accuracy and precision were the square root of the average error (RMSE), determination coefficient (R2) and Pearson correlation (p<0,05). Analysis of the main components (PCA). The variables tmax, tmin, tmean and Qg present the best adjustment of R2 at 0.95, 0.96, 0.91, 0.94, respectively. On the other hand, WS and UR did not present a good adjustment with R2 of 0.34 and 0.38. The results comparing the NASA-POWER platforms (predicted) and surface weather stations (observed) demonstrate that the NASA-Power platform is accurate and accurate for providing the climate data for Qg, Tmax, Tmin, Tmean and PS. Two other Brazilian states (Minas Gerais and São Paulo) were included to evaluate maturity. In MG counted one area and SP 6 areas, in both areas the maturity was assessed in 5 seasons by the Hull-scrape method. From that, four MLP type models were calibrated one for each state (GA, MG and SP) and another from the union of all locations (Global). For each site nine vegetation indices (IVs) were calculated and four climate variables were used in addition to the AGD and the red and NIR bands. In addition, PCA, Pearson correlation (p<0,05) and sensitive analysis were used to select the combinations between the input parameters. The training algorithms used were Backpropagation and Levenberg-marquardt individually and combined. The calibrated networks were evaluated according to RMSE, MAE and R2. The best calibrated model was for the state of Georgia, which presented the values of 0,9428, 0,080 and 0,060 of R2, RMSE and MAE, for the test. Among the training methods the Levenberg-marquardt was the best, and among the IVs and climate data the MNLI, AGD and Qg were the ones that stood out. The other locations showed a variation of 4% and 9% from R2 and 10% (maximum) for RMSE compared to the GA model in the test. Calibrated models are able to estimate PMI, however, studies should be developed for other peanut genotypes and on soil types.