Uso de regresão linear múltipla e redes neurais artificiais na estimativa da interferência de plantas daninhas em cultivos agrícolas

Abstract

The efficiency of weed control in agricultural crops depends on the adoption of the control method at the right time, minimizing the interference of the weed community. However, a definition of the beginning and end of the control season is difficult to measure by the producer due to the particularities of the weed community, the crops, and the management and edaphoclimatic conditions of each cultivation site. One of the alternatives to solve this problem is the use of models based on statistical methods or methods of machine learning developed in the field of artificial intelligence that can be adequate to predict the loss of yield of agricultural cultures due to the interference of weeds. Thus, the objective was to verify if the RLM and RNAs models can predict the beginning of weed control, compare the effectiveness of RNAs with traditional sigmoidal models and also evaluate the ability of RNAs to estimate weed control for different crops and acceptable productivity loss classes and thereby validate a new alternative for modeling and predicting competition between weeds and agricultural crops. The experiments were conducted at the Rafael Fernandes Experimental Farm, Mossoró, RN. The design used was in randomized blocks, with three replications. The treatments consisted of six periods for the beginning of weed control (0, 7, 14, 21, 28, 32 and 42) days after emergence (DAE) for the onion culture and for the sesame and the melon foramen (130, 260, 390, 520, 650, 1300) and (130, 260, 390, 520, 972) degrees days (GD) after sowing sesame and transplanting melon seedlings, respectively. The RLM models based on non-destructive and destructive inputs on the weed community are able to estimate losses of onion productivity, however with low precision. The ANN models considering only the coexistence period and the irrigation system have similar performance to the Multiple Linear Regression models. However, the inclusion of variables related to weed density (non-destructive) and fresh matter (destructive) in RNA models raises the predictive capacity of networks to values close to 99% correct. The RNA model, which made a combination of non-destructive (density) and destructive (Matter Fresh) inputs, dispenses with other more specific inputs such as the C3 / C4 and Monocotyledons / Eudicotyledons (M / E) species. The weed interference period is the main factor to infer the degree of weed interference in the Sesame and Melon cultures. The models of RNAs with the best performance can indicate the beginning of weed control, since they are able to accurately estimate losses caused by weed interference