TouceiraTech: um Farm Management Information System para pecuária de precisão baseado em predição com redes neurais recorrentes

Abstract

To increase productivity in beef cattle ranching it is essential to improve pasture manage ment, which demands periodic measurement of the mass and accumulation of forage and adjustment of stocking rate. Therefore, it is important to develop tools capable of helping cattle ranchers in this process. The present work, supported initially by an exploratory methodology and later by an explanatory one, proposes, implements and evaluates Tou ceiraTech, a prototype of FMIS for Precision Cattle Raising capable of collecting, storing, preprocessing, predicting and visualizing data on forage accumulation rate, necessary for adjusting the stocking rate in a georeferenced area of interest. TouceiraTech was designed in an interdisciplinary way, from the expansion of an original model of forage availabil ity prediction, based on machine learning with LSTM-type Artificial Neural Networks with field vegetation data obtained by historical, direct, destructive sampling in experi ments carried out to evaluate animal production in different management conditions of the native field. In the prediction model are aggregated meteorological data, remotely collected from an automatic weather station near the region of interest, and also mete orological forecast data. These data are automatically pre-processed by TouceiraTech to estimate specific variables, with emphasis on a new automated approach to estimate evapotranspiration. In addition, it allows the use of non-destructive indirect sampling of vegetation, based on NDVI, from the processing of aerial images incorporated into its spa tial database. The results demonstrate the effectiveness of automated periodic collection of meteorological data needed for prediction, from remote databases of INMET, INPE and NOAA. In particular, they prove that changes in the model training approach, in a stratified way, complemented by adjustments in its input variables, allowed the design of specialized models by treatment types with accuracy significantly higher than that of the original model. Additionally, they indicate the potential of incorporating georeferenced aerial images to enable a decision support system to adjust stocking rates based on indirect sampling with low operational cost, replacing the direct method.