Modelagem de variáveis hídricas por meio do sensoriamento remoto orbital e inteligência artificial

Abstract

Soil and water are essential natural resources for sustaining life on Earth, and managing these elements properly is crucial to balancing human needs with environmental conservation. In the hydrological cycle, evapotranspiration is one of the main components of the water balance, and understanding it is key for watershed management and conservation. It is also used to guide irrigation practices, improving water efficiency in agriculture and increasing crop productivity. Besides supporting life, water is the primary cause of soil erosion in Brazil because rainfall supplies the energy for erosive processes. Soil is a non-renewable resource on a human time scale, and erosion rates have been increasing. Rainfall energy—known as erosivity—is a major factor driving sediment loss worldwide. Erosivity is used in erosion prediction models such as the Universal Soil Loss Equation (USLE) and its revised version (RUSLE), but obtaining these values is challenging, as they are generally measured at specific points and cannot be reliably extrapolated to other areas needing soil loss estimates. Traditional spatialization techniques, such as kriging, inverse distance weighting, or Thiessen polygons, often do not capture the variability of natural environments adequately. On the other hand, machine learning models, combined with geographic information system (GIS) products and satellite imagery, have recently been used as effective tools to monitor and model various features of the Earth’s surface.In this study, we developed tools to model reference evapotranspiration (ETo) and rainfall erosivity for the entire country of Brazil using satellite remote sensing and artificial intelligence. The study’s objectives were to: (i) model ETo in real time across Brazil, making images available within 10 minutes after the end of the day; and (ii) model and produce a rainfall erosivity map for Brazil with a spatial resolution of 30 arc-seconds (approximately 1 km2). For ETo modeling, we used WorldClim products, daily extraterrestrial radiation data, and daily evapotranspiration means and standard deviations from meteorological stations of the National Institute of Meteorology (INMET), along with machine learning algorithms. Performance metrics (RMSE, nRMSE, MAPE, d, and NSE) were used to evaluate and select the best model. The Cubist model showed the best performance for ETo, with a MAPE of 17.4%, RMSE of 0.792, nRMSE of 57.3%, d of 0.883, and NSE of 0.634. For rainfall erosivity modeling, we used data from scientific studies (erosivities measured by pluviographs), GIS techniques, and machine learning. Out of 31 pre-selected covariates, seven were used in the model—in order of importance: longitude, solar radiation, annual precipitation, precipitation in the coldest quarter, wind speed, precipitation in the hottest quarter, and annual reference evapotranspiration. After 400 cycles of training and validation, the Random Forest model (using median values) provided the best performance for rainfall erosivity, with an NSE of 0.5823, RMSE of 1,567.17 MJ mm ha−1 h−1 year−1, MAE of 1,135.90 MJ mm ha−1 h−1 year−1, nRMSE of 58.50%, ME of –17.76 MJ mm ha−1 h−1 year−1, and d of 0.8487. We conclude that the ETo product is highly useful and feasible for estimating reference evapotranspiration across Brazil. In addition, climatic and geographic variables can be effectively used to construct rainfall erosivity map’s with known error margins and accuracy on a national scale.