Runoff-erosion modeling at micro-watershed scale: a comparison of self-organizing maps structures

Background In the last decades, everal runoff-erosion models have been proposed to estimate soil erosion, which may lead to loss of fertile land and increase sedimentation and pollution in water bodies. Physically-based erosion models are usually used for such purpose, but a major problem concerning their use is the difficulty to directly measure parameters in the field. This problem can be overcome by exploring empirical models, such as so-called Self-Organizing Maps (SOM). An SOM is a type of Artificial Neural Network (ANN) based on a competitive learning approach for clustering and modeling a variety of databases. Since studies on soil erosion modeling based on SOM are very incipient, we compared some structures of SOM with the purpose of estimating sediment yield based on runoff and climatological data at the micro-watershed scale. The case study was a micro-watershed within the Sumé Experimental Basin, which is located in a semiarid region of Brazil. Different from the conventional ANN, SOM-based models represent a multidimensional data set by means of a bidimensional matrix of features, which may be applied for analysis and estimation purposes. In order to calibrate and validate the proposed SOM structures, we used data from 117 rainfall events that occurred between 1985 and 1991. Results Analyses of the results indicate that all SOM structures were efficiently calibrated with NASH coefficients (Nash & Sutcliffe 1970) varying from 0.88 to 0.90. The SOM structure with 6 × 8 neurons was the most effective for estimating sediment yields when considering the validation data set ( NASH = 0.73). The generated maps showed that sediment yields were directly related to runoff and rainfall intensity and inversely correlated to average vegetation heights. The dry period length did not seem to influence the production of sediments. Conclusions SOM were shown to be very practical and meant to be applied to specific locations. This type of methodology also demands long term data and dynamic recalibration with up-to-date information in order to account for changes in the watershed.