Artificial Neural Networks applied in the determination of Soil Surface Temperature – SST

Maurício Roberto Veronez 1, Adriane Brill Thum 1, Anderson Silva Luz 2
and Deivis R. da Silva
1 Graduate Program in Geology - Vale do Rio dos Sinos University
Avenida Unisinos, 950, São Leopoldo - RS, Brazil – CEP: 93022-000
Tel.: + 55 (51) 3591-1100 – Extension 1769 Fax: + 55 (51) 3590-8177,
2 Students of Civil Engineering. - Vale do Rio dos Sinos University
Avenida Unisinos, 950, São Leopoldo - RS, Brazil – CEP: 93022-000
Tel.: + 55 (51) 3591-1100 – Ramal 1769; Fax: + 55 (51) 3590-8177

Artificial intelligence techniques are being used to facilitate modeling in many different research areas. One example of these techniques is the use of Artificial Neural Networks – ANNs. ANNs are not a new technique and have been studied since the 1940s. The technique was almost forgotten during the 1970s, but reappeared in the 1980s as a possible alternative to traditional computing. Today, ANNs are used in many projects, especially to forecast data, learn algorithms, optimize systems, recognize standards, and others. The surface temperature (ST) is a parameter influenced by changes in weather (temperature and air relative humidity, wind speed, precipitation, etc.) and indicates the hydric state of a plant. Thus, estimating the ST is very useful in monitoring projects that tend to hydric demands of cultures, which will contribute to irrigation programs. Another important application is in the use of the determination of evapotranspiration, where, together with other components of the hydrological cycle, it is important to evaluate the replenishing of underground water-bearing aquifers. Recently, a method used to estimate the ST uses the analysis of NOAA-AVHRR thermal images adapted to the split windows equation. This modeling relates emission amount of variables (generated by the images) and atmospheric data. It is a complex methodology, because besides difficult statistical modeling, it is necessary to digitally process the images to determine the emission amount. ANNs are indicated in this study due to their excellent capacity for generalization, classification, interpolation, extrapolation, tolerance to errors and noise, and because of the fact that they do not require the specification of explicit parameters to complete the modeling process. The purpose of this project is to verify the possibility of the use of ANNs to estimate soil temperatures. The test area selected is a part of an urban area located in Ivoti - RS. In different points of  the area, values of ST were collected using a portable temperature sensor model. For each collecting point, the position was known by the UTM coordinates (E, N) and the elevation (H). Different network topologies were tested in a supervised way by the backpropagation algorithm, using the E, N and H coordinates, as data entry, and giving ST as results. The best topology found was that possible in the limitations of available time for the execution of the successive necessary refinements. Despite the fact that several tests have been done with two and three  intermediate layers, a simple topology was adopted, with only one intermediate layer of 3 and 5 neutrons. Four learning algorithms were tested. The algorithms, Scaled Conjugate Gradient, Levenberg-Marquerdt, and Resilient, provided values for ST with mean errors below 0.9 ºC, during the simulations, while the Gradient Descendent showed a mean error below 1.9 ºC.

Keywords: Artificial Neural Networks, modeling, soil surface temperature

In: Caetano, M. and Painho, M. (eds). Proceedings of the 7th International Symposium on Spatial Accuracy Assessment in Natural Resources and Environmental Sciences, 5 – 7 July 2006, Lisboa, Instituto Geográfico Português

VeronezThum2006accuracy.pdf925.54 KB