Authors:
Henry Kobs
1
;
2
;
Henrique Krever
1
;
2
;
Denilson Ebling
1
;
2
and
Celio Trois
3
Affiliations:
1
Technology Center, Federal University of Santa Maria, Santa Maria, Brazil
;
2
Zeit Artificial Intelligence Solutions Ltd., Santa Maria, Brazil
;
3
Department of Physical and Inorganic Chemistry Engineering, Rovira i Virgili University, Tarragona, Spain
Keyword(s):
Neural Network, Near-Infrared Spectroscopy, Soybean, Artificial Intelligence.
Abstract:
Soybeans are integral to global agriculture and food production, playing a vital role in human and animal nutrition. Accurate assessment of moisture, oil, and protein contents in soybeans is crucial for various applications, including human nutrition, animal feed, and food manufacturing. This paper introduces SpectraNet, a Neural Network architecture designed for predicting soybean contents using Near-infrared Spectroscopy (NIRS) data. NIRS technology provides a cost-effective and non-destructive means of analyzing agricultural samples. Spec-traNet leverages a 1D convolutional Neural Network and multiple prediction heads, demonstrating its efficacy in handling non-linearities present in spectral data. The architecture’s flexibility and adaptability contribute to accurate predictions, automatic feature extraction, and suitability for varying conditions. Comparative analysis with traditional Partial Least Squares Regression (PLSR) models reveals the superior performance of SpectraNet i
n predicting protein, moisture, and oil contents in soybeans. The presented methodology involves comprehensive data collection, laboratory analysis, and model training, showcasing the potential of SpectraNet for real-world applications in agriculture. The results highlight the efficiency and precision of SpectraNet, offering a valuable tool for advancing agricultural practices and ensuring soybean quality.
(More)