A New Approach for Soil Carbon and Nitrogen Assessment through Image Analysis: A Case Study with Oxisols and Inceptisols

• Juliano Tiago Rinaldi UTFPR http://orcid.org/0000-0002-2262-889X
• Dalcimar Casanova UTFPR https://orcid.org/0000-0002-1905-4602
• Larissa Macedo dos Santos Tonial UTFPR https://orcid.org/0000-0002-7046-0788
• Marcelo Teixeira UTFPR https://orcid.org/0000-0002-1008-7838
• Heitor Silvério Lopes UTFPR https://orcid.org/0000-0003-3984-1432

Resumo

This paper presents a new method for predicting carbon and nitrogen content in Oxisols and Inceptisols soils using image analysis. The dataset comprises 40 Oxisols images and 51 Inceptisols images collected in southern Brazil. The images were segmented to isolate the soil area, resulting in the extraction of 12,744 windows. A pre-trained EfficientNetV2-S model was fine-tuned for regression, adding five layers at the end. The model was evaluated using RMSE and R² metrics. Two experiments were conducted: one predicting carbon and nitrogen content for each window and another predicting the median values of the windows for each soil sample. The results show that the median-based prediction significantly improved model performance, achieving an R² of 0.7425 for carbon and 0.7774 for nitrogen. This method offers a faster, non-destructive alternative to laboratory analysis for estimating soil carbon and nitrogen content, contributing to sustainable agricultural practices.

Palavras-chave: Soil, Carbon, Nitrogen, Transfer Learning, Deep Learning