Enhancing Irrigation Management: Unsupervised Machine Learning coupled with Geophysical and Multispectral Data for Informed Decision-Making in Rice Production

Abstract

Integrating diverse data sources for site-specific management zones (SSMZ) in precision agriculture is a complex task. Soil surveys using apparent electrical conductivity (ECa) have proven effective in capturing field variability. However, relying solely on one sensing data type may not fully capture the intricate relationship between soil properties and crop physiological processes. This research assessed the effectiveness of applying multivariate geostatistical analysis and unsupervised machine learning (UML) to geophysical and multispectral data through ECa, NDWI and NDVI indices, for delineating and validating the SSMZ at different crop cycles in five rice field of Tolima department-Colombia. Lee´s correlation was conducted between ECa, NDWI and NDVI data. MULTISPATI-PCA was used to assess the autocorrelation in the input variables before SSMZ delineation. The correlation results revealed two field categories based on the direct or indirect relationship between ECa and the multispectral data. On one hand, ECa values showed an inverse correlation (p<0.01) with multispectral data (-0.25 to -0.45) for fields 1a, 1b and 4a. On the other hand, there was a direct correlation (p<0.01) between ECa and multispectral data (0.31 to 0.51) for fields 2b and 5a. For all rice fields, MULTISPATI-PCA explained a range between 75.01 % to 95.64 % of the total variability. UML results highlighted two, three and four SSMZ, with two SSMZ being the optimal management zones number. The validation suggests that the θ data showed a significant statistical difference for two SSMZs (p < 0.05), regardless of the water content ranges. This study emphasized the importance of integrating geophysical and multispectral datasets in precision agriculture for evaluating field variability and crop conditions.