[A-CG50] Earth & Environmental Sciences and Artificial Intelligence

[ACG50-03] Digital Soil Mapping of Soil Fertility in Thung Kula Rong-Hai Region, Thailand using Machine Learning Algorithm

*Sasirin Srisomkiew¹, Masayuki Kawahigashi¹, Pitayakon Limtong² (1.Tokyo Metropolitan University, 2.Land Development Department)

Keywords:Digital soil mapping, Soil fertility, Machine learning, Principal component analysis

Advancement in the Machine Learning (ML) Algorithm domain in the last few decades with the development of applications of machine learning are now try to apply to the field of Digital Soil Mapping (DSM) to express various soil properties on maps. However, soil fertility which is a complex feature of soil has not been mapped using various physical-chemical soil properties. This study challenges to evaluate soil fertility covering whole through a geographical area using inherited soil properties by the digital soil map with a machine learning algorithm. The target area is the Thung Kula Rong-Hai (TKR) region of Thailand, which covers a total area of 337,230 hectares. The TKR region is famous for producing high-quality jasmine rice, where has been registered as the Protected Geographical Indication (PGI) in 2010. However, the productivity of the jasmine rice in the TKR region is very low as compared to other regions in Thailand. Hence, it is necessary to know the heterogeneity of soil fertility for sustainable land use management to help increase in productivity of rice in the TKR region. In this study, remote sensing data from Landsat8 OLI and DEM data were used to characterize soil fertility covering whole through the studied region using soil properties that were pre-determined by laboratory analyses from 186 soil samples according to soil series in Thailand. Soil samples were collected in the dry season between March and May in 2016. The principal component analysis (PCA) using various soil properties extracted suitable soil parameters to map soil fertility. Following the PCA, the machine learning algorithm for Multiple Linear Regression (MLR) was applied to all available data to establish a predictable model to create the DSM of soil fertility. The obtained PC1 was explained by electrical conductivity (EC), organic matter (OM), phosphorus (P), potassium (K), calcium (Ca), and magnesium (Mg) with significant correlation coefficients, suggesting that PC1 can be a soil fertility parameter. On the other hand, PC2 consisting of soil pH, exchangeable iron (Fe) and copper (Cu) with significant correlations express the availability of micronutrients. The MLR analysis using obtained soil data with the principal component analysis led predictable equations using the near-infrared (NIR) and shortwave infrared (SWIR) channels of Landsat 8, as well as spectral indices of saturation (SI), normalized difference vegetation (NDVI), normalized difference water (NDWI), moisture stress (MSI) and terrain index (DEM) to produce DSM. The predictive accuracy evaluated in terms of R² for MLR models is acceptable to create the DSM which ranged from 0.7 to more than 0.9 for different soil parameters. The estimated DSM demonstrates that most of the cultivated areas in the TKR region are occupied by “very low” to “low” fertilities, where covered approximately 172,313 hectares or 51.1% of the total area.