Digitization of soil organic carbon with artificial neural network and multivariate linear regression in Kurdistan province

Soil organic carbon plays a vital role in climate control and environmental sustainability. It also has a key impact on the physico-chemical and biological properties of the soil as it is considered an indicator of soil health. Therefore, investigating the spatial distribution of soil organic carbon is one of the requirements of climate and soil management planning. Traditional methods of estimating soil organic carbon are costly and time consuming and cannot be replicated and extended to similar locations. With the advancement of technology and the ever-increasing need for cost-effective information, data mining and satellite imagery and land parameters have been digitized by soil features. Digital soil mapping is the development of a numerical or statistical model of the relationship between environmental variables and soil properties that is used for large geographic data to generate a digital map. The three main goals of soil digital mapping are: 1) inferring the relationship between environmental variables and soil characteristics, 2) producing and presenting data that better demonstrate soil-geography coherence, and 3) applying expert knowledge in model design. Digital mapping also develops the potential of pedology and soil geography by creating insights into burial processes.

In this study 110 soil samples along with 101 auxiliary parameters were used to predict soil organic carbon in Kamyaran city (Kurdistan province). Multivariate linear regression models and artificial neural networks were modeled using JMP software.

The results showed that soil organic carbon content was highest in the western and northwestern parts of the study area and was related to forest cover and pasture areas. On the other hand, higher altitudes have higher estimated organic carbon. Auxiliary variables of the channel network base level (40%), band 4 (23%), leaf water content (20%), vector terrain roughness (19%), vertical distance to channel network (18%), catchment slope (18%), Normalized vegetation difference index (17%), catchment area (16%), aspect (16%), dem (16%), band 3 (15%), reflectance absorption index (14%), band 1 (14) %), Rain (13%), band 5 (13%), air temperature (12%), vegetation index (11%), topographic wetness index (10%), vegetation index (10%) and so on had the greatest effect on soil organic carbon modeling, in the artificial neural network model. Modeling soil organic carbon distribution by artificial neural network (R2 = 0.97) performed better than multivariate linear regression (R2 = 0.59).

The results of this study showed that the distribution of organic carbon is more influenced by topographic, vegetation and climate factors. In areas where sampling is not possible in the whole area for any reason, it can be used through environmental data such as topographic, climatic and vegetation parameters and with new data mining methods to estimate soil organic carbon.