Modeling Soil Nitrogen Using Remote Sensing, Regression and Random Forest Models

Soil is one of the important natural resources of any country, which plays an important role in preserving the environment and producing food. Increasing and decreasing the amount of total soil nitrogen due to various agricultural methods, the entry of industrial wastewater into water and other factors, leads to microbial contamination of soil, reduced vegetation and deficiency in agricultural products needed by humans. Mapping soil nutrient distribution helps mangers in decisions. Since laboratory analysis of these parameters is time consuming and costly across large scales, attempts have been made in recent years to study soil nitrogen based on remote sensing techniques. In this regard, the present study investigated the applicability of remote sensing predicting soil total nitrogen in the east of Lenjan city.

Nitrogen reference points were identified by analyzing 50 randomly-selected surface soil samples from 0-20 cm depth. Nitrogen of soil samples was measured by Kjeldahl method after drying soil at 25 ° C, passing through a 2 mm mesh sieve and transferring to the laboratory, to compare the final results obtained from field sampling and remote sensing technology. Landsat 8 OLI Satellite Image of 2019 (Path 164/Row 37) was obtained and geometric and radiometric correction were applied. Cloud cover for image provided was less than 10%. To reduce the effect of atmospheric diffusion on the quality of image, radiation and atmospheric correction were performed using the FLASH model. the Landsat-8 satellite image (rows 164 and 37) taken on 15 Sep. 2019 and along with three topographic indices of elevation, slope and topographic wetness index (TWI) were introduced to the multiple linear regression and random forest models. 

The digital elevation map of the area showed elevation values ​​between 1100 and 2050 meters. The slope of the study area was less than eight percent. Numerical values ​​of TWI index near water bodies were found to be 0.77. DVI and EVI index values ​​increased with increasing vegetation cover. NDVI index showed values ​​higher than 0.3 and NDWI index as a water index showed a maximum value of 0.77. The SAVI index showed high differences from areas without cover to sparse cover and areas with strong vegetation. SBI index and SI salinity indices showed very high variability in terms of soil parameters in barren lands. Nitrogen regression model was built using three indices RVI, DVI and TWI with p-values of 0.049 and 0.00, respectively. In the nitrogen random forest model, however, plant and soil indices played a more important role in model construction with an of r2 value of 0.44.

Total soil nitrogen in soil parameters showed correlation with density and sand and clay from soil texture and in topographic parameters with elevation and spectral indices with EVI RVI, SAVI, NDWI, NDVI and DVI at the level of 0.01 and with SI3 of salinity indices at the 0.05 level. In soil parameters, silt is correlated with sand and clay at the level of 0.05 and sand with clay as well as density with clay are correlated at the level of 0.01. The results of this study showed that the topographic condition of the region along with red and near infrared-based indices had a significant role in predicting soil total nitrogen. Results also showed a slight difference was observed between the two models in predicting soil nitrogen.