engleski

Diffuse reflectance spectroscopy and model analysis for assessing fire affected soil chemical properties in Croatia

Wildfire impacts on soil chemical properties can be detected through spectral signature. Fire severity effects on soil can be assessed using proximal spectroscopy. This method is non-destructive and estimates several soil properties. The objective of this work is to 1) estimate and monitor the effects of wildfire on 11 soil chemical properties using visible and near infrared (VNIR) hyperspectral reflectance and 2) compare model performance obtained by linear calibration methods and data mining techniques. The study was carried out in the Šibenik-Knin County (Croatia) following a summer wildfire that affected an area of 6 ha. A total of 120 topsoil (0 to 3 cm depth) (10 replicates x 3 treatments x 4 time periods) samples from were collected in the interval of 5 days, 3 months, 6 months and 1 year after the wildfire. Samples were collected in plots affected by medium (MS), high severity (HS) and from a non-burned area (control – C). The following soil properties were determined using standard laboratory methods: soil pH, electrical conductivity (EC), carbonates (CaCO3), plant available phosphorus (P2O5) and potassium (K2O), organic carbon (OC), exchangeable calcium (Ca), magnesium (Mg), potassium (K), sodium (Na) and cation exchange capacity (CEC). Soil spectral measurements were obtained using a portable spectroradiometer. Partial least square regression (PLSR) and artificial neural network (ANN) were used to build prediction models of selected soil properties based on original soil reflectance data and first derivative of reflectance. PLSR model was calibrated using cross validation method. In ANN regression analysis, spectra were randomly divided into training and testing sets with proportions of 50% and 50%, respectively. The results showed that spectral reflectance differed (p<0.05) according to the fire severity and sampling time. Tukey’s post- hoc test showed average reflectance of C samples were significantly different from HS and MS, but there was no difference between HS and MS in term of fire severity in all measured time intervals (p<0.05). Average reflectance was significantly lower in MS and HS treatments compared to C during the investigated period, indicating significant changes in soil properties persist one year after the wildfire (p<0.05). PLSR proved to be the better model for K2O, CaCO3, Ca and CEC prediction, while OC, P2O5, exchangeable Mg and K were predicted better using ANN model. Exchangeable Ca and CEC showed the highest R2 in PLSR model, with R2 =0.80, RMSE =4.67 cmol+ kg-1 and R2=0.82, RMSE =4.52 cmol+ kg-1, respectively. OC and P2O5 showed the highest R2 in ANN model, with R2=0.74, RMSE =4.44 % and R2=0.74, RMSE =6.35 mg kg-1, respectively. Soil pH, EC and exchangeable Na showed poor predictability in both models. Overall, both PLSR and ANN models showed satisfactory results for prediction of most of the studied soil properties and represent a good insight in spatio-temporal post-fire soil changes.

soil spectrocsopy ; soil chemistry ; ANN ; PLSR ; modelling

nije evidentirano