Estimation of soil organic matter by UAV hyperspectral remote sensing in coal mining areas

Abstract: This study aims to monitor the soil quality of the cultivated land in different subsidence stages of the coal mining areas, particularly for land reclamation and quality protection. Three types of cultivated land were taken as examples around the Wangzhuang Coal mine in Changzhi City, Shanxi Province, China. The images were first acquired by an Unmanned Aerial Vehicle (UAV) equipped with a hyperspectral camera. The soil samples were then collected to carry out the indoor spectrometry in the study areas. Four transformations of spectral reflectance were also performed on the images, including the reciprocal, the first-order differential, the second-order differential, and multivariate scattering correction. A correlation analysis was conducted to select the sensitive bands with the higher correlation coefficient between the converted spectral reflectance and measured organic matter content. A prediction model was established for the content of soil organic matter using the Multiple Linear Regression (MLR), Partial Least Squares Regression (PLSR), and Back Propagation Neural Network (BPNN). The optimal model was achieved to map the organic matter content in the UAV aerial hyperspectral images, in order to determine the distribution of soil organic matter in the cultivated land range. The accuracy of the prediction model was evaluated to obtain the spatial differences and driving factors of soil organic matter in different subsidence stages of cultivated land. The results show that: 1) The soil organic matter content of the cultivated land in the coal mining subsidence area has the highest correlation with the spectral curve transformed by the multivariate scattering correction and the organic matter content. The sensitive band is 463.75-492.45 nm, 870.79-932.58 nm, the maximum correlation coefficient is 0.63. 2) The prediction accuracy of the organic matter content reached 0.863 and 0.884 in the PLSR and BPNN model in the spectral curve that was processed by multiple scattering correction, respectively, which were significantly higher than that of the MLR model. It infers that the prediction models were feasible to identify the organic matter content. 3) There was a relatively uniform distribution of soil organic matter in the cultivated land in the coal mining subsidence areas. Specifically, the distribution of soil organic matter presented with an average value of 26.94 g/kg in the undisturbed area of coal mining, which was in the upper-middle level in general. Nevertheless, there was great spatial heterogeneity for the overall distribution of soil organic matter in the disturbed and stable subsidence areas of coal mining. More importantly, there was an outstanding differentiation of high and low values, particularly with a proportion from Grade 1 to 6. The disturbed area of coal mining was between the above two grades. The relationship of organic matter content in the mining area was ranked as follows: the cultivated land in the undisturbed > the disturbed > the disturbed and settled, areas of coal mining. Consequently, the reason was the surface deformation, physical and chemical properties of soil, as well as the land evolution in the vegetation and human management before and after coal mining.