CNN inversion model considering texture features and its application to soil selenium content

Soil remote sensing geochemistry typically involves constructing inversion models by correlating geochemical data from samples with spectral data from remote sensing image pixels to infer soil element concentrations. However, the accuracy of modeling using only the emissivity of image element is low. Therefore, this paper incorporates texture information from the images as a modeling factor and constructs a Convolutional Neural Networks (CNN) inversion model that considers texture features, exploring the impact of texture features on the modeling process. Taking selenium (Se) as an example, the study first conducts a correlation analysis between the pretreatment remote sensing data and the soil sample chemical data to select the spectral bands with strong correlations. Then, based on these selected bands, the study uses a 17 × 17 grid of pixels surrounding the sample points as the input and the selenium content at the sample points as the output to construct the CNN inversion model. Finally, the inversion effect of CNN model is compared with Multiple Linear Regression (MLR), Support Vector Machines (SVM), Random Forests (RF) and Backpropagation Neural Networks (BPNN) models conducted by spectral feature alone or a combination of spectral and texture features. This comparison highlights the role of texture features in quantitative remote sensing modeling and the advantages of the CNN inversion model. The results show that compared to the best-performing model based on spectral features alone, SVM (with a test set R² = 0.286), the best model based on spectral and texture features, BPNN (with a test set R² = 0.377), improved the inversion accuracy by nearly 0.1. The CNN model achieved a test set R² of 0.504, significantly outperforming the other models. In conclusion, incorporating texture information into quantitative remote sensing modeling can effectively improve inversion accuracy, and the CNN model demonstrates a clear advantage in soil element inversion.