What question did this study set out to answer?

The study aims to enhance prediction accuracy and generalization of soil chromium content models using CNNs.

February 19, 2026Land Degradation and Development0 citations

Prediction of Chromium Content in Farmland Soil Around Plateau Mining Areas Based on Convolutional Neural Networks and Generalization Performance Study

CFC.M. FuKunming University of Science and Technology AWA.L. WangKunming University of Science and Technology ATAnhong TianKunming University of Science and Technology

Puntos clave

The study aims to enhance prediction accuracy and generalization of soil chromium content models using CNNs.
Collected 121 soil Vis-NIR spectroscopy samples for model training and validation
Tested 62 additional samples for model generalization in downstream environments
Compared performance of five different CNN models: LeNet-5, AlexNet-8, VGGNet-7, GoogleNet-7, and ResNet-13
Utilized DeepExplainer to analyze feature band contributions in the ResNet-13 model
All CNN models showed excellent prediction performance with R2 over 0.86
ResNet-13 model had the highest accuracy with R2 of 0.9194 and RMSE below 150 mg kg-1
ResNet-13 also demonstrated strong generalization with R2 of 0.9065 and RMSE of 176.9822 mg kg-1
Identified key feature bands for predicting soil chromium concentration in the 429 to 2483 nm range

Resumen

ABSTRACT There are few quantitative prediction studies on farmland soil chromium content in Yunnan. In order to improve the accuracy of the prediction model and increase the generalization ability of the prediction model. In this study, a total of 121 soil visible and near‐infrared (Vis–NIR) spectroscopy in upstream and midstream are collected for model training and validation, and 62 soil Vis–NIR samples in downstream are used for testing the generalization ability of the model in new environments. This study systematically compares, for the first time, the performance of five mainstream convolutional neural network (CNN) models with distinct characteristics—traditional LeNet‐5, AlexNet‐8 with large convolution kernels, VGGNet‐7 with small convolution kernels, GoogleNet‐7 incorporating the Inception structure, and ResNet with residual connections—in predicting chromium content in plateau soils, as well as their generalization ability on new environmental datasets. The simulation results showed that the prediction performance of the five different CNN models constructed for soil chromium content was excellent, with R 2 reaching more than 0.86, RPD reaching more than 2.69 and RMSE reached below 195 mg kg −1 on the validation set. In the experiment for quantitatively predicting soil chromium content, the ResNet‐13 model based on residual learning performed the best, with R 2 of 0.9194, RMSE of 147.5360 mg kg −1 , and RPD of 3.5214 on the validation set. In the model generalization ability experiment, the ResNet‐13 model also had the best generalization ability, its R2, RMSE, and RPD were 0.9065, 176.9822 mg kg −1 , and 3.2702 respectively. Meanwhile, the DeepExplainer interpreter based on the SHAP library were used to analyze the contribution of the feature bands in the ResNet‐13 model, it was found that the peaks near 429, 535, 795, 1003, 1407, 1827, 2197, 2291, and 2483 nm were the main feature bands for predicting soil heavy metal chromium concentrations. This study provides a reliable reference value for quantitative prediction of soil heavy metal chromium concentration and model generalizability research.

Preguntar a la IA

Me gusta

Guardar

Ver artículo completo