Abstract:
Objective Accurate estimation of soil organic matter (SOM) content has always attracted a lot of attention. Compared with traditional laboratory analysis methods and professional spectrometers, SOM estimation based on color parameters of smartphone images shows advantages in more economic, higher efficiency and greater convenience. The aim of this study was to assess performance of deep learning on estimating SOM based on color parameters of smartphone images.
Method 728 dry soil samples were photographed in a self-made optical darkroom and determined for SOM. Then, median and mean values of six color channels, i.e., Red (R), Green (G), Blue (B), Hue (H), Saturation (S), Value (V) were obtained after gray processing, threshold segmentation, color conversion and other preprocessing, based on the obtained image through Python and MATLAB software. Subsequently, convolutional neural network (CNN), long short-term memory network (LSTM) and random forest (RF) models were established for SOM estimation, respectively,. Accuracies of the models were evaluated through five-fold cross-validation.
Result The results indicates that CNN, LSTM and RF models had excellent performances with coefficient of determination (R2) rangs from 0.732 to 0.856, the root mean square error (RMSE) ranging from 4.721 to 6.455, and the Lin's concordance correlation coefficient (CCC) ranging from 0.843 to 0.917. The comprehensive performance order was: RF model, CNN model, LSTM. R value and V value were the most important color parameters.
Conclusion CNN and LSTM models demonstrated good accuracy in estimating SOM using color parameters from smartphone images. However, in the case of small sample sizes, the accuracy of CNN and LSTM models was slightly lower than that of RF.
下载: