Search

• Smartphones show excellent ability to predict original soil properties. • Image segmentation significantly improves soil property prediction accuracy. • Significant differences in dominant regions between segmentation methods. • The K-means clustering segmentation method outperforms the OUST method. Prediction of soil properties based on images of original soil acquired from smartphones is considered a potential alternative to conventional soil property determination due to its convenience and low cost. However, images of the original soil may contain distracting information such as fissures, shadows, and rocks, which may affect the image feature extraction and property prediction results. Four unsupervised image segmentation methods, namely the maximum interclass variance (OUST) method, the local adaptive threshold segmentation (LATS) method, the K-means clustering segmentation (KMCS) method, and the natural break point segmentation (NBPS) method, were used to segment 62 images of original soil taken by smartphones in outdoor sunny conditions. The number of segmentation levels (N) for KMCS and NBPS varied from 2 to 10. The dominant region, determined as the region with the highest percentage of image area after segmentation, was selected for subsequent property prediction. The results of the segmentation methods revealed significant differences in the dominant regions, leading to significant variations in color and texture features (p < 0.05). The random forest approach was used to construct predictive models for soil Munsell colour, soil organic matter (SOM) content, and soil texture content. The validation results demonstrated that the image segmentation methods improved the prediction accuracy of all three soil properties compared to that of the unsegmented images. Among the segmentation methods, KMCS (N = 3 and 4) exhibited the highest segmentation performance, surpassing the widely used OUST method. This research highlights the significance of image segmentation methods in predicting soil properties based on images of original soil samples and recommended it as an essential process for future soil image processing. The findings of this study have the potential to advance the development of smartphone-based prediction of properties of original soils. [ABSTRACT FROM AUTHOR]

The ability to quickly and accurately determine soil organic matter (SOM) content is critical for effective soil management decisions. Using the colour of soil images captured by smartphones to predict SOM content has emerged as a promising alternative to traditional wet chemistry methods. However, natural environments can present a complex array of light conditions that can compromise the accuracy and consistency of soil image colour acquisition, thus limiting the method's applicability. To address this issue, we propose five colour calibration patterns (C0 (no calibration), C1 (neutral grey), C2 (RGB), C3 (RGBCMY), and C4 (24 colours)), based on a 24-colour standard card. These patterns were used to calibrate the images of 352 original soil samples obtained from smartphones in three different lighting environments - L1 (100–2000 lx), L2 (35,000–40,000 lx), and L3 (75,000–80,000 lx). Random forest models were used to construct predictive models of soil organic matter (SOM) content based on images. Our findings indicate that smartphones exhibit complex spectral response characteristics, which result in poor image accuracy and stability of uncalibrated (C0) images under varying lighting conditions. The uncalibrated (C0) soil images in different lighting environments exhibited high colour difference (∆E mean = 14.11), resulting in poor SOM model sharing performance (R2 mean = 0.52 and RMSE mean = 20.33 g/kg). The use of colour calibration methods reduced the colour difference between soil images (∆E mean = 8.19) and improved the shared accuracy of the model (R2 mean = 0.61 and RMSE mean = 12.00 g/kg). The pattern of colour calibration has a key impact on the performance of the model application. The model sharing accuracy was found to be higher for the same or similar colour calibration pattern combination compared to different combinations of colour calibration patterns. Overall, the richer the colour calibration blocks, the better the model's shared performance. The findings of this research can enhance the application performance of soil attribute prediction models based on the colour of objects captured by smartphones in natural environments. • Differences in light intensity affect the colour stability of smartphone images. • Colour calibration reduces image colour differences in different light for phones. • Colour calibration boosts the shared ability of the attribute prediction model. • Colour calibration patterns with rich colours for better model-sharing ability. [ABSTRACT FROM AUTHOR]