This study evaluated a novel smartphone-based soil image segmentation technique and subsequent machine learning (ML) optimization methodology with a set of soil images for rapidly predicting soil organic matter (SOM) with minimal soil processing. A smartphone and a custom-made box were used to capture images for 90 soil samples, collected from three different agroclimatic zones of West Bengal, India under three different illumination conditions. To offset the impact of variable illumination, the reflectance component of the image was recovered by removing the illumination from the image. Further, to deceive the ML model without distorting the soil image, an adversarial image was generated by adding Gaussian noise to the image. A Tree-based Pipeline Optimisation Tool was used to find an optimum ML stacking scheme using six different ML models. Model validation statistics indicated that reflectance image-extracted sub-colour space could predict SOM with reasonable accuracy (R² = 0.88, RMSE = 0.28%) using original images in stack one. Moreover, the sub-colour space using perturbed images in stack one could sense noise, worsening the model validation (R² = 0.79, RMSE = 0.36%). Conversely, seven out of eight tested colour spaces in stack two were unable to sense the image noise, producing higher validation performance than the original images. The proposed smartphone-based image acquisition setup combined with the computer vision and ML pipeline produced an important advance in affordable optical tool-based SOM prediction with significant time and cost savings. More research is warranted to extend this approach by incorporating field images of variable soil types taken under variable illuminations.

本研究评估了一种新的基于智能手机的土壤图像分割技术和随后的机器学习 (ML) 优化方法，其中包含一组土壤图像，用于以最少的土壤处理快速预测土壤有机质 (SOM)。使用智能手机和定制​​盒子捕获 90 个土壤样本的图像，这些样本是在三种不同照明条件下从印度西孟加拉邦的三个不同农业气候区收集的。为了抵消可变照明的影响，通过从图像中去除照明来恢复图像的反射分量。此外，为了在不扭曲土壤图像的情况下欺骗 ML 模型，通过向图像添加高斯噪声来生成对抗性图像。基于树的管道优化工具用于使用六种不同的 ML 模型找到最佳的 ML 堆叠方案。²  = 0.88，RMSE = 0.28%）使用堆栈一中的原始图像。此外，在堆栈 1 中使用扰动图像的子颜色空间可以感知噪声，使模型验证恶化（R ²  = 0.79，RMSE = 0.36%）。相反，堆栈 2 中八个测试颜色空间中有七个无法感知图像噪声，产生比原始图像更高的验证性能。提议的基于智能手机的图像采集设置与计算机视觉和 ML 管道相结合，在经济实惠的基于光学工具的 SOM 预测方面取得了重要进展，并显着节省了时间和成本。需要更多的研究来扩展这种方法，通过结合在可变光照下拍摄的可变土壤类型的现场图像。