Gyasi Emmanuel Kwabena, Purushotham Swarnalatha
School of Computer Science and Engineering, VIT University, Vellore 632014, India.
Sensors (Basel). 2023 Jul 27;23(15):6709. doi: 10.3390/s23156709.
Scholars have classified soil to understand its complex and diverse characteristics. The current trend of precision agricultural technology demands a change in conventional soil identification methods. For example, soil color observed using Munsell color charts is subjective and lacks consistency among observers. Soil classification is essential for soil management and sustainable land utilization, thereby facilitating communication between different groups, such as farmers and pedologists. Misclassified soil can mislead processes; for example, it can hinder fertilizer delivery, affecting crop yield. On the other hand, deep learning approaches have facilitated computer vision technology, where machine-learning algorithms trained for image recognition, comparison, and pattern identification can classify soil better than or equal to human eyes. Moreover, the learning algorithm can contrast the current observation with previously examined data. In this regard, this study implements a convolutional neural network (CNN) model called Soil-MobiNet to classify soils. The Soil-MobiNet model implements the same pointwise and depthwise convolutions of the MobileNet, except the model uses the weight of the pointwise and depthwise separable convolutions plus an additional three dense layers for feature extraction. The model classified the Vellore Institute of Technology Soil (VITSoil) dataset, which is made up of 4864 soil images belonging to nine categories. The VITSoil dataset samples for Soil-MobiNet classification were collected over the Indian states and it is made up of nine major Indian soil types prepared by experts in soil science. With a training and validation accuracy of 98.47% and an average testing accuracy of 93%, Soil-MobiNet showed outstanding performance in categorizing the VITSoil dataset. In particular, the proposed Soil-MobiNet model can be used for real-time soil classification on mobile phones since the proposed system is small and portable.
学者们对土壤进行分类以了解其复杂多样的特性。精准农业技术的当前趋势要求改变传统的土壤识别方法。例如，使用孟塞尔色卡观察到的土壤颜色具有主观性，且观察者之间缺乏一致性。土壤分类对于土壤管理和可持续土地利用至关重要，从而促进不同群体（如农民和土壤学家）之间的交流。分类错误的土壤可能会误导相关过程；例如，它会阻碍肥料的施用，影响作物产量。另一方面，深度学习方法推动了计算机视觉技术的发展，经过图像识别、比较和模式识别训练的机器学习算法能够比人眼或与人眼一样好地对土壤进行分类。此外，学习算法可以将当前的观察结果与先前检查的数据进行对比。在这方面，本研究实现了一种名为Soil-MobiNet的卷积神经网络（CNN）模型来对土壤进行分类。Soil-MobiNet模型采用了与MobileNet相同的逐点卷积和深度卷积，只是该模型使用逐点和深度可分离卷积权重，再加上另外三个全连接层进行特征提取。该模型对韦洛尔理工学院土壤（VITSoil）数据集进行了分类，该数据集由属于九个类别的4864张土壤图像组成。用于Soil-MobiNet分类的VITSoil数据集样本是在印度各邦收集的，它由土壤科学专家准备的九种主要印度土壤类型组成。Soil-MobiNet在对VITSoil数据集进行分类时表现出色，训练和验证准确率为98.47%，平均测试准确率为93%。特别是，所提出的Soil-MobiNet模型可用于手机上的实时土壤分类，因为所提出的系统体积小且便于携带。
