Soil Moisture Classification Using Computer Vision and Deep Learning Techniques

Soil moisture monitoring is crucial for optimizing irrigation practices and enhancing crop productivity in agriculture. Traditional methods using sensors have limitations in terms of cost, durability, and area coverage. This paper proposes a Convolutional Neural Network (CNN) based framework for soil moisture classification using a large dataset consisting of 1500 soil images divided into three classes: dry soil (500 images), moist soil (500 images), and water-logged soil (500 images). The CNN model architecture includes a repetition of multiple layers, such as convolution, rectified linear unit (ReLU), and pooling layers, to extract intricate features and classify soil moisture content accurately. Data augmentation techniques, including mirroring, random cropping, rotation, and colour shifting, were employed to increase the dataset's diversity and reduce overfitting. The model achieved an accuracy of 94.42% with a loss of 0.14 using a batch size of 32. The confusion matrix demonstrated high correct prediction rates for each soil moisture class. The model's performance was compared using different activation functions, with the Exponential Linear Unit (ELU) proving more efficient than ReLU. The proposed CNN is superior to ResNet50 and support vector machine (SVM) classifiers in terms of accuracy and speed. Future scope includes incorporating the system into edge devices for real-time automatic irrigation. It has to be noted that the proposed work is different from the ones that predict soil moisture to decide the irrigation schedule or classification of soils (red soil, black soil, alluvial soil, clay, loam etc.).