Soil Moisture Detection for Precision Agriculture using Deep Learning Techniques

Soil moisture detection is vital for optimised crop growth and agricultural output. The conventional manual testing techniques are challenging and time-consuming. Recent advancements in computer vision and deep learning can reduce human effort significantly. The proposed methodology bridges the gap between conventional practices and modern development by using convolutional models, specifically InceptionV3, VGG16, and MobileNetV2, with some additional layers to the architecture for our specific use case. Additionally, deep learning techniques can detect real-time moisture in dry and wet soils based on visual features, such as colour, brightness, and texture variations. A novel dataset comprising 1318 images of dry and wet soils created under different lighting conditions, terrains and soil types to cover the different scenarios. The images were preprocessed, augmented and split into training and validation sets, while test sets were taken from the original dataset. The training, validation and test accuracies for the InceptionV3 model achieved $99.88 %, 94.32 %$, and $94.31 %$, respectively. However, the VGG16 model achieved $99.76 %, 92.41 %$ and $90.53 %$ and MobileNetV2 obtained $99.76 %, 95.26 %$ and $93.56 %$, respectively, for the same metrics. The above results depict that the InceptionV3 model could outperform the VGG16 and MobileNetV2 models in classifying soil moisture content, and this leads to the path for intelligent farming systems that maximise yields while conserving our planet's finite soil and water assets.