Precision agriculture is an important approach in modern farming that increases crop productivity, reduces fertilizer waste, and maintains soil health. Nitrogen, Phosphorus, Potassium, pH, and organic matter are essential for plant growth. Traditional lab testing is accurate but expensive, time-consuming, and not accessible to every farmer. Therefore, a fast and cost-effective intelligent system is needed for soil analysis. In this project, a Soil Nutrients Detector was developed using deep learning and computer vision to predict nutrients based on soil images. The dataset includes three soil types: Loam, Sandy Loam, and Clay Loam. Multiple images of each sample were taken from different angles to better capture the texture. Along with this, pH, N, P, K, and organic matter values were collected from the lab and linked to the corresponding images. Before training, images were resized and normalized, and augmented using rotation and flipping. The data was divided into training, validation, and test sets. Initially, different CNN architectures were tested; they showed good training accuracy but overfitting on validation data. A Vision Transformer was also tried but underfitted due to limited data. The best result came from transfer learning. Fine-tuning was done using VGG16, which is pre-trained on ImageNet. This allowed the model to benefit from pre-learned features of textures and shapes. VGG16 achieved the highest validation accuracy, better generalization, and more stable
Tools: Soil nutrients detection, Deep Learning, Machine Learning, Computer Vision, Image Processing, Soil classification, CNN (Convolutional Neural Network), VGG16
Department: Department of Mathematics
Poster
