Agricultural soil characterization and crop recommendation using deep learning algorithms: Model Selection and AI Application Development

Abstract

Agriculture is vital, and soil is a fundamental component with unique characteristics for different crops. This thesis research aimed to develop a deep learning-based system for soil type classification. It explores the performances of eight CNNs architectures namely, DenseNet201, MobileNetV3Large, VGG16, VGG19, InceptionV3, ResNet50, Xception, and a novel architecture referred to as simple_architecture, for accurately classifying agricultural soil types found in Maradi, Niger. The research methodology encompasses data collection, cleaning, preprocessing, model building, hyperparameter optimization, model compilation, and the development of an AI-based application. The findings highlight that ResNet50 and DenseNet201 were better than other models for all performance metrics. Thus, the developed application is meant to empower farmers to optimize their practices in the face of land degradation and climate change challenges.