Soil image classification and segmentation: A survey from deep learning, multimodal data and hybrid models

This survey explores recent advances in the automated analysis of soil images, with a focus on classification and segmentation techniques driven by machine learning and deep learning. The paper provides a comprehensive overview of methods ranging from traditional approaches, such as handcrafted descriptors and statistical classifiers to state-of-the-art neural architectures, including Convolutional Neural Networks, transformers and hybrid multimodal models. While early methods relied heavily on texture and color features, modern deep learning approaches demonstrate enhanced performance through end-to-end learning and the ability to capture complex spatial patterns in soil images. The survey distinguishes between different classification paradigms : object detection, pixel-wise segmentation and general category classification. It emphasizes their relevance for soil analysis tasks such as identifying soil types, mineral compositions or predicting soil properties from RGB or spectral data. Self-supervised and transfer learning techniques are highlighted as promising solutions to the challenge of limited annotated datasets. Beyond classification, the paper discusses the critical role of segmentation in analyzing soil grain morphology, granulometry and spatial distribution. Deep learning-based segmentation methods offer significant improvements over traditional image processing techniques, particularly in heterogeneous or occluded conditions. These techniques enable more accurate particle size estimation and detailed soil texture analysis, information essential for geotechnical, hydrological and environmental applications. Increasingly, hybrid, multimodal and segmentation-aware approaches are being explored, not as an endpoint, but as the next frontier in addressing current analytical limitations to enhance robustness and generalization.