Soil image analysis towards automated interpretation of soil properties based on Munsell soil colour guidelines using a vision language model

Soil colour is widely used as an indicator of soil’s physical and chemical properties. Traditionally, soil colour is assessed by visually matching samples to the Munsell Soil Colour Book (MSCB) to infer key soil properties. Although informative, this manual process is subjective and inconsistent, and often supplemented by laboratory analysis, which is both time-consuming and costly. Advances in smartphone imaging and vision language models (VLMs) create new opportunities for automated soil interpretation from soil imagery. This study presents a framework for automated interpretation of soil properties from soil images based on Munsell Soil colour guidelines using a fine-tuned VLM within a Visual Question Answering (VQA) framework. For a given soil image and a structured question set, the model generates concise, soil domain aligned descriptions of key soil characteristics. Among the evaluated tasks, Munsell colour prediction achieved the highest accuracy (87%), while properties such as iron state, redox condition and mineral indicators showed moderate performance. These results provide insight into which soil attributes are most visually distinctive and therefore most amenable to VLM based reasoning. Overall, the proposed framework provides a digitised and accessible approach for colour-guided soil interpretation from soil imagery, reducing reliance on subjective human assessment.