Enhancing Spatiotemporal Data Harmonization Through Semantic Enrichment Using SciBERT LLM for Soil Analytic

This research introduces an innovative methodology to enhance spatiotemporal data harmonization in soil analytics by leveraging SciBERT [1], an advanced pre-trained large language model specialized in scientific text comprehension. The approach capitalizes on SciBERT's detailed contextual understanding to extract essential scientific concepts from accompanying textual descriptions in spatiotemporal datasets, addressing challenges stemming from limited annotated data. The proposed standardized representation of spatial and temporal dimensions, achieved through semantic enrichment and entity recognition, leads to significant improvements in predicting, aligning, and integrating spatiotemporal information. The methodology involves systematic data gathering on diverse soil types in India, including nutrient content and physical properties, creating a comprehensive fertilizer database through web scraping, preprocessing, and employing NLTK for lemmatization and tokenization. The finetuned SciBERT model is currently employed for manual querying processes, with future plans to integrate this into semantic enrichment. This research explores the application of semantic enrichment using SciBERT to address challenges in soil analytics, for improving spatiotemporal harmonization in soil data.