TIME-AWARE MACHINE LEARNING WITH MULTI-SOURCE DATA FOR FIELD-SCALE AGRICULTURAL DECISION SUPPORT

Field scale decision support for smallholder farms in southern India requires models that respect both space and time. This study aimed to develop a time aware machine learning pipeline that fuses electromagnetic soil conductivity, laboratory soil analyses, Sentinel-2 multi-date imagery and local weather records to provide actionable management zone delineation and nutrient advisory. We process EM surveys and soil cores with ordinary kriging to create spatial rasters, generate weekly Sentinel-2 indices from 2018 to 2022, and align daily weather with crop phenology. A set of models including Random Forest, LSTM, Temporal Convolutional Network and a spatio-temporal Transformer are trained under leave-one-year-out and parcel based cross validation. The Transformer yields the best predictive performance for yield and nutrient response with mean RMSE = 0.42 t ha⁻¹ and R² = 0.78 on hold out years, improving RMSE by ~12 percent over the Random Forest baseline. Management zone stability measured by Cohen Kappa and Jaccard indices is highest for kriging + fuzzy k-means zonation and supports a 3-class mapping as the best trade-off between stability and interpretability. Practical decision outputs include per parcel N application bands and short advisories designed for field managers. The pipeline reduces the uncertainty of seasonal advisories and delivers maps and short textual guidance usable by extension services in the southern Indian context.