Early Paddy Yield Forecasting in Sri Lanka Using SARIMA and LSTM for Farmer and Collector Decision Support

Early paddy-yield information can support cultivation and procurement planning in Sri Lanka, where farmers and collectors make decisions before final harvest outcomes are known. This study developed and evaluated a web-based decision-support system (DSS) that integrates district-level paddy-yield forecasting with separate farmer and collector workflows. Historical district-seasonal records from 2004–2023 covered 21 retained districts and 797 observations. District-specific Seasonal Autoregressive Integrated Moving Average (SARIMA) and Long Short-Term Memory (LSTM) models were compared using an archived final-four-observation holdout. Mean Absolute Percentage Error (MAPE), averaged across districts, was 22.06% for SARIMA and 17.38% for LSTM, with LSTM recording lower MAPE in 12 of 21 districts; a paired district comparison reported *p* = 0.0333. However, chronology, scaling, and unequal forecast-information horizons limit this archived comparison, so the result is not interpreted as validated LSTM superiority. A problem-validation survey of 122 participants (103 farmers and 19 collectors) identified reported planning uncertainty and strong perceived need for earlier forecasts. The DSS was implemented using FastAPI and React/Vite with English and Sinhala presentation. A post-demonstration evaluation with 34 intended users produced a mean System Usability Scale score of 73.53. A shared planning-uncertainty item did not differ significantly from the independent baseline sample. The contribution is therefore an integrated and bounded forecasting-DSS prototype with preliminary stakeholder evidence and explicitly stated validation limitations.