MSc

This research delves into the refinement of seasonal climate forecasts downscaled to district levels, employing state-of-the-art AI/ML methods for optimal integration in district-level agriculture applications. Initial attempts of downscaling seasonal climate forecasts using the SRCNN method provided unsatisfactory results.
Hence DL4DS, a Python module developed by Gonzalez, C.A.G.(2023) for downscaling was used. GPCP (low- resolution) & TRMM (high-resolution), ERA5 precipitation data and GFS Model output precipitation data (both low-resolution and high-resolution) along with some predictor variables were used for the study.
Downscaling was done with these three sets of data and 4 different methods were employed. The research methodology involves a thorough analysis of these datasets to understand the intricate relationship between data sources and model performance.
Promising results were achieved with the models that were trained with ERA5 precipitation and GFS model output high-resolution precipitation data, contrasting with less reliable outcomes for the models that were trained with low-resolution and high- resolution data simultaneously. The study provides valuable insights into the challenges
and successes of employing AI/ML methods for downscaling seasonal forecasts for use in agriculture. The challenges encountered with certain datasets emphasize the importance of careful consideration and selection of input data for optimal performance of the AI/ML-based models for downscaling seasonal forecasts.
Keywords: Downscaling, AI/ML, Deep learning, Seasonal forecasts, Convolutional Neural Networks