DSSAT with Differential Evolution: A Receding Horizon Approach for Online Fertilizer Optimization

Achieving sustainable agriculture requires balancing economic viability with environmental responsibility, particularly in fertilizer management. The Decision Support System for Agrotechnology Transfer (DSSAT) is a software application program widely used among researchers and professionals due to its accuracy and flexibility in simulating crop growth, development, and final yield across diverse soil-plant-atmosphere dynamics. It contains the models for many different crops, and a suite of utilities for soil, weather, crop management, and experimental data management. Despite its strengths, the DSSAT crop simulation model exhibits certain limitations. Its whole-season simulation approach precludes real-time adjustments throughout the growing season, due, for example, to unexpected weather events or inaccurate forecasts. Furthermore, its automatic management tool is limited to handling automatically only irrigation, sowing, and harvesting, and does not provide the option to define the fertilization strategy. Furthermore, the automatic irrigation, for instance, is threshold-based, so it is not designed for irrigation optimization. This thesis specifically addresses these problems by designing a simulation strategy based on a receding horizon variation of a differential evolution optimization algorithm on top of the DSSAT crop simulation model. This also includes a weather forecast in the loop to make the model capable of defining a (sub-)optimal online strategy for the fertilizer application. In this way, the DSSAT could be used as an online decision-support tool, with the objective of maximizing the yield given a fertilizer budget. The algorithm is supposed to be robust against weather variations and the choice of fertilizer application days. First, the optimization algorithm is applied and tested with real seasonal weather. This phase is made to test the algorithm and evaluate its performance. After that, the hypothesis of knowing the weather is reduced and the optimization is refined throughout the season. The evaluation of the results is not trivial, due to the high dependence of the growth to many aspects that are difficult to reproduce. This is in fact a challenge for all this kind of research. The results are compared to an extensive simulation campaign with real weather data, and the shape of the treatments is compared to some suggestions derived from research papers or equivalent documentation. Preliminary findings demonstrate the algorithm's ability to identify patterns consistent with paper suggestions, achieving performance similar to the simulated maximum.