ASHS 2024 Conference

Accurate prediction of cocoa yields is critical for farmers, governments, and industry as it influences logistics and supports decision-making. In tropical agriculture, there has been a recent trend toward integrating sensor technology, data science, and machine learning to enhance management and boost crop production. However, cocoa yield prediction models rely on the quality and availability of public datasets and genetic differences among cocoa genotypes. Moreover, current cocoa models require technical skills for satellite image processing or a significant investment in sensors and software. Different statistical models in cocoa predict yield independently of physiological processes or disease pressure. Therefore, we propose a mechanistic model that uses historical yield and weather data from 2010 to 2023 and the in-field sampling from 61 orchard plots from four farms in Guayas, Ecuador. Time series measures of cocoa pods and disease incidence per tree in the plots were recorded for the 2023 and 2024 growing seasons. Cocoa pod counts and diseased pods, as well as tree photographs for biomass calculation, were recorded using a customized mobile application. Ecuador´s cocoa production in this location has a bimodal annual distribution, with the highest peak following the start of the rainy season. Moniliasis disease also presented the highest incidence within the next two months of precipitaion peaks. Several varieties of cocoa are grown in Ecuador, but production is dominated by two main groups: Fine cocoa or national flavor and CCN-51. The national type of cocoa is characterized by its unique flavor profile. However, it is prone to diseases and has a lower yield. To overcome these challenges, our study aims to develop a machine learning-based model geared towards Ecuador's distinct national type cocoa varieties, taking into account local climate patterns, soil characteristics, biomass, and direct cocoa pod field counting. The analysis reveals that cocoa yield variability is affected mainly by moniliasis disease incidence, tree biomass, and environmental factors such as temperature, solar radiation, and precipitation. In contrast, soil texture, pH, and electrical conductivity had minor variations and a negligible effect on yield changes. The proposed model was compared with other machine learning algorithms based on Mean Absolute Error and Mean Square Error. The validation phase, employing the Mean Absolute Percentage Error, indicates our model's substantial predictive accuracy with a confidence interval of 73.4 percent at the 0.1 significance level and confirms the model's effectiveness in forecasting cocoa yields under Ecuadorian conditions.