In Zambia, smallholder fish farmers need access to climate services for fish farming. Managing climate-driven environmental factors and stressors, such as temperature and water supply, is challenging with limited resources and technical skills and yet critical for smallholder productivity. This paper presents the results of an exploratory analysis of environmental data from smallholder fish farmers in Zambia to develop a predictive model for tilapia fish farmers. The model variables included time of day, air maximum temperature, air minimum temperature and pond temperature as the response variable .
Five models were compared, including linear regression, stochastic regression, deep learning, random forest, and Decision tree. The data was modelled for a pond designed according to best aquaculture practices, and therefore, pond size and pond depth are constants. A linear regression model for pond temperature is: Y(t) = αX1 + βX2 + C + ϒ, where α and β are the gradients (coefficients) of the maximum and minimum air temperature features, respectively. Additionally, C denotes the intercept and ϒ is the err coeff. Y(t) is the predicted pond temperature in degrees centigrade at time t. The R-squared value for the linear regression was 0.6, hence the linear regression was a good fit. The seasonal pattern of pond temperatures and predicted pond temperatures using the linear regression model displayed a close resemblance.
Three machine learning models (deep learning, decision tree, and random forest) were also compared. The decision tree model is most applicable and yielded good model fit results and a close resemblance of predicted pond temperature to actual pond temperature. It is one way to display an algorithm that only contains conditional control statements. The performance evaluation results of the decision-tree model displayed a R-squared value of 0.6 which is lesser than that of deep learning model (i.e. 0.85). The model did, however, display high performance metrics.
The decision tree model is a good foundation for farmers to understand and predict pond temperatures on their farms. This model is flexible and can be refined for specific operations so that farmers can update scenarios, actions and their predicted outcomes. The decision tree model will be refined over time as it currently incorporates discrete pond temperature measurements for the morning and evening. Additionally, modelling the return of temperature to optimal ranges through water replacement actions would be desirable. For this, pond temperature measurements need to be linked to the rate of water replacement (using flow meters) and the temperature of the source water. With continuous data of pond temperatures over 24 h-cycles, it may be possible to refine models to predict pond temperatures one day in advance.