Site-specific management zones delineation and Yield prediction for rice based cropping system using on-farm data sets in Tolima (Colombia)

In the valley of the Alto Magdalena, Colombia, intensive agriculture and inefficient soil and water management techniques have generated a within field yield spatial variability, which have increased the production costs for the rice-based cropping system (rice, cotton and maize crops rotation field). Crop yield variations depend on the interaction between climate, soil, topography and management, and it is strongly influenced by the spatial and temporal availabilities of water and nutrients in the soil during the crop growth season. Understanding why the yield in certain portions of a field has a high variability is of paramount importance both from an economic and an environmental point of view, as it is through the better management of these areas that we can improve yields or reduce input costs and environmental impact. The aim of this study was 1) to predict rice yield using on farm data set and machine learning and 2) to compare delimited management zones (MZ) for rice-based cropping system with physiological parameters and within field variation yield.

A 72 sampling points spatially distributed were defined in a 5 hectares plot at the research center Nataima, Agrosavia. For each sampling point, physical and chemical properties, biomass and relative chlorophyll content were determined at different vegetative stages. A multispectral camera mounted to an Unmanned Aerial Vehicle (UAV) was used to acquire multispectral images over the rice canopy in order to estimate vegetation indices. Five nonlinear models and two multilinear algorithms were employed to estimate rice yield. The fuzzy cluster analysis algorithm was used to classify soil data into two to six MZ. The appropriate number of MZ was determined according to the results of a fuzziness performance index and normalized classification entropy.

Results of the rice yield prediction model showed that the best performance was obtained by K-Nearest Neighbors (KNN) regression algorithm with an average absolute error of 10.74%. Nonetheless, the performance of the other algorithms was acceptable except the Multiple Linear regression (MLR). The MLR showed the highest RMSE with 2712.26 kg.ha^-1 in the testing dataset, while KNN regression was the best with 1029.69 kg.ha^-1. These findings show the importance of machine learning could have for supporting decisions in agriculture processes management.

The cluster analyses revealed that two zones was the optimal number of classes based on different criteria. Delineated zones were evaluated and revealed significant differences (p≤0.05) in sand, apparent density, total porosity, pH, organic matter, phosphorus, calcium, magnesium, iron, zinc, cover and boron. The relative chlorophyll content of cotton and maize crops showed a similar spatial distribution pattern to delimited MZ. The results demonstrate the ability of the proposed procedure to delineate a farmer’s field into zones based on spatially varying soil and crop properties that should be considered for irrigation and fertilization management.