Seasonal and inter-annual climate variability have become important issues for farmers, and climate change has been shown to increase them. Simultaneously farmers and agricultural organizations are increasingly collecting observational data about in situ crop performance. Agriculture thus needs new tools to cope with changing environmental conditions and to take advantage of these data. Data mining techniques make it possible to extract embedded knowledge associated with farmer experiences from these large observational datasets in order to identify best practices for adapting to climate variability. We introduce new approaches through a case study on irrigated and rainfed rice in Colombia. Preexisting observational datasets of commercial harvest records were combined with in situ daily weather series. Using Conditional Inference Forest and clustering techniques, we assessed the relationships between climatic factors and crop yield variability at the local scale for specific cultivars and growth stages. The analysis showed clear relationships in the various location-cultivar combinations, with climatic factors explaining 6 to 46% of spatiotemporal variability in yield, and with crop responses to weather being non-linear and cultivar-specific. Climatic factors affected cultivars differently during each stage of development. For instance, one cultivar was affected by high nighttime temperatures in the reproductive stage but responded positively to accumulated solar radiation during the ripening stage. Another was affected by high nighttime temperatures during both the vegetative and reproductive stages. Clustering of the weather patterns corresponding to individual cropping events revealed different groups of weather patterns for irrigated and rainfed systems with contrasting yield levels. Best-suited cultivars were identified for some weather patterns, making weather-site-specific recommendations possible. This study illustrates the potential of data mining for adding value to existing observational data in agriculture by allowing embedded knowledge to be quickly leveraged. It generates site-specific information on cultivar response to climatic factors and supports on-farm management decisions for adaptation to climate variability.

Agriculture research uses “recommendation domains” to develop and transfer crop management practices adapted to specific contexts. The scale of recommendation domains is large when compared to individual production sites and often encompasses less environmental variation than farmers manage. Farmers constantly observe crop response to management practices at a field scale. These observations are of little use for other farms if the site and the weather are not described. The value of information obtained from farmers’ experiences and controlled experiments is enhanced when the circumstances under which it was generated are characterized within the conceptual framework of a recommendation domain, this latter defined by Non-Controllable Factors (NCFs). Controllable Factors (CFs) refer to those which farmers manage. Using a combination of expert guidance and a multi-stage analytic process, we evaluated the interplay of CFs and NCFs on plantain productivity in farmers’ fields. Data were obtained from multiple sources, including farmers. Experts identified candidate variables likely to influence yields. The influence of the candidate variables on yields was tested through conditional forests analysis. Factor analysis then clustered harvests produced under similar NCFs, into Homologous Events (HEs). The relationship between NCFs, CFs and productivity in intercropped plantain were analyzed with mixed models. Inclusion of HEs increased the explanatory power of models. Low median yields in monocropping coupled with the occasional high yields within most HEs indicated that most of these farmers were not using practices that exploited the yield potential of those HEs. Varieties grown by farmers were associated with particular HEs. This indicates that farmers do adapt their management to the particular conditions of their HEs. Our observations confirm that the definition of HEs as recommendation domains at a small-scale is valid, and that the effectiveness of distinct management practices for specific micro-recommendation domains can be identified with the methodologies developed.

Wrapper methods are a type of feature selection method that finds a subset of variables to improve the performance of a classifier by removing redundant and irrelevant variables. The use of a wrapper implies that each time a candidate solution is explored, the classifier is evaluated on the quality measures selected (e.g. accuracy or precision). Though robust, this iteration across several candidate solutions can become computationally intensive and time-consuming. In this paper we propose a wrapper, that is based on binary Covering Arrays (CAs), and binary Incremental Covering Arrays (ICAs), that have been widely used for experimental design and fault detection in software and hardware testing. The new wrapper was evaluated with six classifiers on seven data sets. The results show that the CAs and ICAs with strength 6 significantly improve the performance and reduces the number of variables required by the classifier. A comparative analysis of the proposed method against wrappers based on other search approaches such as genetic algorithms (GA) and particle swarm optimization (PSO), shows that the proposed method yields results similar to GA, but not to PSO, with differences to PSO, in accuracy, which in the majority of cases is below 0.04. This lack of accuracy, by which the new wrapper fails to match PSO, is offset by the fact that the user does not need to fine tune algorithm parameters, such as velocity ranges, timing, cognitive coefficient, and social coefficient, while it is also much easier to program in parallel.

For several decades, maize (Zea mays L.) management decisions in smallholder farming in tropical regions have been a puzzle. To best balance alternative management practices' environmental and economic outcomes, an extensive dataset was gathered through CIMMYT's knowledge hub in Chiapas, a state in southern Mexico. In a knowledge hub, farmers, with the support of farm advisors, compare conventional and improved agronomic practices side-by-side and install demonstration fields where they implement improved practices. In all these fields data on on-farm operations and results is collected. The dataset was assembled using field variables (yield, cultivars, fertilization and tillage practice), as well as environment variables from soil mapping (slope, elevation, soil texture, pH and organic matter concentration) and gridded weather datasets (precipitation, temperature, radiation and evapotranspiration). The dataset contains observations from 4585 fields and comprises a period of 7 years between 2012 and 2018. This dataset will facilitate analytical approaches to represent spatial and temporal variability of alternative crop management decisions based on observational data and explain model-generated predictions for maize in Chiapas, Mexico. In addition, this data can serve as an example for similar efforts in Big Data in Agriculture.