WEATHER LORE VALIDATION TOOL USING FUZZY COGNITIVE MAPS BASED ON COMPUTER VISION

Abstract

The creation of scientific weather forecasts is troubled by many technological challenges (Stern & Easterling, 1999) while their utilization is generally dismal. Consequently, the majority of small-scale farmers in Africa continue to consult some forms of weather lore to reach various cropping decisions (Baliscan, 2001). Weather lore is a body of informal folklore (Enock, 2013), associated with the prediction of the weather, and based on indigenous knowledge and human observation of the environment. As such, it tends to be more holistic, and more localized to the farmers’ context. However, weather lore has limitations; for instance, it has an inability to offer forecasts beyond a season. Different types of weather lore exist, utilizing almost all available human senses (feel, smell, sight and hearing). Out of all the types of weather lore in existence, it is the visual or observed weather lore that is mostly used by indigenous societies, to come up with weather predictions. On the other hand, meteorologists continue to treat this knowledge as superstition, partly because there is no means to scientifically evaluate and validate it. The visualization and characterization of visual sky objects (such as moon, clouds, stars, and rainbows) in forecasting weather are significant subjects of research. To realize the integration of visual weather lore in modern weather forecasting systems, there is a need to represent and scientifically substantiate this form of knowledge. This research was aimed at developing a method for verifying visual weather lore that is used by traditional communities to predict weather conditions. To realize this verification, fuzzy cognitive mapping was used to model and represent causal relationships between selected visual weather lore concepts and weather conditions. The traditional knowledge used to produce these maps was attained through case studies of two communities (in Kenya and South Africa).These case studies were aimed at understanding the weather lore domain as well as the causal effects between metrological and visual weather lore. In this study, common astronomical weather lore factors related to cloud physics were identified as: bright stars, dispersed clouds, dry weather, dull stars, feathery clouds, gathering clouds, grey clouds, high clouds, layered clouds, low clouds, stars, medium clouds, and rounded clouds. Relationships between the concepts were also identified and formally represented using fuzzy cognitive maps. On implementing the verification tool, machine vision was used to recognize sky objects captured using a sky camera, while pattern recognition was employed in benchmarking and scoring the objects. A wireless weather station was used to capture real-time weather parameters. The visualization tool was then designed and realized in a form of software artefact, which integrated both computer vision and fuzzy cognitive mapping for experimenting visual weather lore, and verification using various statistical forecast skills and metrics. The tool consists of four main sub-components: (1) Machine vision that recognizes sky objects using support vector machine classifiers using shape-based feature descriptors; (2) Pattern recognition–to benchmark and score objects using pixel orientations, Euclidean distance, canny and grey-level concurrence matrix; (3) Fuzzy cognitive mapping that was used to represent knowledge (i.e. active hebbian learning algorithm was used to learn until convergence); and (4) A statistical computing component was used for verifications and forecast skills including brier score and contingency tables for deterministic forecasts. Rigorous evaluation of the verification tool was carried out using independent (not used in the training and testing phases) real-time images from Bloemfontein, South Africa, and Voi-Kenya. The real-time images were captured using a sky camera with GPS location services. The results of the implementation were tested for the selected weather conditions (for example, rain, heat, cold, and dry conditions), and found to be acceptable (the verified prediction accuracies were over 80%). The recommendation in this study is to apply the implemented method for processing tasks, towards verifying all other types of visual weather lore. In addition, the use of the method developed also requires the implementation of modules for processing and verifying other types of weather lore, such as sounds, and symbols of nature. Since time immemorial, from Australia to Asia, Africa to Latin America, local communities have continued to rely on weather lore observations to predict seasonal weather as well as its effects on their livelihoods (Alcock, 2014). This is mainly based on many years of personal experiences in observing weather conditions. However, when it comes to predictions for longer lead-times (i.e. over a season), weather lore is uncertain (Hornidge & Antweiler, 2012). This uncertainty has partly contributed to the current status where meteorologists and other scientists continue to treat weather lore as superstition (United-Nations, 2004), and not capable of predicting weather. One of the problems in testing the confidence in weather lore in predicting weather is due to wide varieties of weather lore that are found in the details of indigenous sayings, which are tightly coupled to locality and pattern variations(Oviedo et al., 2008). This traditional knowledge is entrenched within the day-to-day socio-economic activities of the communities using it and is not globally available for comparison and validation (Huntington, Callaghan, Fox, & Krupnik, 2004). Further, this knowledge is based on local experience that lacks benchmarking techniques; so that harmonizing and integrating it within the science-based weather forecasting systems is a daunting task (Hornidge & Antweiler, 2012). It is partly for this reason that the question of validation of weather lore has not yet been substantially investigated. Sufficient expanded processes of gathering weather observations, combined with comparison and validation, can produce some useful information. Since forecasting weather accurately is a challenge even with the latest supercomputers (BBC News Magazine, 2013), validated weather lore can be useful if it is incorporated into modern weather prediction systems. Validation of traditional knowledge is a necessary step in the management of building integrated knowledge-based systems. Traditional knowledge incorporated into knowledge-based systems has to be verified for enhancing systems’ reliability. Weather lore knowledge exists in different forms as identified by traditional communities; hence it needs to be tied together for comparison and validation. The development of a weather lore validation tool that can integrate a framework for acquiring weather data and methods of representing the weather lore in verifiable forms can be a significant step in the validation of weather lore against actual weather records using conventional weather-observing instruments. The success of validating weather lore could stimulate the opportunity for integrating acceptable weather lore with modern systems of weather prediction to improve actionable information for decision making that relies on seasonal weather prediction. In this study a hybrid method is developed that includes computer vision and fuzzy cognitive mapping techniques for verifying visual weather lore. The verification tool was designed with forecasting based on mimicking visual perception, and fuzzy thinking based on the cognitive knowledge of humans. The method provides meaning to humanly perceivable sky objects so that computers can understand, interpret, and approximate visual weather outcomes. Questionnaires were administered in two case study locations (KwaZulu-Natal province in South Africa, and Taita-Taveta County in Kenya), between the months of March and July 2015. The two case studies were conducted by interviewing respondents on how visual astronomical and meteorological weather concepts cause weather outcomes. The two case studies were used to identify causal effects of visual astronomical and meteorological objects to weather conditions. This was followed by finding variations and comparisons, between the visual weather lore knowledge in the two case studies. The results from the two case studies were aggregated in terms of seasonal knowledge. The causal links between visual weather concepts were investigated using these two case studies; results were compared and aggregated to build up common knowledge. The joint averages of the majority of responses from the case studies were determined for each set of interacting concepts. The modelling of the weather lore verification tool consists of input, processing components and output. The input data to the system are sky image scenes and actual weather observations from wireless weather sensors. The image recognition component performs three sub-tasks, including: detection of objects (concepts) from image scenes, extraction of detected objects, and approximation of the presence of the concepts by comparing extracted objects to ideal objects. The prediction process involves the use of approximated concepts generated in the recognition component to simulate scenarios using the knowledge represented in the fuzzy cognitive maps. The verification component evaluates the variation between the predictions and actual weather observations to determine prediction errors and accuracy. To evaluate the tool, daily system simulations were run to predict and record probabilities of weather outcomes (i.e. rain, heat index/hotness, dry, cold index). Weather observations were captured periodically using a wireless weather station. This process was repeated several times until there was sufficient data to use for the verification process. To match the range of the predicted weather outcomes, the actual weather observations (measurement) were transformed and normalized to a range [0, 1].In the verification process, comparisons were made between the actual observations and weather outcome prediction values by computing residuals (error values) from the observations. The error values and the squared error were used to compute the Mean Squared Error (MSE), and the Root Mean Squared Error (RMSE), for each predicted weather outcome. Finally, the validity of the visual weather lore verification model was assessed using data from a different geographical location. Actual data in the form of daily sky scenes and weather parameters were acquired from Voi, Kenya, from December 2015 to January 2016.The results on the use of hybrid techniques for verification of weather lore is expected to provide an incentive in integrating indigenous knowledge on weather with modern numerical weather prediction systems for accurate and downscaled weather forecasts.