小型氣象盒子結合資料探勘方法應用於短期降雨預測

Abstract

一些農作物，如：玉米、柿子或藥材，傳統上藉由陽光曝曬降低含水率以延長儲存時間，此時，若發生突發性的降雨將會導致農作物受損，造成經濟上的損失。降雨亦可能影響某些果樹花朵的授粉成功率，影響後續的收成。因此，更加精準的降雨預測將有助於改善作物的產量與品質。 本研究開發的小型氣象盒子可蒐集溫度、濕度、大氣壓力、光照度與雨滴感測等環境資訊，具有資料儲存與傳送降雨即時提醒功能。系統設置於台灣大學與桃園市，進行資料收集與耐久性測試。 建立降雨預測模型的歷史數據有三種來源，分別為：1. 台大與桃園市兩處氣象盒子的持續蒐集 2. 宜蘭大學於上將梨果園設置的監測系統以及 3. 中央氣象局公開的歷史數據。針對歷史資料使用了資料探勘技術，監督式學習二元分類方法，用以建立目前氣象狀態與特定時間後降雨情形的關係。使用Decision Tree、Bagged Tree、KNN、SVM四種演算法，與10-fold cross validation的驗證方法。結果顯示，氣象盒子的資料由KNN建立的模型預測效果最佳，氣象局資料則為Bagged Tree之效果最佳。

Rainfall is important for food production plan, water resource management and all activity plans in the nature. The occurrence of prolonged dry period or heavy rain at the critical stages of the crop growth and development may lead to significant reduce crop. For tea farmers and fruit farmers, under proper soil conditions and management. The quality and yield of crops are more affected by temperature and precipitation. It can be very helpful to reduce the effects of sudden and extreme rainfall by taking effective security measures in advance. Therefore, Rainfall prediction is very important for farmers. This research has developed a small weather box that collects environmental information such as temperature, humidity, atmospheric pressure, illuminance and raindrop. The purpose is to monitor the agricultural environment and collect environmental information from local areas. The study set up the weather box in National Taiwan University (NTU), Taoyuan city, Yilan city as experimental fields. The analysis of the experiment was carried out in two ways: 1. Weather box 2. Central Weather Bureau (CWB). The dataset was obtained from Central Weather Bureau (CWB) and consists of several atmospheric attributes, and data mining techniques was used. Data mining techniques can effectively predict the rainfall by extracting the hidden patterns among available features of past weather data. This research performed rainfall prediction in Taipei city using five classifier : Support Vector Machine, K Nearest Neighbor, Decision Tree and Bagged Tree. According to results, the prediction model performed well for no-rain class however for rain class wasn’t well. The results show that the model of the weather box is the best for KNN, and the data of the weather bureau is the best for the Bagged Tree. The CWB trained model predicted the rainfall accuracy was generally poor, the improvement is made by adjusting the sample proportion. The experimental results show that KNN and SVM have the most performance improvement under the equilibrium ratio.