| 摘要: | 氣候變遷影響日益加劇，其中受到衝擊之一的便是農業。因為土壤水分直接影響植物生長和農業產量，對土地生態系統的穩定性和水資源的可持續利用也都具有一定的影響性。因此若能提高對於土壤含水量變化趨勢的預測準確性，對於農業決策有很大的幫助。本研究旨在利用機器學習中的隨機森林方法來預測深度10公分和20公分的土壤含水量變化情況。以臺中霧峰農試所的歷史土壤水分數據，結合其他環境因素變數如時雨量、累積降雨量組合等，構建隨機森林模式。在模式訓練過程中，由人工窮舉找出較佳的參數組合，如訓練天數、預測天數等。此場址之預測結果顯示，累積降雨量對模式的影響最大。不論是考慮全部資料時間段、僅考慮雨季時期或是透過擬合曲線，均可以發現在深度10公分和20公分下，累積降雨期為6到8天時預測結果較準確，除了在深度10公分時僅考慮雨季時期無法得出最佳降雨天數，其餘皆可得出。考慮全部資料時間段深度10公分和20公分時選擇下降轉折點作為最佳累積天數，MAPE(%)值為25.18和5.13；僅考慮雨季時期，在深度20公分其MAPE(%)為6.63；透過擬合曲線在深度10、20公分，預測與訓練誤差皆小的條件下其預測結果之RMSE(%)值可達2.37和2.03。於未來研究中可以考慮添加更多氣象變數，或是將隨機森林模式結果與水文物理模式相比較，或者進一步探討乾旱時期的應用，以提高預測準確性，為農業應用提供更好的數據供決策者參考。;The impact of climate change is increasingly severe, particularly in the agricultural sector. Soil moisture, a key factor in plant growth and agricultural yield, also plays a significant role in the stability of land ecosystems and the sustainable use of water resources. Therefore, enhancing the accuracy of soil moisture prediction trends is crucial for informed agricultural decision-making. This study, utilizing the Random Forest method in machine learning, aims to predict soil moisture changes at depths of 10 cm and 20 cm. By leveraging historical soil moisture data from the Wufeng Agricultural Research Station in Taichung, along with other environmental variables such as hourly rainfall and cumulative rainfall, a Random Forest model was meticulously constructed. The model training process involved determining optimal parameter combinations, such as training days and prediction days, through a careful process of manual trial and error, ensuring the reliability of the study′s findings.
The prediction results for this site indicate that cumulative rainfall has the greatest impact on the model. Whether considering the entire data period, only the rainy season, or fitting a curve, it can be observed that at depths of 10 cm and 20 cm, predictions are more accurate when the cumulative rainfall period is 6 to 8 days. The exception is at a depth of 10 cm during the rainy season, where an optimal rainfall period could not be determined. When considering the entire data period at depths of 10 cm and 20 cm, choosing the inflection point of the decline as the optimal cumulative days, the MAPE (%) values are 25.18 and 5.13, respectively. During the rainy season, at a depth of 20 cm, the MAPE (%) is 6.63. At 10 cm and 20 cm depths, the prediction RMSE (%) values are 2.37 and 2.03 for the appropriate fitting range concerning the difference between the training and predicting results, respectively.
Future research could consider adding more meteorological variables, comparing the results of the Random Forest model with hydrological and physical models, or further exploring applications during drought periods to improve prediction accuracy. This would provide better data for agricultural decision-makers to reference. |
