Akademik Veri Yönetim Sistemi
ISPEC 18. ULUSLARARASI TARIM, HAYVANCILIK VE KIRSAL KALKINMA KONGRESİ, Konya, Türkiye, 24 - 26 Ekim 2025, ss.178-179, (Özet Bildiri)
In ths study, the am was to predct daly sol temperature at a depth of 5 cm usng a deep learnng model, Long Short-Term Memory (LSTM), based on meteorologcal data from the Yüreğr/Çukurova Agrcultural Research Automatc Meteorologcal Observaton Staton (staton no. 18054) for the perod between 2016 and 2025. The nput varables of the model ncluded average wnd speed, total precptaton, average temperature, maxmum temperature, mnmum temperature, relatve humdty, and daly evapotranspraton values. Mssng data n the dataset were completed usng the lnear nterpolaton method. The nput varables of the model were scaled usng mn-max normalzaton. Durng model development, varous sldng wndow szes (3, 5, 7, 10, 14, 21, and 30 days) were tested to examne ther effect on predcton accuracy. The model performance was evaluated usng Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and the Coeffcent of Determnaton (R²) metrcs. As a result of the evaluaton, the hghest accuracy was acheved usng a 7-day sldng wndow, wth MAE, RMSE, and R² values determned as 1.565 °C, 1.890 °C and 0.943 respectvely. Models wth shorter sldng wndows (3 and 5 days) yelded lower accuracy due to nsuffcent hstorcal nformaton (R² ≈ 0.885). On the other hand, n models wth longer sldng wndows (21 and 30 days), a decrease n performance was observed due to the loss of nformaton tmelness (R² ≈ 0.916–0.921). The fndngs ndcate that sol temperature responds wth a delay to short-term clmate dynamcs, and the sldng wndow sze s a determnng factor n the performance of tme seres predctng models.