Analisis Fenologi dan Prediksi Water Stress Lahan Tebu Menggunakan Random Forest Regressor dengan Data Multi Indeks Vegetasi

Abstract

Tanaman tebu (Saccharum spp.) merupakan komoditas strategis di Indonesia yang rentan terhadap cekaman air, khususnya pada lahan tadah hujan. Penelitian ini menganalisis dinamika fenologi dan mengembangkan model prediksi Crop Water Stress Index (CWSI) berbasis integrasi multi-indeks vegetasi Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDWI), Normalized Difference Drought Index (NDDI), Land Surface Water Index (LSWI), Optimized Soil Adjusted Vegetation Index (OSAVI) dengan variabel cuaca harian menggunakan algoritma Random Forest Regressor (RFR). Data berasal dari citra Landsat 8 dan data meteorologis BMKG periode 2021–2023, melalui tahapan pra- pemrosesan, pemulusan harmonik, serta konstruksi fitur berbasis lag time dan rolling mean. Evaluasi sepuluh skema kombinasi fitur menunjukkan kinerja terbaik pada model gabungan indeks vegetasi dan cuaca dengan lag_cwsi tanpa Land Surface Temperature (LST), menghasilkan R² sebesar 91,01% dan MAPE 9,08%. Meski akurasi turun ±6% dibanding model berbasis LST, pendekatan ini tetap efektif dan praktis mengingat keterbatasan data LST di wilayah tropis. Analisis SHapley Additive exPlanations (SHAP) mengidentifikasi fitur penting LSWI, NDWI, NDDI, NDVI, OSAVI, curah hujan, lama penyinaran, dan kelembapan relatif sebagai kontributor utama. Fitur lag dan rolling mean memperkuat prediksi pola water stres musiman pada lahan tebu, sehingga temuan ini berpotensi mendukung keputusan agronomis dan sistem pemantauan stres air berbasis spasial-temporal.

Sugarcane (Saccharum spp.) is a strategic commodity in Indonesia that is vulnerable to water stress, particularly on rainfed land. This study analyzes phenological dynamics and develops a Crop Water Stress Index (CWSI) prediction model based on the integration of multi-vegetation indices: Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDWI), Normalized Difference Drought Index (NDDI), Land Surface Water Index (LSWI), Optimized Soil Adjusted Vegetation Index (OSAVI), and daily weather variables using the Random Forest Regressor (RFR) algorithm. Data comes from Landsat 8 imagery and BMKG meteorological data for the 2021–2023 period, through preprocessing, harmonic smoothing, and feature construction based on lag time and rolling mean. Evaluation of ten feature combination schemes demonstrated the best performance in the combined vegetation and climate index model with lag_cwsi without Land Surface Temperature (LST), resulting in an R² of 91.01% and a MAPE of 9.08%. Although accuracy decreased by ±6% compared to LST-based models, this approach remains effective and practical given the limited LST data available in tropical regions. SHapley Additive exPlanations (SHAP) analysis identified key features: LSWI, NDWI, NDDI, NDVI, OSAVI, rainfall, solar radiation duration, and relative humidity as key contributors. The lag and rolling mean features strengthen the prediction of seasonal water stress patterns in sugarcane fields, potentially supporting agronomic decisions and spatio-temporal water stress monitoring systems.