Pemodelan Prediksi Kerawanan Kebakaran Lahan Gambut berdasarkan Tinggi Muka Air dengan Pendekatan Machine Learning

Abstract

Kebakaran lahan gambut di Asia Tenggara merupakan bencana berulang yang semakin parah pada musim kemarau, diperparah oleh fenomena El Niño yang meningkatkan kekeringan, terutama di Provinsi Jambi, Indonesia. Dampak kebakaran gambut meliputi degradasi lingkungan, hilangnya habitat, kerusakan ekosistem dan peningkatan emisi akibat kabut asap, sehingga pemantauan kerawanan dan mitigasi menjadi sangat penting. Pemantauan berbasis tinggi muka air, terutama melalui teknologi Internet of Things (IoT), telah terbukti efektif untuk menilai kelembapan tanah secara real-time dan mendukung peringatan dini risiko kebakaran. Penelitian ini bertujuan untuk membangun model prediksi kerawanan berdasarkan tinggi muka air dan mengategorikan kerawanan kebakaran berbasis data lingkungan dengan pendekatan machine learning, mengevaluasi performa model, dan menerapkan model pada wilayah lain. Data dikumpulkan dari stasiun pemantauan berbasis Internet of Things (IoT) yang merekam delapan peubah lingkungan setiap 15 menit selama 3 bulan di Desa Pematang Rahim. Peubah prediktor yang diamati yaitu curah hujan, kelembapan udara, suhu, intensitas cahaya, kelembapan tanah, arah angin dan kecepatan angin dengan target peubah tinggi muka air. Secara ringkas, model dinyatakan dalam bentuk aturan IF–THEN dominan : (1) IF SM < 28 % THEN Y = -75 cm; (2) IF 28 = SM = 62 % AND RF = 1,9 mm THEN -41 = Y = -22 cm; dan (3) IF SM = 62 % AND = 23 TEM = THEN Y = -14 cm. Hasil prediksi menunjukkan nilai performa dengan nilai R² 0,93, RMSE 10,42, MAE 5,98, dan MAPE 0,76 % yang menunjukkan model yang tidak bias. Selain itu, simulasi jumlah peubah dan waktu data menunjukkan bahwa memungkinkan reduksi peubah prediktor menjadi tiga peubah utama, yaitu kelembapan tanah, suhu, dan curah hujan dengan waktu terbaik selama 3 bulan. Selanjutnya, klasifikasi kerawanan dibagi berdasarkan ambang batas kritis tinggi muka air < -40 cm untuk menentukan kategori kerawanan "rendah" dan "tinggi". Hasil klasifikasi mencapai akurasi 97%. Model kemudian diterapkan pada data Desa Teluk Dawan. Hasil penerapan model menunjukkan bahwa model belum mampu diterapkan di daerah lain, sehingga perlu dilakukan pelatihan data agar model dapat menyesuaikan peubah lingkungan di wilayah lain. Klasifikasi kerawanan pada data baru dengan target titik panas menunjukkan bahwa hari tanpa hujan, kelembapan tanah, tinggi muka air, dan suhu tanah secara berurutan mempengaruhi kerawanan kebakaran. Hasil penelitian ini juga memberi rekomendasi untuk menyeleksi peubah lingkungan dalam menduga tinggi muka air menjadi tiga, yaitu kelembapan tanah, suhu, dan curah hujan sebagai strategi monitoring pencegahan kebakaran hutan dan lahan gambut yang lebih efektif.

Peatland fires in Southeast Asia are recurring disasters that become increasingly severe during the dry season and are exacerbated by El Niño, which intensifies drought and expands affected areas, particularly in Jambi Province, Indonesia. The impacts of peat fires include environmental degradation, habitat loss, ecosystem damage, and increased haze-related emissions, underscoring the need for vulnerability monitoring and mitigation. Groundwater level-based monitoring, particularly through the Internet of Things (IoT), has proven effective for assessing soil moisture in real time and for supporting early warning of fire risk. This study aims to develop a prediction model based on groundwater levels, categorize fire risk using environmental data using a machine learning approach, evaluate model performance, and apply the model to other regions. Data were collected from an Internet of Things (IoT)-based monitoring station that recorded eight environmental variables every 15 minutes for 3 months in Pematang Rahim. The observed predictor variables were rainfall, relative humidity, temperature, light intensity, soil moisture, wind direction, and wind speed, with groundwater level as the target variable. The model is expressed in the form of dominant IF–THEN rules: (1) IF soil moisture < 28% THEN groundwater level is in a very low condition (= -75 cm); (2) IF soil moisture is 28–62% AND rainfall is = 1.9 cm THEN the groundwater level is in a transition state from dry to wet (-41 to -22 cm); and (3) IF soil moisture is = 62% AND soil temperature is low THEN the groundwater level is relatively shallow (= -14 cm). The prediction results show an R2 of 0.93, RMSE of 10.42, MAE of 5.98, and MAPE of 0.76%, indicating an unbiased model. In addition, the simulation of the number of variables and data time shows that it is possible to reduce the predictor variables to three main variables, namely soil moisture, temperature, and rainfall, with the best time being 3 months. Furthermore, vulnerability classification is divided based on a critical threshold of groundwater level < -40 cm to determine the categories of "low" and "high" vulnerability. The classification results achieved 97% accuracy. The model was then applied to data from Teluk Dawan Village. The results of the model application showed that it could not yet be applied to other areas; training data were needed to enable the model to adapt to environmental variables. The classification of vulnerability in the new data with hotspot targets shows that, in sequence, days without rain, soil moisture, water level, and soil temperature affect fire vulnerability. The results of this study also provide recommendations for selecting environmental variables for estimating water levels, namely soil moisture, temperature, and rainfall, as a strategy for more effective monitoring of forest and peatland fire prevention.