AGRICULTURAL CROP MAPPING USING SENTINEL-1 TIME SERIES: EVALUATION OF DEEP LEARNING AND MACHINE LEARNING MODELS

MERVE HİLAL ARSLAN

AGRICULTURAL CROP MAPPING USING SENTINEL-1 TIME SERIES: EVALUATION OF DEEP LEARNING AND MACHINE LEARNING MODELS

Tezin Türü: Yüksek Lisans

Tezin Yürütüldüğü Kurum: Hacettepe Üniversitesi, Fen Bilimleri Enstitüsü, Geomatik Mühendisliği A.B.D., Türkiye

Tezin Onay Tarihi: 2026

Tezin Dili: İngilizce

Öğrenci: MERVE HİLAL ARSLAN

Danışman: Mustafa Türker

Özet:

Accurate and timely crop type maps are essential for agricultural monitoring, production planning, and sustainable land management. Sentinel-1 time-series SAR imagery is a valuable data source for crop classification because it enables regular observations regardless of cloud cover and daylight conditions.

In this study, crop types in an agricultural area in Kırklareli, Türkiye, were classified using the Sentinel-1 time-series SAR data acquired in 2021. A Convolutional Long Short-Term Memory (ConvLSTM)-based classification approach was applied and its performance was compared with those of Gated Recurrent Unit (GRU), 3D Convolutional Neural Network (3D CNN), and Random Forest (RF) models. The crop types classified included barley, sunflower, wheat, meadow, paddy rice, rapeseed, maize, and triticale. Farmer Registration System (FRS) data were used as reference data for model training and accuracy assessment. A total of ten Sentinel-1 time-series images acquired between April and September 2021 were selected. In the classification, the VV (vertical-vertical) and VH (vertical-horizontal) backscatter bands, and the derived VH/VV ratio feature, were used as input features.

Classification performance was evaluated in two ways. First, the trained models were evaluated using the independent test dataset. Second, the resulting classification maps were assessed at the parcel level. In the test-set evaluation of the models, RF achieved the highest accuracy of 84.78%, followed by GRU with 83.94%, 3D CNN with 83.80%, and ConvLSTM with 83.57%. In the parcel-based assessment of the classification maps, RF also achieved the highest overall accuracy (OA) of 85.38%, followed by 3D CNN with 84.78%, ConvLSTM with 84.31%, and GRU with 84.03%. Based on the results achieved in this study, RF ranked first in both evaluations. However, the relatively small differences among the results indicate that all four models provided comparable crop classification performance.