AN OBJECT-BASED LIGHTWEIGHT CNN MODEL FOR CROP CLASSIFICATION USING SATELLITE IMAGE TIME-SERIES

---

Tezin Türü: Doktora

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

Tezin Onay Tarihi: 2026

Tezin Dili: İngilizce

Öğrenci: MÜSLÜM ALTUN

Danışman: Mustafa Türker

Özet:

Timely and accurate crop type mapping is critical for yield estimation, food security assessments, and effective agricultural management. Recent developments in crop classification from satellite imagery have shifted from using single-date static images to employing spectral data from image time-series. Object-based classification approaches are commonly preferred for crop mapping tasks. Through the innovations in machine learning, deep convolutional neural networks (CNNs) have become prominent state-of-the-art methods in crop detection. This study proposes a lightweight CNN deep learning model for agricultural crop type detection from satellite image time-series. The proposed model was evaluated within the context of object-based (segment-based and field-based) and pixel-based image analysis frameworks using the Sentinel-2 image time-series. The experiments were conducted in two distinct agricultural regions in Türkiye (Manisa and Kırklareli) characterized by diverse crop types and distributions. Image classification was carried out using all bands of Sentinel-2 image time-series, as well as the selected optimal bands determined through the Support Vector Machine–Recursive Feature Elimination (SVM-RFE) algorithm. For object-based approaches, a rotation process was applied on the objects (fields and segments) to align them with their minimum bounding boxes. The results were evaluated without rotation and with rotation cases of the objects to assess the impact of this geometric alignment.

The results of both study areas revealed that the proposed lightweight CNN model based on all bands and with rotation technique yielded the highest model accuracies (train, validation, and test) and the best classification accuracies in terms of overall accuracy (OA) and Kappa coefficient. Specifically, for this configuration, the proposed model achieved training accuracies of 97.2% and 92.0%, validation accuracies of 86.6% and 88.0%, and test accuracies of 83.1% and 86.6%, for Manisa and Kırklareli respectively. In terms of crop classification accuracy, the proposed model yielded OA values of 89.8% for Manisa and 88.8% for Kırklareli, with the corresponding Kappa coefficients of 0.879 and 0.871.

For both study areas, classifications based on all spectral bands consistently outperformed the classifications based on the selected optimal bands by approximately 3%, and the rotation technique improved results by about 2% compared to without rotation technique. The field-based classification approach demonstrated superior OA performance over the object-based approach by approximately 6% in Manisa and 9% in Kırklareli. The field-based approach also outperformed the pixel-based approach by approximately 10% in Manisa and 15% in Kırklareli. On the other hand, the object-based approach surpassed the pixel-based approach by about 4% and 6% in the respective study areas. Moreover, the results of the proposed model demonstrated advantages compared to VGG-16, ResNet-50, and U-Net in all tested configurations. Overall, the findings of this study demonstrate that the proposed light CNN model is promising for accurate and cost-effective field-based and object-based crop type classification using Sentinel-2 image time-series.