Mapping Nationwide Subfield Division Dynamics in Saudi Arabia Using Temporal Patterns of Sentinel-2 NDVI and Machine Learning

Abstract

Object-based analysis is widely used for extracting information from satellite data using machine learning, offering reduced sensitivity to fine-scale variability, noise, and computational cost compared to pixel-based methods. However, segmentation algorithms for center-pivot fields often treat fields as single units, neglecting that a field can be subdivided into different sections caused by varied management practices, such as differing planting and harvesting dates, crop types, and rotations. This variability is particularly prevalent in hot, arid regions, such as Saudi Arabia, where precise water and crop management are crucial for sustaining agricultural productivity. However, such subfield division reduces the accuracy of object-based agroinformatics insights and the effectiveness of large-scale analyses. A machine learning-based approach combining Kmeans clustering and cosine similarity was developed to quantify subfield divisions using temporal features derived from Sentinel-2 normalized difference vegetation index (NDVI) time series. The performance of discrete wavelet transformation(DWT) and Savitzky–Golay filtering was compared for processing the NDVI time series. When evaluated against a reference dataset, the approach achieved a maximum accuracy of 93.38% with DWT level 1 decomposition using the “haar” wavelet function. These parameters were applied to map the nationwide center-pivot subfield division dynamics across Saudi Arabia from 2019 to 2023. Results revealed that approximately 20% of center-pivot fields exhibited subfield divisions, ranging from 5740 fields (2083 km$^{2}$) in 2020 to 7342 fields (2770 km$^{2}$) in 2023. Larger fields were more prone to subfield divisions, with a median acreage of 40 ha compared to 20 ha for undivided fields. Dominant management strategies included half-to-half and 5:3:2 divisions. This approach enhances object-based agroinformatics products and facilitates more accurate food security assessments.

Keywords

• Agroinformatics
• cosine similarity
• discrete wavelet transformation
• Kmeans clustering
• Savizky–Golay filter (SG)
• subfield division