Enhancing Machine Learning Crop Classification Models through SAM-Based Field Delineation Based on Satellite Imagery

Accurate crop classification is vital for various agricultural applications, including yield estimation, land use monitoring, and precision farming. Machine learning models trained on satellite imagery have shown promising results in this domain. However, the accuracy of these models heavily depends on the quality of input features, particularly the delineation of individual fields. Traditional methods for field delineation often face challenges in complex landscapes and heterogeneous agricultural patterns. In this study, we propose a novel approach to enhance machine learning crop classification models by integrating the Segment Anything Model (SAM) for field delineation by utilizing satellite imagery. The Segment Anything Model (SAM) algorithm is a versatile segmentation method capable of segmenting images into meaningful regions based on a wide range of features. By leveraging SAM, we aim to accurately delineate field boundaries from satellite imagery, providing more refined input features for machine learning models. The proposed methodology involves several key steps. First, we preprocess satellite imagery to enhance spectral signatures and differentiation between parcels. Next, we apply SAM to segment the imagery into distinct field boundaries based on various features, extracted from Sentinel-2 satellite imagery. These delineated fields serve as input features for training machine learning classification models. One of the primary advantages of SAM-based field delineation is its flexibility in segmenting diverse agricultural landscapes. SAM can adapt to different environmental conditions and crop types, effectively capturing the spatial variability present in satellite imagery. This comprehensive delineation enables machine learning models to learn more representative features, thereby improving classification accuracy. To evaluate the effectiveness of our approach, we conducted experiments on a diverse dataset comprising satellite imagery from various regions and crop types. We achieved 81 % accuracy in crop identification, while in this work we show the effectiveness of parcel-based filtering, which removes isolated misclassified pixels. Since multiple crops may be cultivated within the same cadastral unit, our field delineation parcels proved to be more effective for postprocessing (filtering) purposes compared to cadastral data, thus ensuring accurate attribution of crops to corresponding land parcels. Our results demonstrate significant improvements in crop classification maps when incorporating SAM-derived field boundaries into the training data. Our study highlights the potential of SAM-based field delineation techniques to enhance machine learning crop classification models using Sentinel-2 satellite imagery. By accurately delineating field boundaries, we provide more informative input features, leading to improved classification performance and better insights into agricultural landscapes. This research contributes to advancing the capabilities of remote sensing technologies for precision agriculture and land management applications.