Analysis of Machine Learning Algorithms’ Performance in Multitemporal Image Classification for Agricultural Management (Case Study: Qazvin Plain)

Abstract

Introduction

Due to the world's population growth and the severe problems caused by climate change, there is an immediate demand for sustainable farming methods and effective natural resource management. Agriculture forms the foundation of economic stability and food security, so methods that lead to maximum resource utilization are needed. However, they cause minimal environmental disturbance. This goal requires the ability to carry out fast and accurate crop mapping since it is used in strategizing agricultural activities, assessing yield, and ensuring sustainability.

Monitoring and mapping crop types by conventional field-based approaches have been the norm for a long time. They are primarily laborious, time-consuming, and expensive, especially on large scales. Remote sensing approaches, fueled by new-generation satellite imagery and machine-learning capabilities, offer a viable alternative to enabling large-scale monitoring and detailed classification of land cover and crop types. The most promising tools in this respect are the Sentinel-1 and Sentinel-2 satellites, with high spatial, temporal, and spectral resolution data, ideally suited for agricultural monitoring.

Sentinel-1 provides radar imagery, which is especially useful for monitoring vegetation structure even in cloudy conditions, while Sentinel-2 provides high-resolution optical images with multiple spectral bands suited for vegetation study. These datasets can be combined to capture complementary information on the physical and spectral characteristics of the land surface.

There are different kinds of ML algorithms. Here, the performance of three most common algorithms are compared: RF, SVM, and XGBoost. The relative strengths of each of the algorithms in classification provide insights that are critical to their suitability in diverse agricultural scenarios.



Materials and Methods

This study was conducted in the Qazvin Irrigation Network, an agriculturally significant area of about 80,000 hectares in Iran. The study area includes different types of land cover, such as wheat, alfalfa, fallow land, urban, and bare land, which are the most cultivated crops in the area. This heterogeneity makes classification difficult, especially in semi-arid regions where crops and other land cover classes have similar spectral signatures. The data from the Sentinel-1 and Sentinel-2 satellites were used for such challenges. The former acquires reliable radar data that captures surface roughness and soil moisture even under cloudy conditions. Meanwhile, the latter delivers high-resolution optical imagery with many spectral bands, which is excellent for studying vegetation health and structure. Several necessary data-preprocessing steps were carried out to ensure that accurate classifications were developed. Atmospheric and sensor noise reduction was accomplished via radiometric and geometric correction, respectively, for radar and optical imagery. Derived spectral indices such as NDVI, SAVI, and LAI aided the detection of vegetation characteristics under study by enhancing separability at spectral levels. Furthermore, temporal fusion was performed by combining images taken at different times to account for the phenological changes in vegetation over the growing seasons. These preprocessing steps allowed for a robust dataset representing spatial and temporal land cover changes.

Finally, three machine-learning algorithms were implemented for classifying the preprocessed satellite images: RF, SVM, and XGBoost. The reasons for choosing RF, an ensemble-based approach, include its robustness to noise and its ability to handle complicated datasets. On the other hand, SVM was adopted because it optimizes classification boundaries through its kernel-based feature. XGBoost is a highly accurate advanced gradient boosting technique that can realize large-scale computing with low expenses. The dataset was then divided into a training and testing set in a 70/30 ratio to prevent overfitting and ensure the model's reliability. Classification accuracy was assessed based on overall accuracy and the kappa coefficient, while the Jeffries-Matusita (JM) test quantified spectral separability between land cover classes.



Results and Discussion

The results demonstrated that integrating optical and radar data significantly improved classification accuracy. Among the three algorithms, RF outperformed the others, achieving an overall accuracy of 93.98% and a kappa coefficient of 0.996. These results highlight RF's ability to handle spectrally overlapping classes and complex datasets effectively.

The XGBoost algorithm also performed well, achieving an overall accuracy of 93.94%. However, its performance was slightly hindered by its inability to distinguish between classes with similar spectral characteristics, such as wheat and alfalfa. While providing reasonable results, SVM achieved a lower overall accuracy of 83.79%, mainly due to its sensitivity to spectral overlap.

The JM test revealed that certain classes, such as wheat and alfalfa, exhibited low spectral separability. This limitation underscores the importance of integrating radar data and spectral indices to enhance differentiation. The study also highlighted the potential of temporal data fusion to capture phenological changes, further improving classification performance.



Conclusion

This study indicated the potential of integrating multi-source remote sensing data and machine learning algorithms for crop classification in semi-arid regions. The RF algorithm proved the most accurate and robust method, showing its adaptability to the heterogeneous and complicated nature of the datasets. XGBoost and SVM are also very promising, but their performance could be improved further with additional parameter optimization.

Future research should investigate the application of more advanced techniques, such as CNNs and deep learning frameworks, to improve classification accuracy further. A deeper understanding of the dynamics of crops and land use changes can be achieved by including multi-temporal and multi-spectral datasets.

The results of such a study can have substantial implications for sustainable agriculture and resource management. In this context, remote sensing and machine learning technologies offer means to address critical challenges related to food security and environmental conservation in the most climate-vulnerable regions.