Physical changes on the earth’s surface due to natural processes and human interventions are continuous phenomena. The human induced changes in agricultural land use appear more frequently due to seasonal cropping patterns. Efficient technologies and advanced methodologies are required for recording and analyzing the extent and pace of such changes for monitoring and management of economic and environmental aspects. As the Remote Sensing [RS] platforms providing images of the Earth’s surface at various spatial and temporal resolutions have evolved over time, so have developed the image processing and analysis approaches ¹. One of the prominent image classification methods, Geographic Object-Based Image Analysis [GEOBIA], segments the target RS image by identifying the pixels of similar characteristics and then groups them into objects corresponding to various land use land cover [LULC] classes ². This approach offers more accurate representation of landscape structures and patterns by reducing the noise associated with single-pixel based analysis ³. GEOBIA is particularly effective in the processing of high-resolution images as it can capture fine details and patterns more precisely ⁴. Overcoming the effects of shadows, it utilizes various object attributes like color, shape, size, and texture for efficient differentiation and classification of complex objects ⁵. Moreover, it can be integrated with the technologies like Machine Learning [ML] and Artificial Intelligence [AI] to enhance analytical algorithms. As mentioned, GEOBIA involves two main steps, i.e. image segmentation and object classification. A variety of algorithms are available for segmentation and delineation of objects, and ‘Simple Non-Iterative Clustering [SNIC]’ is used widely. The common ML classification algorithms used include Support Vector Machines [SVM], Gradient Boosting Trees [GBT] and Random Forests [RF]. The widespread use of these classifiers in remote sensing applications is recognized for their robust performance ^{6, 7}. Thus, Comparing the relative performance of these classifiers is desirable for selecting the most suitable model considering the crucial factors like accuracy, complexity, resource usage, and practical application flexibility. Proper implementation of ML classification algorithms improves accuracy, increases operational flexibility and allows efficient processing of large datasets. Additionally, these minimize the chances of human errors and enhance the reliability of the results 8.

Google Earth Engine [GEE], based on Google Cloud Platform, is an invaluable online facility for storage, retrieval, processing and applications of remote sensing datasets of various spatial and temporal scales. The platform's ability to quickly access high-resolution images allows the users to efficiently process large size datasets. Additionally, GEE offers numerous pre-built tools and algorithms for data processing, analysis and display, making it ideal for studying and monitoring contemporary environmental changes through online platforms ⁹. The integration of Google Cloud ML technology with RS, particularly utilizing Sentinel-1 datasets, involves extracting key features from Synthetic Aperture Radar [SAR] data. This provides a highly efficient tool and technique, especially effective for classifying rice fields. The SAR imaging capably captures data during day and night times as well as in all weather conditions i.e. cloudy, rainy, snowy etc. Also, the SAR images provide VV [Vertical Transmit Vertical Receive] and VH [Vertical Transmit Horizontal Receive] polarizations, and the difference values of these two channels enhance the probability of identifying the physical characteristics of rice crop area more accurately ^{10, 11}.

Classifying post-harvest rice fields using satellite imagery poses significant challenges due to varying field conditions and the diverse nature of satellite data. Identifying the optimal combination of classification methods and parameters for accurate detection remains a complex task. Therefore, this study aims to explore and compare various classification methods for identifying post-harvest rice fields using Sentinel-1 satellite imagery within the GEE platform.

Chae Chang and Buak Kang subdistricts, located in the San Kamphaeng District of Chiang Mai Province, Thailand, cover an area of approximately 47.225 km² [Figure 1].

The study area located between 99°04'32"E, 99°10'54"E longitudes and 18°40'07"N, 18°44'30"N latitudes includes 23 villages. Its physiography is a mixture of flat plains and low hills, and the cultural landscape is predominantly composed of agricultural fields with intermittent settlements and roads. It has tropical monsoon climate with an average annual rainfall of about 1,100 millimeters, the large part of which is received during the rainy season from May to October. The average monthly temperature is around 31°Celsius with a mean monthly minimum about 18°C in January and mean monthly maximum nearly 38°C in April. The area also benefits from two well-developed irrigation systems, namely Mae-On and Mae Kuang, ensuring the water supply essential for winter-season rice cultivation. The advantageous topography, favorable climatic conditions and availability of irrigation facilities make these subdistricts well-suited for large-scale rice cultivation supporting the livelihoods of the residents. The population of these subdistricts is about 11,000 persons, mostly engaged in rice cultivation which is their staple food and primary source of income ¹².

This study attempts to identify the most suitable machine learning algorithm for delineating the area under rice crop adopting GEOBIA approach. It investigates the outcomes of several crucially relevant factors which include 5 combinations of 6 multi-sensor satellite image bands, 5 sets of 4 image segmentation parameters and 3 classifier algorithms using GEE for processing and analysis. The SNIC operation was run for segmenting the selected multi-sensor image bands and LULC classification was performed through SVM, GBT and RF algorithms.

The results indicate that minimal image segmentation provides better alignment of the segment boundaries with the boundaries of the corresponding sample areas collected in fieldwork. Although, an optimal set of 4 SNIC parameters was identified but these may or may not be suitable for other areas because shape and size of agricultural fields require adaptation of these parameters.

While comparing the 3 ML classifier algorithms, RF outperformed SVM and GBT in terms of OA [97%] and K [0.94]. SVM was less effective due to its computational intensity and reduced accuracy with distinct class boundaries, but GBT performed better particularly in case of imbalanced spatial pattern. Integration of the 3 optical bands of Sentinel-2 with the VV and VH polarization data of Sentinel-1 enhanced the classification performance significantly, highlighting the benefits of multi-sensor data fusion.

The capability of segments in capturing the objects accurately is fundamental in determining the accuracy of classification. Each classifier's capabilities vary depending on the segmentation parameters. The results show that using different classifiers with the same parameters significantly impacts classification outcomes, with RF performing best compared to SVM and GBT. Therefore, it can be concluded that classification accuracy and precision should derive from both good segmentation and good classifier.