MACHINE LEARNING FOR MAPPING OF CROPS USING SATELLITE DATA

Abstract

One of the major challenges of the 21st century is to ensure global food security. Such challenges are especially prominent in developing countries, where food supply is already an issue. However, climate variability (very less rainfall condition cause drought or heavy rainfall cause flood) is also a key factor for increasing the vulnerability to food insecurity (Food and Agriculture Organization (FAO), 2013). In developing countries like India, urbanization turned rich agricultural land into residential area. Boosting agricultural practice may be one alternative to fulfil this increasing demand for food security. Therefore, accurate and timely information about crops is necessary for sustainable agricultural development. Mapping of crops at regional and national level gives input to the various agencies like insurance companies, geoportal, regional agricultural boards for decision making purpose. Crop classification is also important to understand how crop diversity affects the environmental condition within the region. Traditional methods of data collection are time consuming and expensive approach to acquire information for agricultural mapping. Advanced remote sensing techniques are very useful and cost-effective tool for acquiring a large amount of information regarding crops. During the last few decades, coarse to medium resolution data have been widely used for crop mapping at the local and global level. The recently launched Sentinel-2 opened a new door of opportunities for an application like crop classification. However, there is a lack of a framework to select Earth Observation (EO) data for a specific application. The suitability of remote sensing data for crop mapping at a specific site is also important, which may affect the classification results. The spectral similarity of the various crop makes crop identification a challenging task specifically in a heterogeneous agricultural environment. Machine Learning (ML) techniques are effectively used for various classification problems in the remote sensing domain. However, the selection of an appropriate ML algorithm is still a question of research due to the diversity of data sources, different types of problems to address, complexity of target classes and characteristics of the study area, etc. Therefore, this study aims to test the potential of several ML techniques for mapping of crop using Landsat-8 OLI and Sentinel-2A data. The considered ML techniques are k-Nearest Neighbor (k-NN), Naïve Bayes, Support Vector Machine (SVM), Stochastic Gradient Boosting (SGB), Adaboost.M1, Extreme Gradient Boosting (Xgboost), and Random Forest (RF). The literature review suggests that RF and SVM are the most popular ML techniques, whereas, limited i research has been carried out to test the potential of SGB and Xgboost for crop classification. In addition, this study also proposes an Ensemble Object Based Image Analysis (EOBIA) framework for mapping of crops. In this work, the selected study area ‘Roorkee’ is located in the Haridwar district of Uttarakhand India, and covers a total area of 1049.31 km2. The selected region is heterogeneous as it consists of hills having dense forests at the northern side, whereas, rich agricultural land at the southern side. This area consists of small agricultural fields usually separated by planted trees or other crops. The selected study area is partitioned into eleven Land Use Land Cover (LULC) classes namely High Density Forest (HDF), Low Density Forest (LDF), orchard, sand, built-up, water, fallow, and crops categorised as wheat, sugarcane, fodder, and other crops. Here, Sentinel-2A and Landsat-8 OLI data have been used for crop classification. In case of Sentinel-2A data, resampling has been carried out on six bands to convert the data from 20m to 10m spatial resolution. Whereas, Landsat-8 OLI data have been used at 30m spatial resolution consist of six spectral bands (Blue, Green, Red, NIR, SWIR1, and SWIR2). The overall methodology may be divided into three subparts. Initially, all ML techniques have been implemented for both the datasets. Parameter optimization has been carried out for each ML classifier and final models have been developed with the optimal set of tuning parameters. Performance evaluation has been performed based on accuracy measures to select best three ML classifiers for further processing. Spatial and spectral features may be beneficial to distinguish classes more accurately. Therefore, textural features i.e. Gray-Level Co-occurrence Matrix (GLCM) and selected Vegetation Indices (VI) have been generated. In order, to select the most relevant features for crop mapping, feature selection has been done using Recursive Feature Elimination (RFE), and Sequential Forward Floating Selection (SFFS) approaches. The selected optimal feature set is used to train the ML classifiers and performance has been tested by using various accuracy measures. In addition, data suitability analysis has been carried out in each classification scenario. In the proposed object-based framework (EOBIA), image segmentation has been carried out to create objects from pixels. Multiresolution image segmentation method has been chosen for object creation using both datasets i.e. Landsat-8 OLI and Sentinel-2A. Various parameters associated with segmentation like shape (𝑤𝑠ℎ𝑎𝑝𝑒), spectral (𝑤𝑠𝑝𝑒𝑐𝑡𝑟𝑎𝑙), ii compactness (𝑤𝑐𝑜𝑚𝑝𝑡), smoothness (𝑤𝑠𝑚𝑜𝑜𝑡ℎ), etc. are used to create objects from input data. Here, different features such as spectral, contextual, shape, and texture-based features have been computed based on the Mean Decrease Gini score which may be beneficial for the classification. Thereafter, the selected ML classifiers (RF, SGB, and Xgboost) are trained on the segmented objects, and prediction is done using the trained ML classifiers. Finally, the assessment of accuracy has been performed using various measures as Overall Accuracy (OA), Precision, Recall, F1-score, Quantity Disagreement (𝑄𝑑) and Allocation Disagreement (𝐴𝑑). The mean plot analysis has been performed by computing the mean reflectance values of training samples for all considered eleven LULC classes on both Landsat-8 OLI and Sentinel-2A dataset. This analysis is beneficial to understand the intermixing of various classes. It is found that in case of Landsat-8 OLI data, classes are more separable in the NIR band. Results demonstrated that additional information provided by Sentinel-2A data (in terms of Red-Edge bands) may be beneficial to distinguish complex LULC classes, more specifically crop classes. Results of the ML classifier’s performance analysis demonstrated that Xgboost outperformed for both the datasets. Results indicate that when only spectral features are used, Xgboost obtained the Overall Accuracy (OA) of 81.79% along with Allocation Disagreement (𝐴𝑑) of 12.30% and Quantity Disagreement (𝑄𝑑) of 5.91%. Whereas, for Sentinel-2A data maximum OA of 86.91% has been obtained with 𝐴𝑑 and 𝑄𝑑 value of 8.70% and 4.38% respectively. It is found that RF and SGB also performed well as compared to other classifiers, whereas, k-NN produced the lowest accuracy of 70.82% for Landsat-8 OLI data, and Naïve Bayes has given the minimum OA of 75.96% using Sentinel-2A data. The outcome of this analysis indicates that Xgboost, RF, and SGB have given the best classification results, therefore, these ML classifiers are selected for further analysis. This research work also aims to test the suitability of Sentinel-2A over Landsat-8 OLI data for crop classification in a heterogeneous agricultural environment. Results indicate that Sentinel-2A data given a significant improvement of 5.56 percentage points (average increase of all classifiers) in OA in comparison to Landsat-8 OLI data. The main reason for this accuracy rise is the additional information provided by the Red-Edge bands of Sentinel-2A. Additionally, the higher spatial resolution of Sentinel-2A also contributed to increase the iii iv

accuracy of classification in such an agricultural environment where crop fields are small in size and usually mixed with other crops.

Further, the impact of the inclusion of textural features (GLCM) and Vegetation Indices (VI) has been examined on both datasets, and optimal features are determined using feature selection methods. Results indicate that the SFFS method obtained minimum misclassification error by considering fewer features as compared to the RFE method. It is found that an average increase of 2.14 and 2.62 percentage points has been observed using optimal features of Landsat-8 OLI and Sentinel-2A respectively. The maximum increase of 3.08 percentage point in OA is observed by SGB classifier using optimal features of Sentinel 2A data.

Finally, the outcomes of EOBIA framework have been evaluated for the datasets. Results indicate that the proposed EOBIA approach achieved approximately 4 percentage points (average of RF, SGB, and Xgboost) higher OA than ML models using optimal features for Sentinel-2A data. The highest OA of 93.77% has been achieved for Xgboost classifier. Remarkable improvements have been noted in the class-specific accuracy of both major crops (wheat and sugarcane). However, in the case of Landsat-8 OLI data, only a marginal improvement of approximately 1 percentage point has been noted in the classification accuracy.

This study focused on the mapping of crops using ML techniques and the proposed EOBIA framework. It is found that ensemble classifiers such as SGB and Xgboost have great potential for crop classification in complex classification scenario. The proposed EOBIA framework effectively identified major crops using Sentinel-2A data.