Supervised Machine Learning Algorithms for Land Cover Classification in Casablanca, Morocco

In the domain of environmental monitoring and land cover analysis, the application of supervised algorithms for classifying Landsat satellite imagery [1] stands as a critical and rapidly evolving area of research. This investigation aims to conduct a comprehensive assessment of the effectiveness of various supervised algorithms in the classification of Landsat images, with a specific emphasis on the urban landscape of Casablanca, Morocco [2]. Characterized by its rapid urbanization and intricate land-use dynamics, Casablanca provides a complex setting for this analysis, marked by a wide array of urban structures and evolving land-use patterns.

The imperative for this study emerges from the increasing necessity to establish robust classification methodologies capable of delineating distinct land cover types within dynamic urban environments. The consequences of hasty urban expansion necessitate an in-depth examination of spatial transformations, particularly for applications in natural resource management, urban planning, and environmental challenge mitigation.

This research explores a spectrum of state-of-the-art supervised classification algorithms [3, 4], namely SVM [5], RF [6], GTB [7], DT [8], MD [9], and CART [8]. Each algorithm possesses unique attributes; however, their aptitude in the specific context of Casablanca's variable and complex environmental conditions necessitates a thorough evaluation.

The study aims to transcend mere analysis of overall accuracy, delving into finer aspects such as sensitivity to varied land cover classes, resilience under spatial and temporal variations, and generalizability across diverse urban settings. Land cover classification is integral to informed decision-making in sustainable development, urban planning, and environmental preservation. Nevertheless, the task is challenged by the complexity and rapid transformation of land characteristics and the demand for precise large-scale data acquisition.

Land use classification stands as a fundamental element in urban planning [10-12], natural resource management, and risk assessment. The ability to discern how land is utilized in different regions is imperative for informed decision-making processes related to sustainable development. Yet, achieving precision in land use classification is laden with challenges.

The accuracy of land use classification is paramount. It ensures that the information employed in decision-making, encompassing urban planning, environmental conservation, or natural risk management, is relevant and reliable. The primary difficulties in this endeavor arise from the complexity and rapid evolution of soil characteristics, coupled with the necessity of procuring accurate data on a substantial scale.

In addressing these challenges, the role of satellite technology, particularly Landsat 8 OLI, is indispensable. These satellites provide critical data for land cover classification, facilitating extensive observation and consistent data collection on a large scale.

To effectively tackle the intricacies of land classification, the deployment of specific algorithms is essential. This study will evaluate the efficacy of algorithms such as SVM [5], RF [6], GTB [7], DT [8], MD [9], and CART [8]. These algorithms will be assessed for their ability to accurately classify land cover in varying contexts.

SVM: SVMs are a type of supervised learning algorithm employed for both classification and regression challenges. Their primary objective is to identify an optimal hyperplane that categorizes the data into distinct classes, maximizing the margin between these classes. This separation is key to SVM's efficacy in handling various classification tasks.

CART: CART algorithms utilize tree structures for decision-making, based on specific data characteristics. They function by recursively segmenting data into homogeneous subgroups, thereby facilitating precise classification or the prediction of continuous variables. This method is particularly effective in scenarios where decision rules can be hierarchically structured.

RF: RF comprises an ensemble of decision trees, each constructed using a random subset of the training data. This approach enhances the model's accuracy and robustness by mitigating the variance typically associated with individual decision trees, making it a robust choice for complex classification tasks.

GTB: GTB involves sequentially building multiple models, often weak decision trees, to rectify errors from preceding models. This cumulative approach results in a more effective overall model, proving valuable in scenarios requiring nuanced error correction and refinement.

DT: DTs are a form of tree structure that represent various decisions and their potential outcomes. Each node signifies a feature, each branch symbolizes a decision pathway, and each leaf denotes a final output or classification. This method is widely used for its simplicity and interpretability.

MD: The Minimum Distance classifier operates by categorizing objects based on their proximity to predefined prototypes. It computes the similarity between the characteristics of an object and those of the prototypes, classifying the object into the class of the nearest prototype. This technique is particularly useful in applications where proximity to known categories is a reliable indicator of class membership.

The principal aim of this research is to critically compare and assess the performance of selected algorithms when applied to Landsat 8 OLI imagery within the Casablanca region. Utilizing the Google Earth Engine platform [13], the study evaluates the capability of these algorithms to accurately categorize various land cover types, including Forest, Barren, Built-Up, Water, and Cropped areas. This comprehensive assessment endeavors to elucidate the efficacy, suitability, and limitations of each algorithm in the context of land cover classification in Casablanca.

The significance of this investigation lies in its potential to furnish decision-makers, urban planners, and environmentalists with essential insights regarding the reliability and precision of land use data derived from these algorithms. Such information is crucial for making informed decisions in urban development and environmental conservation.

This document is methodically structured to facilitate a clear understanding of the research process: Section 2 offers a review of prior studies in the realm of satellite image classification. Section 3 introduces the study area and delineates the data sources, providing a solid foundation for the research. Section 4 outlines the experimental methodology employed in this study. Section 5 is dedicated to the interpretation and analysis of the results, including a thorough discussion of the experimental findings. The final section, Section 6, articulates the conclusions drawn from this investigative endeavor.

3.1 Study area

Casablanca is a vibrant city located in the western part of Morocco, situated on the Atlantic coast. Its geographical coordinates are approximately 33.5731° N latitude and 7.5898° W longitude. Figure 1 shows Casablanca area. As the largest city in Morocco and one of the major cities in North Africa, Casablanca serves as an essential economic, cultural, and industrial hub in the region.

The surrounding areas of Casablanca consist of suburbs, agricultural lands, and natural features, making it an ideal study area for assessing land cover dynamics and urban growth using satellite imagery.

1.png

Figure 1. Casablanca study area

3.2 Overview of Landsat 8 OLI satellite data used

The Landsat 8 OLI satellite data used in this study were acquired over multiple time periods, enabling temporal analysis to monitor changes in land cover over time. The spatial resolution of Landsat 8 OLI imagery ranges from 15 meters for the panchromatic band to 30 meters for the multispectral bands. Table 1 shows an overview of temporal and spatial resolution of Landsat 8.

Table 1. Spatial and temporal resolution of the Landsat 8 OLI imagery [24]

Prior to analysis, the Landsat 8 OLI satellite data [24] underwent necessary preprocessing steps, such as atmospheric correction and radiometric calibration, to ensure data accuracy and reliability. These preprocessed images served as the foundational dataset for applying supervised algorithms [25] to classify land cover in the Casablanca region.

The Landsat 8 OLI satellite data, with its high-resolution and multispectral capabilities, provides a valuable resource for understanding land cover dynamics, urban expansion, and environmental changes in Casablanca, contributing to informed decision-making for sustainable development and resource management in the city.

The experimental results highlight the varying performance of the different classification algorithms evaluated. Each of these algorithms presents strengths and weaknesses that merit in-depth analysis to guide the selection of the most appropriate model for specific applications. Figure 4 shows the overall accuracy (OA) of different methods.

Various metrics [35] have been used to assess the performance of each classification method [4], including overall accuracy, user accuracy for each class, manufacturer accuracy for each class and the Kappa coefficient obtained from the confusion matrix. The equations used to calculate these indicators are shown in Table 5. To assess the accuracy of classified land cover maps produced by supervised machine learning algorithms in GEE [11, 12].

We carried out an in-depth analysis of the confusion matrix for each method. The results are presented in Tables 3 and 4, which give a comprehensive overview of the accuracy assessment.

The confusion matrix is an essential tool for assessing the performance of a classification model [36], providing a detailed insight into how it classifies data according to its reality. It is generally square in shape, with the entries representing the different possible outcomes of a classification, as shown in Figure 3. In the case of a binary classification problem (with two classes), it is generally divided into four boxes [27]:

·True positives (VP): The model has correctly predicted that the element belongs to the positive class (actual and predicted values are both positive).

·True negatives (VN): The model has correctly predicted that the element does not belong to the positive class (actual and predicted values are both negative).

·False positives (FP): The model has incorrectly predicted that the element belongs to the positive class when it doesn't (actual value is negative and predicted is positive).

·False negatives (FN): The model has wrongly predicted that the element does not belong to the positive class, when in fact it does (actual value is positive and predicted is negative).

To improve the classification results, spectral indices such as NDVI [37], BSI, MNDWI and NDBI were incorporated. These indices were used during the feature extraction phase and included in the training data, helping to improve the classifier's performance. This approach enabled more accurate identification of land use categories.

4.png

Figure 4. Overall accuracy (OA) of different methods

5.png

Figure 5. Legend of classification results

Random Forest clearly emerged as the leading algorithm, posting a remarkable overall accuracy of 95.42%. This high performance can be attributed to the ensemble nature of this model, which combines predictions from multiple decision trees to reduce over-fitting and improve generalization. This is particularly beneficial in complex contexts such as land cover classification. However, it is crucial to assess the model's sensitivity to variations in the training data, and to test its robustness in the face of different data sets. In addition, the impact of hyperparameters, such as the number of trees in the forest, should be examined to optimize model performance.

Minimum Distance also stands out, with an overall accuracy of 94.77%. Its strength lies in its conceptual simplicity and efficiency, particularly in class discrimination. However, the performance of this algorithm may be sensitive to the spatial distribution of the data and differences in class variability. Further analysis of the discriminating features identified by the MD could help to understand the underlying mechanisms of classification. It could also help identify scenarios where the simplicity of the MD approach might be preferable to more complex models.

Table 3. References data /classes

Table 4. Land cover evaluation metrics of each method

6a.png

6b.png

6c.png

6d.png

6e.png

6f.png

Figure 6. Land cover classification map for Casablanca

7a.png

7b.png

7c.png

7d.png

7e.png

7f.png

Figure 7. The producer’s accuracy (PA), and the user’s accuracy (UA) values of each classifier

Table 5. Kappa coefficient and overall accuracy for land use classification of each classifier used

Table 6. Equations of evaluation metric

Although the SVM achieved 100% accuracy for the "Water" class, its overall accuracy of 83% raises questions about its ability to handle variability in other classes. SVM is renowned for its ability to handle high-dimensional data, but its success often depends on the judicious choice of kernels and parameters. A thorough analysis of support vectors and separation margin could reveal information about the relative difficulty of classification for different classes. Optimization of SVM parameters could also be explored to improve its performance.

Gradient Tree Boost presents an interesting balance between producer and user accuracy, with an overall accuracy of 93.46%. This algorithm can adjust its predictions based on previous errors, which can be advantageous in scenarios where classes are unbalanced. However, GTB's interpretability may be less than that of other, simpler algorithms, raising questions about the model's transparency. For applications where model explicability is crucial, GTB-specific interpretation techniques could be explored [7].

Decision trees (CART and DT) have shown solid performance, albeit inferior to that of RF. Their simplicity and interpretability make them attractive candidates in contexts where understanding the decision process is crucial. However, the propensity of decision trees to overfitting can be a challenge, and regularization techniques may be needed to improve their generalizability. The use of cross-validation and other model validation strategies could help alleviate this problem. Table 3 and 4 illustrate the Kappa coefficient and Overall accuracy for land use classification of each classifier used.

Integrating the results of previous studies [3, 19, 22, 26, 37-42] and our own research, the main achievement of this study lies in the successful adaptation of the methodology to land use mapping and evaluation in different geographical contexts.

The results underline the superiority of the Random Forest classifier over other classifiers, with an impressive accuracy of 95.42% and a Kappa coefficient of 0.94, underscoring its substantial superiority over random classification. Figure 5 shows the legend of classification results and Figure 6 gives a visual presentation of the classified cards.

In summary, this research has contributed significantly by successfully adapting the methodology for mapping and assessing land use in diverse regions or countries. The results underlined the importance of spectral indices in improving classification accuracy and highlighted the need to consider the inherent complexity of land cover classification in diverse contexts. Figure 7 shows the producer’s accuracy (PA), and the user’s accuracy (UA) values of each classifier.