With the rapid development of remote sensing and geographic information systems, accurate land cover classification has become essential for ecological monitoring and land use management. This study evaluates three classification algorithms—KMeans, Random Forest, and Multi-Layer Perceptron—using Sentinel-2 multispectral imagery across three regions: urbanized Shanghai, the agricultural hills of Tuscany, and the alpine area of Val d'Aosta. A standardized workflow was implemented, including image preprocessing, manual annotation, feature extraction, model training, result visualization, and accuracy assessment. The results show that supervised models clearly outperform the unsupervised KMeans algorithm. Among them, MLP achieved the highest accuracy and spatial consistency, producing coherent and meaningful land cover maps. Random Forest also performed well in distinguishing spectrally distinct classes but was less effective in refining boundaries and representing minority categories. KMeans, by contrast, only captured coarse patterns and showed high levels of misclassification. The comparison of sampling strategies further revealed that uniform sampling improved recognition of minority classes and reduced imbalance, strengthening model robustness. Overall, the findings indicate that MLP with uniform sampling provides the most effective framework for diverse land cover classification, offering guidance for large-scale remote sensing applications and ecological assessment. The thesis is organized into seven chapters. Chapter 1 outlines the Earth Observation background, research objectives, and the Sentinel-2 dataset. Chapter 2 introduces the Google Earth Engine platform, describing data catalogs, classifier implementations, and evaluation metrics. Chapter 3 presents the study areas and training samples, while Chapter 4 details the preprocessing workflow, including cloud masking, band selection, and spectral index computation. Chapter 5 reports the classification results for KMeans, Random Forest, and Multi-Layer Perceptron across the study regions. Chapter 6 compares proportional and uniform sampling strategies, highlighting their impact on classification robustness. Finally, Chapter 7 synthesizes the overall results, emphasizing model performance and the role of sampling strategies, and provides recommendations for remote sensing applications and ecological assessments.
Con il rapido sviluppo del telerilevamento e dei sistemi informativi geografici, una classificazione accurata della copertura del suolo è diventata fondamentale per il monitoraggio ecologico e la gestione del territorio. Questo studio valuta tre algoritmi di classificazione—KMeans, Random Forest e Multi-Layer Perceptron—utilizzando immagini multispettrali Sentinel-2 in tre regioni: l’area urbanizzata di Shanghai, le colline agricole della Toscana e la zona alpina della Valle d’Aosta. È adottato un flusso di lavoro standardizzato che comprende la pre-elaborazione delle immagini, l’annotazione manuale, l’estrazione delle caratteristiche, l’addestramento dei modelli, la visualizzazione dei risultati e la valutazione dell’accuratezza. I risultati mostrano che i modelli supervisionati superano l’algoritmo non supervisionato KMeans. L’MLP ha raggiunto la massima accuratezza e coerenza spaziale, producendo mappe coerenti e significative. Random Forest ha dato buoni risultati nella distinzione di classi spettralmente diverse, ma è stato meno efficace nel definire i confini e rappresentare categorie minoritarie. KMeans, invece, ha colto solo schemi grossolani con elevata misclassificazione. Il confronto tra strategie di campionamento ha evidenziato che quello uniforme migliora il riconoscimento delle classi minoritarie e riduce lo squilibrio, rafforzando la robustezza del modello. Nel complesso, i risultati indicano che l’MLP con campionamento uniforme rappresenta l’approccio più efficace per la classificazione della copertura del suolo, fornendo orientamenti per applicazioni di telerilevamento su larga scala e valutazioni ecologiche. La tesi è articolata in sette capitoli: il Capitolo 1 presenta il contesto dell’Osservazione della Terra, gli obiettivi di ricerca e il dataset Sentinel-2; il Capitolo 2 introduce la piattaforma Google Earth Engine, descrivendo cataloghi di dati, implementazioni dei classificatori e metriche di valutazione; il Capitolo 3 illustra le aree di studio e i campioni di addestramento; il Capitolo 4 descrive il flusso di pre-elaborazione, comprendendo mascheramento delle nuvole, selezione delle bande e calcolo di indici spettrali; il Capitolo 5 riporta i risultati ottenuti con KMeans, Random Forest e Multi-Layer Perceptron; il Capitolo 6 confronta il campionamento proporzionale e uniforme, sottolineandone l’impatto sulla robustezza; infine, il Capitolo 7 sintetizza i risultati complessivi, evidenziando prestazioni dei modelli e implicazioni di strategie di campionamento, e propone raccomandazioni per applicazioni di telerilevamento e valutazioni ecologiche.
Comparison between different approaches for Land Cover classification
Chen, Yuxin
2024/2025
Abstract
With the rapid development of remote sensing and geographic information systems, accurate land cover classification has become essential for ecological monitoring and land use management. This study evaluates three classification algorithms—KMeans, Random Forest, and Multi-Layer Perceptron—using Sentinel-2 multispectral imagery across three regions: urbanized Shanghai, the agricultural hills of Tuscany, and the alpine area of Val d'Aosta. A standardized workflow was implemented, including image preprocessing, manual annotation, feature extraction, model training, result visualization, and accuracy assessment. The results show that supervised models clearly outperform the unsupervised KMeans algorithm. Among them, MLP achieved the highest accuracy and spatial consistency, producing coherent and meaningful land cover maps. Random Forest also performed well in distinguishing spectrally distinct classes but was less effective in refining boundaries and representing minority categories. KMeans, by contrast, only captured coarse patterns and showed high levels of misclassification. The comparison of sampling strategies further revealed that uniform sampling improved recognition of minority classes and reduced imbalance, strengthening model robustness. Overall, the findings indicate that MLP with uniform sampling provides the most effective framework for diverse land cover classification, offering guidance for large-scale remote sensing applications and ecological assessment. The thesis is organized into seven chapters. Chapter 1 outlines the Earth Observation background, research objectives, and the Sentinel-2 dataset. Chapter 2 introduces the Google Earth Engine platform, describing data catalogs, classifier implementations, and evaluation metrics. Chapter 3 presents the study areas and training samples, while Chapter 4 details the preprocessing workflow, including cloud masking, band selection, and spectral index computation. Chapter 5 reports the classification results for KMeans, Random Forest, and Multi-Layer Perceptron across the study regions. Chapter 6 compares proportional and uniform sampling strategies, highlighting their impact on classification robustness. Finally, Chapter 7 synthesizes the overall results, emphasizing model performance and the role of sampling strategies, and provides recommendations for remote sensing applications and ecological assessments.| File | Dimensione | Formato | |
|---|---|---|---|
|
Comparison between different approaches for Land Cover classification_YuxinChen.pdf
non accessibile
Dimensione
20.24 MB
Formato
Adobe PDF
|
20.24 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/243145