Machine learning based combinatorial analysis for land use and land cover assessment in Kyiv City (Ukraine)

Abstract

The main goal of this study is to evaluate different models for further improvement of the accuracy of land use and land cover (LULC) classification on Google Earth Engine using random forest (RF) and support vector machine (SVM) learning algorithms. Ten indices, namely normalized difference vegetation index, normalized difference soil index, index-based built-up index, biophysical composition index, built-up area extraction index (BAEI), urban index, new built-up index, band ratio for built-up area, bare soil index, and normalized built up area index, were used as input parameters for the machine learning algorithms to improve classification accuracy. The combinatorial analysis of the Sentinel-2 bands and the aforementioned indices allowed us to create four combinations based on surface reflectance characteristics. The study includes data from April 2020 to September 2021 and April 2022 to June 2022. The multitemporal Sentinel-2 data with spatial resolutions of 10 m were used to determine the LULC classification. The major land use classes such as water, forest, grassland, urban areas, and other lands were obtained. Generally, the RF algorithm showed higher accuracy than the SVM. The overall accuracy for RF and SVM was 86.56% and 84.48%, respectively, and the mean Kappa was 0.82 and 0.79, respectively. Using the combination 2 with the RF algorithm and combination 4 with the SVM algorithm for LULC classification was more accurate. The additional use of vegetation indices allowed to increase in the accuracy of LULC classification and separate classes with similar reflection spectra.

Temporal Variations in Land Surface Temperature within an Urban Ecosystem: A Comprehensive Assessment of Land Use and Land Cover Change in Kharkiv, Ukraine

“…To do this, we created additional samples: 376 points, evenly distributed over the image Table 2. In the classification process, our methodology extended beyond the mere application of spectral bands [111]. It included a comprehensive analysis of multiple vegetative indices, including NDVI, NDBSI (Normalized Difference Bare Soil Index), BAEI (Built-up Area Extraction Index), NDWI (Normalized Difference Water Index), NDBI (Normalized Difference Built-up Index), BRBA (Band Ratio for Built-up Area), NBAI (Normalized Built up Area Index), IBI (Index-Based Built-up Index), NBI (New Built-up Index), and UI (Urban Index) to enhance the differentiation of spectral similarities among classes [112][113][114][115][116].…”

“…Temporal classification analysis of Landsat-derived land cover datasets from 1984 to 2023 revealed significant changes in LULC, highlighted by high classification accuracy for water, urban, and vegetation classes. Notably, water classification achieved an exceptional accuracy of 99%, while urban areas showed high accuracy, likely improved by the inclusion of additional spectral indices [10,18,46,68,70,111]. Each vegetative index plays a distinct role in differentiating the classes: NDVI is essential for identifying dense vegetation and sparse vegetation by measuring the health and density of vegetation [11,16,18,31].…”

“…The use of the RF algorithm was pivotal in achieving high classification accuracy, especially when comparing the thermal regimes of different LULC classes [20,45,68]. This method's efficacy is juxtaposed with other classifiers like the support vector machine [8,9,23,26,36] and maximum likelihood classifier [11,16,21,24,29], and these algorithms, which, despite their advantages, exhibit varying degrees of effectiveness in LULC classification.…”

“…Many studies are specifically dedicated to comparing the performance of existing methods for obtaining and improving LST data [61][62][63][64]66]. The main statistical approaches and methods commonly used in the study of LST and especially LULC interdependence are supervised and unsupervised techniques [9,20,25,[67][68][69][70]; Mann-Kendall statistics [13,71]; principal component analysis end ordinary least squares [72,73]; cellular-automata [21,33,48,74] and most widely used linear and multiple linear regression analyses [9,11,15,28,[33][34][35][36][37][38][39]49,57,76]. Particular attention is merited by studies focused on establishing linear and nonlinear dependencies between LST, UHI effects, and various vegetation indices [31,67,[77][78][79][80][81][82][83].…”

Evaluation of different machine learning algorithms for LULC classification in heterogeneous landscape by using remote sensing and GIS techniques

“…GEE allows users to evaluate all freely available high volume LANDSAT and Sentinel data without having to download them to their local Personal computer [3]. The advantage of GEE includes parallel processing environment with high volume of data [4].…”