|
A Comparative Analysis of Machine Learning Algorithms for Monitoring Land Cover Changes (Case Study: Eyvan County, Ilam)
|
Mahboobe Gholami   , Ali Mahdavi * |
|
|
|
Abstract: (383 Views) |
Extended Abstract
Introduction
Monitoring land-use/land-cover (LULC) changes is crucial for sustainable management of fragile ecosystems such as Iran's western Zagros forests. These forests, covering about 5 million hectares, are of great importance for soil conservation, hydrological balance, biodiversity, and carbon storage. However, they are increasingly vulnerable to the negative impacts of population growth, agricultural encroachment, over grazing, illegal logging and recurrent drought. These issues are all evident in central Zagros Eyvan County (Eyvan-e Gharb) of Ilam Province. This study compares the performance of two machine-learning algorithms; Artificial Neural Network (ANN) and K-Nearest Neighbors (KNN) for classifying and monitoring LULC changes over a 20-year period (2000–2020) using multi-temporal Landsat imagery. By integrating 14 spectral indices with field-validated training data, the research quantifies spatiotemporal dynamics of five major classes (water, orchard, forest, rangeland, and cropland) and evaluates which algorithm better captures the complex, non-linear interactions typical of Zagros landscapes. The findings provide a scientifically robust basis for evidence-based conservation planning and highlight the superiority of non-linear models in ecologically heterogeneous and data-limited environments. This work addresses a key knowledge gap by systematically comparing ANN and KNN specifically within the Zagros context, where few studies have evaluated algorithm suitability under similar topographic and anthropogenic pressures.
Materials and Methods
Landsat TM (2000, 2010) and OLI (2020) images (path/row 168/37) were acquired from the USGS EarthExplorer database. All scenes were pre-processed in ENVI 5.3, including geometric correction to sub-pixel accuracy using ground control points and topographic maps, followed by atmospheric correction via the Dark Object Subtraction (DOS) method. To enhance classification accuracy, a total of 14 spectral indices were calculated, including NDVI, SAVI, NDWI, EVI, GNDVI, RVI, DVI, GDVI, TVI, RDVI, OSAVI, NormR, NormNir, and NormG. Five land cover classes were identified: water, orchard, forest, rangeland, and cropland. Training and validation samples (795 samples in total) were collected using GPS-based field surveys with sub-3-meter accuracy, Google Earth imagery, and visual interpretation of satellite data. Two machine learning algorithms were implemented for supervised classification. The KNN algorithm, with K=15 determined through trial and error and Euclidean distance as the distance metric, was selected for its simplicity and interpretability. The ANN algorithm employed a multilayer perceptron with backpropagation, featuring an input layer corresponding to input bands, a learning rate of 0.1, 1000 iterations, and sigmoid activation function. The models were trained using 70% of the samples (557 samples), with the remaining 30% (238 samples) reserved for accuracy assessment. Model performance was evaluated using overall accuracy and Kappa coefficient derived from error matrices. Image processing and data analysis were conducted using TerrSet 2020 and ArcGIS 10.8 software. All spatial layers were standardized to ensure consistency and minimize potential bias in the analysis.
Results and Discussion
Accuracy assessment revealed clear superiority of the ANN algorithm. In 2020, ANN achieved the highest performance (OA = 83 %, Kappa = 0.78), outperforming KNN (OA = 74 %, Kappa = 0.63). ANN also performed better in 2000 (OA = 74 %, Kappa = 0.67) than KNN (OA = 72 %, Kappa = 0.61), while KNN was marginally superior in 2010 (OA = 69 %, Kappa = 0.56 vs. ANN OA = 62 %, Kappa = 0.51). The 2020 ANN error matrix showed excellent water classification (100 % producer’s accuracy) and strong performance for forest and cropland, with main confusions occurring between orchard–forest and rangeland–cropland. Post-classification change detection using the superior ANN model indicated a 9 % net loss of forest cover (34 277 ha → 31 125 ha) between 2000 and 2020. Cropland expanded by 23 % and orchards by 18 %, reflecting conversion of natural vegetation to agriculture. Water bodies increased overall by 23 %, primarily due to impoundment of the Kangir Dam (notable +68 % rise 2000–2010, followed by a 27 % decline 2010–2020 linked to drought and increased downstream withdrawals). In contrast, KNN produced more conservative estimates and, in some periods, contradictory trends (e.g., 47 % cropland increase and opposing rangeland dynamics). These discrepancies underscore ANN’s superior ability to model non-linear spectral relationships in heterogeneous Zagros terrain.
Conclusion
This study demonstrates the clear advantage of Artificial Neural Networks over K-Nearest Neighbors for LULC classification and change detection in the complex Zagros landscape. The ANN model delivered consistently higher accuracies (peak OA 83%, Kappa 0.78) and more reliable trend quantification, effectively capturing subtle and abrupt changes that the simpler KNN algorithm underestimated or misrepresented. Over two decades, Eyvan County experienced a concerning 9 % decline in forest cover accompanied by 23% and 18% expansions in cropland and orchards, respectively. These shifts are driven by population growth, agricultural encroachment into forest margins, overgrazing that prevents regeneration, fuelwood collection, and climate-induced stresses (successive droughts and altered precipitation). The observed 23% net increase in water bodies is largely attributable to the Kangir Dam, though subsequent reductions highlight the interplay of dam operation, drought, and rising irrigation demand. The substantial divergence between ANN and KNN results, particularly the 47% cropland increase reported by KNN versus 23% by ANN, highlights the critical importance of algorithm selection in ecologically sensitive regions. ANN’s non-linear modeling capacity and lower sensitivity to class imbalance make it markedly more suitable for Zagros-type environments. The approach validates the integration of Landsat imagery, spectral indices, and machine learning as a powerful, cost-effective tool for long-term monitoring. Future studies should incorporate higher-resolution data (e.g., Sentinel-2), ensemble methods, and deeper networks to further enhance accuracy and enable predictive modeling. These findings provide a solid scientific foundation for developing targeted conservation strategies and sustainable land-management policies to safeguard the Zagros forests. |
Article number: ۵ |
| Keywords: Satellite imagery, Artificial Neural Network, Land use, Landsat satellite, Land management |
|
|
|
Type of Study: Research |
Subject:
Special
Received: 2025/10/29 | Revised: 2026/06/13 | Accepted: 2026/02/15 | Published: 2026/02/15 | ePublished: 2026/02/15
|
|
|
|
|
|
|
| Send email to the article author |
|
|
|
| Add your comments about this article |
|
|
|
|
gholami M, Mahdavi A. A Comparative Analysis of Machine Learning Algorithms for Monitoring Land Cover Changes (Case Study: Eyvan County, Ilam). jwmseir 2025; 19 (71) : ۵ URL: http://jwmsei.ir/article-1-1216-en.html
|
