Land-Cover and Land-Use Mapping Under Limited Data Highlights Hyperparameter Stability and Predictor Design

Abstract

Land-cover (LC) and land-use (LU) mapping is essential for environmental monitoring, yet supervised multi-decadal classification remains constrained by heterogeneous agroforestry mosaics, uneven observation quality, and limited consistent reference data. This study presents a Landsat-based workflow for robust multi-temporal LC/LU classification, combining quality-focused preprocessing, radiometric harmonization, predictor-set evaluation, and Random Forest (RF) hyperparameter-stability assessment. Implemented in south-central Chile, the workflow combines masks with a locally calibrated cloud–snow index and applies per-band histogram matching for color balancing. It also tests progressive predictor designs (seasonal spectral bands, spectral indices, and territorial variables) under sample scarcity. Classification used Google Earth Engine and an RF grid search over hyperparameter ranges. Performance was evaluated through validation-kappa (κ) distributions and a hyperparameter-dispersion metric assessing accuracy and stability. Color balancing improved cross-period consistency and can partly offset the absence of spectral indices, but high performance was achieved only with the full predictor set, even under limited observations. High-performing treatments showed lower hyperparameter dispersion, supporting model selection that jointly considers accuracy and stability rather than one tuned configuration. Under limited reference data, classification depends more on predictor design and robustness-based selection than on increasing sample size alone.