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

Highlights

What are the main findings?

• Low-performing treatments exhibit greater hyperparameter variability, whereas high-performing treatments show greater stability.
• The full predictor set yields the best performance, with limited gains from larger sample sizes or radiometric harmonization.

What are the implications of the main findings?

• Model selection should evaluate both accuracy and stability, rather than focusing solely on peak accuracy.
• Predictor design—especially territorial variables—is a more decisive factor for performance than sample size.

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.

1. Introduction

Land cover (LC) and land use (LU) are core inputs for environmental monitoring and land-management decisions because they provide spatially explicit representation of interactions between human activities and ecosystem processes across space and time [1,2]. Despite advances in Earth-observation archives, open data, machine-learning workflows, and cloud processing, supervised multi-decadal LC/LU classification at national-to-regional scales remains constrained by heterogeneous mosaics and ambiguous class boundaries. It is also limited by the uneven availability of high-quality observations due to clouds and atmospheric effects and by temporally inconsistent reference labels from multi-source inventories [1,2,3,4]. These limitations are particularly pronounced in agroforestry systems, where agriculture, plantation forestry, livestock grazing, native vegetation, and wetlands coexist at fine scales and often share spectral and structural signatures, making class separation and change attribution sensitive to methodological choices. South-central Chile is suitable because its landscapes have undergone substantial LC/LU change, including plantation expansion, agricultural reconfiguration, and pressure on native vegetation [5,6].

The Landsat archive supports multi-decadal LC/LU mapping [1,2], while Google Earth Engine (GEE) enables scalable machine-learning processing [3]. However, robust mapping depends less on nominal data volume than on extracting stable, comparable signals across seasons, atmospheric conditions, and sensors [2,7]. Even atmospherically corrected products can retain contamination, radiometric distortions, and non-random gaps that compromise temporal comparability and create apparent class differences unrelated to real LC/LU variation [4,8]. These constraints motivate workflows that treat observation quality and cross-scene consistency as primary determinants of classification reliability.

Cloud, cloud-shadow, snow, water, and saturation masking is fundamental. However, predefined masks can produce omission and commission errors, particularly in humid or mountainous areas where clouds, snow, haze, and thin cloud are difficult to separate [4,8,9]. Because undetected contamination can introduce extreme reflectance values, bias composites, and distort feature space, conservative masking is often preferable when temporal stability is prioritized over pixel retention [4]. Masking performance also depends on local tuning to regional illumination, LC conditions, and seasonality [10,11]. Semi-transparent aerosols and haze further affect imagery, while multi-sensor differences in spectral response and radiometric calibration create additional inconsistencies [2]. Together, these effects represent measurement-scale mismatch that shifts or rescales reflectance distributions and may cause models to learn scene- or sensor-specific artifacts, reducing cross-period generalization and inflating spurious change. Histogram matching and related distribution-based methods are therefore used to reduce residual discontinuities [12,13,14]. Accordingly, the workflow combines predefined masks with a locally calibrated cloud–snow index, prioritizing contaminant exclusion even when valid pixels are discarded, and per-band histogram matching to align target images with radiometric-reference mosaics while maintaining temporal coverage.

Training data also constrain historical LC/LU reconstruction because labels often come from inventories with inconsistent legends and temporal definitions, requiring harmonization and stability assumptions that add uncertainty [7]. Class frequencies are usually imbalanced, biasing models toward dominant, stable classes and reducing sensitivity to rare classes and transitions central to LC/LU-change assessment [15]. Yet perfect balance may reduce accuracy by discarding abundant observations or oversampling rare classes and introducing artificial dependence [7,16,17]. Sample imbalance should therefore be diagnosed explicitly and incorporated into sampling design.

Predictor design further affects separability in heterogeneous agroforestry mosaics. Multi-season spectral bands describe surface properties, while indices target vegetation vigor, moisture, burn signals, and soil background [18]. Their benefits are context-dependent because indices can be redundant with raw bands, sensitive to residual atmospheric effects, and unstable when masking or compositing creates seasonal or spatial gaps. Their marginal value should therefore be assessed against the original bands rather than inferred from established ecological interpretations. Territorial variables such as topography, climate, accessibility, and edaphic proxies can add stable spatial context where spectral confusion persists [19]. They complement rather than replace spectral information by encoding landscape position and environmental gradients linked to vegetation, management, and disturbance. Yet they may also impose stable spatial priors that favor persistent patterns. However, larger predictor sets increase dimensionality, computational demand, and overfitting risk when effective per-class samples are limited or imbalanced [7]. Progressive predictor sets can therefore balance information gains against model complexity, especially when historical reference observations are sparse.

Random Forest (RF) is widely used for LC/LU mapping because it captures nonlinear class boundaries, handles mixed predictors, and tolerates noise and collinearity in high-dimensional feature spaces [20,21]. Its performance nevertheless depends on hyperparameters controlling tree structure, feature subsampling, and ensemble size. These settings affect the bias–variance trade-off and predictor-importance stability, with effects that vary by predictor dimensionality, class structure, and sampling design; optimal values are therefore context-dependent [20,22,23]. Remote-sensing outcomes can therefore shift with parameterization and predictor selection, especially under ambiguous classes or limited training data [21,22]. Evaluation should move beyond “single best model” tuning toward structured exploration of plausible hyperparameter ranges, allowing treatment effects to be identified as patterns that persist across configurations.

Existing studies thus face three broad limitations. First, uneven observation quality, residual radiometric mismatch, inconsistent reference labels, and class imbalance reduce the reliability of multi-decadal mapping. Second, spectral indices or auxiliary variables are often added without controlled assessment of their marginal contribution, obscuring informative, redundant, and context-dependent predictors. Third, evaluation commonly relies on fixed or narrowly tuned RF settings and single accuracy estimates, while stability across preprocessing, sampling, predictor-set, and hyperparameter choices is rarely quantified. The combined effects of radiometric harmonization, sample scarcity, predictor design, and RF sensitivity therefore remain insufficiently understood for multi-decadal LC/LU mapping in agroforestry regions.

To address this gap, this study evaluates an operational workflow for supervised multi-temporal LC/LU mapping under imperfect data conditions by quantifying methodological choices as performance gains. The preprocessing techniques are established rather than individually novel, but their combined implementation was designed for the study’s requirements. The workflow integrates Landsat winter and summer observations, predefined masks, a locally calibrated cloud–snow index, reference-mosaic generation, and per-band histogram matching for radiometric harmonization.

The study makes three core scientific contributions. First, it integrates these preprocessing steps into a reproducible multi-decadal LC/LU classification workflow. Second, it provides a controlled evaluation of three treatment axes: radiometric harmonization (color balancing), progressive predictor-set composition (seasonal bands, spectral indices, and territorial variables), and training-sample size under reference-data scarcity. Third, it assesses RF stability through distributions of the kappa statistic (κ) across a structured hyperparameter space rather than through a single optimized configuration. The practical value of the workflow lies in its supporting moderate-resolution LC/LU mapping and environmental monitoring in heterogeneous agroforestry landscapes where reference data are temporally uneven, spatially incomplete, or limited.

2. Materials and Methods

2.1. Study Area

The 50,490 km² study area is in south-central Chile and includes the provinces of Ñuble, Biobío, and Malleco, excluding the island (Figure 1). The region spans the Coastal Range, Central Depression, and Andes and includes Mediterranean, humid temperate, and mountain climates; the Itata and Biobío rivers have pluvio-nival regimes [24,25,26,27]. During the last four decades, the landscape has shifted from predominantly agricultural land and degraded native vegetation toward expanded forest plantations and grazing areas, while agriculture has declined and intensified locally [5,28,29,30,31]. Current LU includes rainfed and irrigated agriculture, forestry, livestock production, native forest, scrubland, natural meadows, and wetlands, several of which support multiple LU functions [6,29,31,32].

2.2. Individual Images and Temporal Design

Landsat imagery was used because it combines medium spatial resolution with a long-term historical archive [2]. The study area is covered by nine scenes from paths 232, 233, and 1 and rows 85, 86, and 87. Level-2 surface-reflectance collections for Landsat 5, 7, and 8 (L5, L7, and L8) were obtained from GEE [3]: LT05_C01_T1_SR, LE07_C01_T1_SR, and LC08_C01_T1_SR [3,33,34,35,36]. Predictors included blue ($Bb$), green ($Bg$), red ($Br$), near-infrared ($Bnir$), shortwave-infrared ($Bswir1$ and $Bswir2$), and brightness-temperature ($Bbt$) bands. Wavelengths and access links are provided in Appendix A.1,

Table A1. Surface-reflectance and brightness-temperature bands.

Code	Description	L5 and L7 Wavelength (μm)	L8 Wavelength (μm)
$Bb$	Blue band	0.45–0.52	0.452–0.512
$Bg$	Green band	0.52–0.60	0.533–0.590
$Br$	Red band	0.63–0.69	0.636–0.673
$Bnir$	Near-infrared band	0.77–0.90	0.851–0.879
$Bswir1$	Shortwave-infrared band 1	1.55–1.75	1.566–1.651
$Bswir2$	Shortwave-infrared band 2	2.08–2.35	2.107–2.294
$Bbt$	Brightness-temperature band	10.4–12.5	10.60–12.51

L5, L7, L8 = Landsat 5, 7, 8; $Bbt$ = average of two bands.

.

The 1997–2020 study period was divided into 12 biennial periods (P1–P12) defined by hydrological years. Winter and summer windows were extended into adjacent seasons to improve seasonal coverage. For P1, both windows covered 183 days: winter, day of year (DOY) 126–309; summer, DOY 310–125. For P2–P12, each covered 137 days: winter, DOY 150–287; summer, DOY 333–102. Separate winter and summer median mosaics were generated for each period, and pixels with insufficient valid observations were removed. All processing used a 30 m grid.

Image availability varied by season, period, cloud threshold, and sensor history (

Table 1. Landsat image availability and potential captures by period, season, and maximum-cloudiness threshold.

Period or Statistic	Landsat Satellites	Overpasses per Season	Captures per Season	Winter 10%	Winter 75%	Winter 100%	Summer 10%	Summer 75%	Summer 100%	Winter Capture Rate (%)	Summer Capture Rate (%)
P1 1997–1998 †	5	23	206	14	35	40	45	77	83	19	40
P2 1999–2000	5, 7	34	308	29	74	88	117	168	175	29	57
P3 2001–2002	5, 7	34	308	34	125	153	111	168	183	50	59
P4 2003–2004	5, 7 ‡	34	308	41	112	132	160	258	265	43	86
P5 2005–2006	5, 7 ‡	34	308	39	157	172	133	208	218	56	71
P6 2007–2008	5, 7 ‡	34	308	57	125	136	175	237	243	44	79
P7 2009–2010	5, 7 ‡	34	308	34	86	98	118	201	217	32	70
P8 2011–2012	5 §, 7 ‡, 8 §	34	308	41	97	103	88	151	156	33	51
P9 2013–2014	7 ‡, 8	34	308	62	203	235	210	284	294	76	95
P10 2015–2016	7 ‡, 8	34	308	48	177	225	164	268	284	73	92
P11 2017–2018	7 ‡, 8	34	308	67	184	217	171	261	280	70	91
P12 2019–2020	7 ‡, 8	34	308	67	176	206	171	265	281	67	91
Total		400	3597	533	1550	1805	1663	2545	2679	592	882

† = more extended period; ‡ = L7 scan-line gaps; § = few images. P8 assumes two satellites because L7 was fully available and L5 and L8 operated partially. Overpasses = estimated acquisition dates under 16-day repeat cycle; Captures = estimated scene captures assuming nine scenes per overpass; Capture rate = images at 100% cloudiness/Captures; L5, L7, L8 = Landsat 5, 7, 8.

). Approximately 3597 scene captures were potentially available. Under the 10% cloud threshold, 533 winter and 1663 summer images were retained; under the 75% threshold, 1550 winter and 2545 summer images were retained. The record also included the addition of L7, the post-2003 L7 scan-line failure, reduced sensor availability in P8, and the incorporation of L8. Median mosaics required repeated spatial coverage equivalent to approximately 75–125 images per period. Because no winter period met this criterion under the 10% threshold and P1 also lacked sufficient summer coverage, the P1 seasonal windows were extended, and all available images were used.

2.3. Homogenization and Cloud/Snow/Water Masking

Cross-sensor preprocessing first standardized the Landsat inputs. The two L8 thermal bands were averaged, and post-2003 L7 scan-line gaps were filled by local-mean extrapolation from one-pixel neighborhoods [37]. Predefined quality filters removed pixels affected by atmospheric opacity, clouds, cloud shadows, adjacent clouds, snow, water, or saturation [4,36]. Because cloud–snow separation remains difficult [8,9], a cloud–snow index ($CsInd$), adapted from the Simple Cloud Score [10,11], was calculated as the mean of five locally normalized sub-indices:

$$CsInd=1/5·{\sum }_q^5\left({Ind}_q-\mathrm{R}\mathrm{e}\mathrm{f}\mathrm{M}\mathrm{i}\mathrm{n}\left({Ind}_q\right)\right)/\left(\mathrm{R}\mathrm{e}\mathrm{f}\mathrm{M}\mathrm{a}\mathrm{x}\left({Ind}_q\right)-\mathrm{R}\mathrm{e}\mathrm{f}\mathrm{M}\mathrm{i}\mathrm{n}\left({Ind}_q\right)\right)$$

(1)

Higher values indicated a greater probability of cloud or snow. $\mathrm{R}\mathrm{e}\mathrm{f}\mathrm{M}\mathrm{i}\mathrm{n}$ and $\mathrm{R}\mathrm{e}\mathrm{f}\mathrm{M}\mathrm{a}\mathrm{x}$ defined local minimum and maximum reference values and were estimated separately for each sub-index over the study region. The sub-indices ${Ind}_q$ are:

• ${Ind}_1=Bb$ (clouds are bright in blue band);
• ${Ind}_2=Bb+Bg+Br$ (clouds are bright in the visible range);
• ${Ind}_3=Bnir+Bswir1+Bswir2$ (clouds are bright in the infrared range);
• ${Ind}_4=\mathrm{R}\mathrm{e}\mathrm{f}\mathrm{M}\mathrm{i}\mathrm{n}\left(Bbt\right)+\mathrm{R}\mathrm{e}\mathrm{f}\mathrm{M}\mathrm{a}\mathrm{x}\left(Bbt\right)-Bbt$ (clouds are generally colder);
• ${Ind}_5=\left(Bg-Bswir1\right)/\left(Bg+Bswir1\right)$ (normalized snow index).

Two full-period mosaics, $Winter\_97\_20$ and $Summer\_97\_20$, were created from scenes with less than 10% metadata cloudiness using median compositing; pixels represented by fewer than seven images were discarded. $Winter\_97\_20$ was used to calibrate one regional $\mathrm{R}\mathrm{e}\mathrm{f}\mathrm{M}\mathrm{i}\mathrm{n}$ and $\mathrm{R}\mathrm{e}\mathrm{f}\mathrm{M}\mathrm{a}\mathrm{x}$ value per sub-index. A single $CsInd$ threshold was then selected by visually comparing masked and unmasked versions of $Winter\_97\_20$, prioritizing removal of cloud- or snow-contaminated pixels over retention of ambiguous pixels. The global mask ($Gm$) was derived from $CsInd\left(Winter\_97\_20\right)$ by removing values above the selected threshold and isolating low-coverage patches smaller than 2.7 ha. $Winter\_97\_20$, $Summer\_97\_20$, $Gm$, and the calibrated $CsInd$ were retained for subsequent image and mosaic processing.

2.4. Radiometric Harmonization by Histogram Matching

Per-band histogram matching, here termed color balancing, was used to reduce residual cross-sensor and cross-period radiometric differences [12,13,38,39,40]. This method defined transformation functions to map surface-reflectance values ($SR$) to cumulative histograms ($H$) and vice versa. For each target band $B_{Tar}$, the forward function $f$ is defined as $H_{Tar}=f\left({SR}_{Tar}\right)$; for a reference band $B_{Ref}$, the backward function $g$ is ${SR}_{Ref}=g\left(H_{Ref}\right)$. Both $f$ and $g$ were modeled as continuous, non-monotonic curves using RF regression—a secondary application distinct from the main RF described later. Applying $f$ to the $B_{Tar}$ yields a cumulative-histogram band, ${BH}_{Tar}=f\left(B_{Tar}\right)$. The color-balanced target band is then obtained by applying $g$ to align it with the reference histogram:

$$B_{Tar,Ref}=g\left({BH}_{Tar}\right)=g\left(f\left(B_{Tar}\right)\right)$$

(2)

$Summer\_97\_20$ masked by $Gm$ formed the global summer reference, $SumRef\_97\_20$, whereas $Winter\_97\_20$ was first masked by $Gm$ and then balanced against $SumRef\_97\_20$ to form the global winter reference, $WinRef\_97\_20$. The radiometric-reference mosaics were centered near 2011 rather than the 2009.3 midpoint.

Winter and summer mosaics were also produced for four six-year intervals: 1997–2002, 2003–2008, 2009–2014, and 2015–2020. They were balanced against the global radiometric references and served as partial references for shorter periods. Individual images were balanced only over pixels overlapping the appropriate partial reference, providing spatially adaptive matching. Hierarchical reference construction, local overlap matching, and RF-regression implementation are reported in Appendix A.1, Figure A1 and Figure A2.

2.5. Biennial Mosaics

For each biennial period, winter and summer collections were filtered to less than 75% metadata cloudiness (Figure 2a, step 1). This threshold balanced spatial coverage and computation; higher thresholds added mostly unusable images. Each image was masked using the predefined quality masks, $Gm$, and the corresponding $CsInd$ mask (step 2). Images with less than 5000 ha of unmasked area or average patch size were excluded because they did not support reliable histogram matching.

2.5.1. Color-Balancing Treatment Axis (Bal)

Color-unbalanced mosaics (“No” level) were produced as per-band medians without radiometric adjustment (steps 3–4). Color-balanced mosaics (“Yes” level) were generated by balancing each image against its partial reference before compositing, calculating the per-band median, and balancing the complete mosaic once more against the corresponding reference (steps 5–7).

2.5.2. Predictor-Set Treatment Axis (Var)

The winter and summer bands were supplemented with six seasonal indices and eight territorial variables (steps 8–10). The indices represented vegetation features ($Evi$, $Gci$, and $Sipi$) and LC-related features ($Bsi$, $Nbri$, and $Ndwi$). Territorial predictors represented terrain ($AspectE$, $AspectN$, $Chil$, and $Mtpi$), climate ($Prep06$, and $Temp02$), surface flow ($WbInd$), and human accessibility ($AccesCity$). Definitions, equations, sources, and access links are reported in

Table A2. Spectral indices.

Type	Code	Description	Equation	Scale	Source
Vegetation	$Evi$	Enhanced vegetation index	$2.5·\left(Bnir-Br\right)/\left(Bnir+6·Br+7.5·Bb+1\right)$	≈(−1, 1)	[76]
	$Gci$	Green chlorophyll index	$Bnir/Bg-1$	[−1, ∞)	[77]
	$Sipi$	Structure-insensitive pigment index	$\left(Bnir-Bb\right)/\left(Bnir+Br\right)$	≈(−1, 1)	[78]
Other LC	$Bsi$	Bare soil index	$\left(Bswir1+Br-Bnir-Bb\right)/\left(Bswir1+Br+Bnir+Bb\right)$	≈(−1, 1)	[79]
	$Nbri$	Normalized burned ratio index	$\left(Bnir-Bswir2\right)/\left(Bnir+Bswir2\right)$	[−1, 1]	[80]
	$Ndwi$	Normalized difference water index	$\left(Bg-Bnir\right)/\left(Bg+Bnir\right)$	[−1, 1]	[81]

Band codes follow Table A1; all indices are dimensionless.

and

Table A3. Territorial variables.

Type	Code	Description	Resolution	Scale/Unit	Source	Direct Link
Terrain	$AspectE$	East aspect multiplied by slope	30 m	[−1, 1]	[82,83] †	https://developers.google.com/earth-engine/datasets/catalog/JAXA_ALOS_AW3D30_V3_2
	$AspectN$	North aspect multiplied by slope	30 m	[−1, 1]	[82,83] †	https://developers.google.com/earth-engine/datasets/catalog/NASA_NASADEM_HGT_001
	$Chil$	Continuous heat-insolation load index	90 m	[−1, 1]	[84] ‡	https://developers.google.com/earth-engine/datasets/catalog/CSP_ERGo_1_0_Global_ALOS_CHILI
	$Mtpi$	Multi-scale topographic position index	270 m	[−1, 1]	[84] ‡	https://developers.google.com/earth-engine/datasets/catalog/CSP_ERGo_1_0_Global_ALOS_mTPI
Climate	$Prep06$	June average monthly precipitation	1 km	mm	[85]	https://developers.google.com/earth-engine/datasets/catalog/OpenLandMap_CLM_CLM_PRECIPITATION_SM2RAIN_M_v01
	$Temp02$	February average monthly temperature	1 km	°C	[86] §	https://developers.google.com/earth-engine/datasets/catalog/OpenLandMap_CLM_CLM_LST_MOD11A2-DAY_M_v01
Surface flow	$WbInd$	Water-body distance index	180 m	[0, 1]	[87,88] †	https://geoportal.cl/geoportal/catalog/35252/Catastro%20de%20Lagos https://geoportal.cl/geoportal/catalog/34955/Humedales
Human	$AccesCity$	Accessibility to cities (travel time)	928 m	Minute	[89]	https://developers.google.cn/earth-engine/datasets/catalog/Oxford_MAP_accessibility_to_cities_2015_v1_0

† = calculated using the corresponding equation above; ‡ = normalized to 0–1; § = scaled to °C.

(Appendix A.1).

Each predictor $X$ was normalized to the 0–1 range according to: $X_{\mathrm{n}\mathrm{o}\mathrm{r}\mathrm{m}\mathrm{a}\mathrm{l}\mathrm{i}\mathrm{z}\mathrm{e}\mathrm{d}}=\left(X-\mathrm{m}\mathrm{i}\mathrm{n}\mathrm{i}\mathrm{m}\mathrm{u}\mathrm{m}\left(X\right)\right)/\left(\mathrm{m}\mathrm{a}\mathrm{x}\mathrm{i}\mathrm{m}\mathrm{u}\mathrm{m}\left(X\right)-\mathrm{m}\mathrm{i}\mathrm{n}\mathrm{i}\mathrm{m}\mathrm{u}\mathrm{m}\left(X\right)\right)$. For bands and indices, minimum and maximum values were calculated across periods to preserve temporal comparability; for invariant territorial variables, they were calculated by variable. This produced 12 biennial mosaics each containing 34 predictor layers (

Table 2. Independent variables.

Winter Band	Summer Band	Winter Index	Summer Index	Territorial
$Bb\_Win$	$Bb\_Sum$	$Evi\_Win$	$Evi\_Sum$	$AspectE$
$Bg\_Win$	$Bg\_Sum$	$Gci\_Win$	$Gci\_Sum$	$AspectN$
$Br\_Win$	$Br\_Sum$	$Sipi\_Win$	$Sipi\_Sum$	$Chil$
$Bnir\_Win$	$Bnir\_Sum$	$Bsi\_Win$	$Bsi\_Sum$	$Mtpi$
$Bswir1\_Win$	$Bswir1\_Sum$	$Nbri\_Win$	$Nbri\_Sum$	$Prep06$
$Bswir2\_Win$	$Bswir2\_Sum$	$Ndwi\_Win$	$Ndwi\_Sum$	$Temp02$
$Bbt\_Win$	$Bbt\_Sum$			$WbInd$
				$AccesCity$

Bands: $Bb$, $Bg$, $Br$ = blue, green, red; $Bnir$ = near-infrared; $Bswir1$, $Bswir2$ = shortwave-infrared 1, 2; $Bbt$ = brightness-temperature. Indices: $Evi$, $Gci$, $Sipi$ = enhanced vegetation, green chlorophyll, structure-insensitive pigment; $Bsi$, $Nbri$, $Ndwi$ = bare soil, burned ratio, difference water. $Win$, $Sum$ = winter, summer. $AspectE$, $AspectN$ = east-, north-aspects; $Chil$ = continuous heat-insolation load; $Mtpi$ = multi-scale topographic position; $Prep06$ = June precipitation; $Temp02$ = February temperature; $WbInd$ = water-body index; $AccesCity$ = city accessibility.

). Three nested predictor sets were used: B, seasonal bands; BI, B plus seasonal indices; and BIT, BI plus territorial variables.

Mean valid temporal coverage across the 12 mosaics was 85.5% and nearly complete in the core agroforestry area. Most missing data formed high-elevation fringes associated with persistent winter cloud or snow; fixed gaps also occurred over large water bodies and urban areas. Coastal limitations reflected the 1 km resolution of climatic predictors. Remaining gaps were addressed during post-classification processing in Section 2.9.

2.6. LC Reference Data and Observation Sampling

LC reference observations were extracted from Land-Use Change Monitoring polygons produced by the National Forest Corporation (CONAF) [41,42,43] (Figure 2b, step 11). Three nested legends served different purposes: C0 comprised 27 detailed LC classes used for classification; C1 pooled them into 12 classes for sample balancing; and C2 aggregated them into seven LU classes for transition analysis. Harmonization rules, category definitions, dataset links, and temporal references are in

Table A5. Land-cover (LC) and land-use (LU) categories.

Code 0: C0.	Unpooled LC category	Code 1: C1.	Pooled LC Category	Code 2: C2.	Pooled LU Category
1	Permanent agricultural	1	Permanent agricultural	1	Agricultural
2	Rotating agricultural	2	Rotating agricultural
3	Annual grassland	3	Improved agricultural grassland
4	Perennial grassland
5	Meadow	4	Natural meadow	2	Natural meadow
6	Meadow steppe
7	Scrubland	5	Scrubland	3	Scrubland
8	Arborescent scrubland
9	Undefined adult forest plantation: adult forest plantation; forest plantation with feral forest plantation	6	Adult forest plantation	4	Forest plantation
10	Young or harvested forest plantation	7	Young or harvested f. p.
11	Adult native forest	8	Native forest	5	Native forest
12	Secondary native renewal forest
13	Scrubby native forest
14	Undefined native forest: native forest; adult native forest with secondary native renewal forest
15	Undefined wetland: wetland; Marsh wetland; Ñadi wetland; Vega wetland; another wetland	9	Wetland	6	Wetland
16	Beach and dune	10	Water, riverbank and shore
17	Riverbank
18	River
19	Lentic water body: water body; lake, lagoon, reservoir and dam
20	Urban and industrial	11	Urban and industrial	7	Without vegetation
21	City, town and industrial zone
22	Industrial mining
23	Rocks without vegetation	12	Without vegetation
24	Altitudinal gradient without vegetation
25	Scoria without vegetation
26	Landslide without vegetation
27	Snow and glacier: snow and glacier; snow; glacier

and

Table A6. Land-cover (LC) polygonal datasets.

Dataset Name or Source	Labeled Year	Likely Characterized Period	Static Assumed Period (Recent Past)		Dynamic Assumed Period (Inter-Statics Periods)
			Years	Code	Years	Code
“...period 2013–2016” [42] https://ide.minagri.gob.cl/descarga-de-capas-shp/planificacion-catastral	1997	1995–1996	1995–1996	T1 (not used)
					1997–2004	T1-T2
	2008	2006–2007	2005–2006 †	T2
					2007–2010	T2-T3
	2014	2012–2013	2011–2012 †	T3 (not used)
					2013–2014	Not used
	2017	2015–2016	2015–2016	Not used
“...period 2001–2019” [43] https://ide.minagri.gob.cl/descarga-de-capas-shp/planificacion-catastral	2001	1999–2000	1999–2000	T4
					2001–2010	T4-T5
	2013	2011–2012	2011–2012	T5
					2013–2013	Not used
	2016	2014–2015	2014–2015	Not used
					2013–2014	T5-T6
	2017	2015–2016	2015–2016	T6
	2019	2017–2018	2017–2018	T7

† = assumed lag; T1–T7 = temporal reference codes. Static periods use one reference; dynamic periods use LC polygons unchanged between two references.

(Appendix A.2).

Two temporal modes were used (step 12). In static mode, inventory dates matched the target period and all eligible polygons were retained. In dynamic mode, only polygons whose C0 class remained unchanged between the surrounding static inventories were used. P12 lacked reference data and was excluded from sampling and balancing, but it remained in the final LU reconstruction.

Random points were generated by period, with lower point density in large polygons and higher density in small polygons, and then merged into one dataset. Category and period frequencies were ranked to identify overrepresentation. No points were removed from the rarest category; category-based discard rates increased with frequency, and observations from overrepresented periods were removed above a predefined threshold. Appendix A.2 provides the point-generation, ranking, and discard rules.

Observation-Percentage Treatment Axis (Obs)

Four observation levels represented 12.5%, 25%, 50%, and 100% of the maximum available sample. Each level used a distinct random seed and was drawn from the full observation dataset before intersection with biennial mosaics to extract predictors and period identifiers (Figure 2c, 13). The computational limit was tested using large samples, the full predictor set, and extreme RF-hyperparameter values. Before percentage subsampling, seed-specific datasets averaged 10,922 points, with 140–2547 observations per C0 category and 245–1143 per period. Random subsets were then selected for each Obs level (step 14). Training and validation counts are reported in

Table 3. Training and validation point counts by observation level and summary group.

Group	Items	12.5% Train	12.5% Validate	25% Train	25% Validate	50% Train	50% Validate	100% Train	100% Validate
(a) Total	1	886	443	1882	941	3621	1811	7266	3632
(b) Periods	11	81	40	171	86	329	165	661	330
(c) Pooled LC categories	12	74	37	157	78	302	151	606	303
(d) Pooled LC categories–periods	132	7	3	14	7	27	14	55	28

LC = land cover; Train = training points; Validate = validation points.

(defined in Section 2.7).

2.7. RF Classification and Performance Metrics

The Obs, Bal, and Var axes generated 24 ungrouped treatments (Figure 2c, step 14). For each treatment, observations were divided into three parts: two for training and one for validation. Alternating the held-out part produced three sample variations.

RF classification used ee-Classifier-smileRandomForest in GEE [44] and predicted the 27 C0 classes (step 15). RF hyperparameters were tuned using a grid search [45]. The grid evaluated bag fraction from 0.50 to 1.00 in 0.05 increments, number of trees from 25 to 300 in 25 increments, and variables per split from 2 to 24 in 2 increments (step 16). Minimum leaf population and maximum leaf nodes were unconstrained, allowing retention of rare classes. Across treatments, sample variations, and grid combinations, 31,944 classification measurements were obtained, with 757–1584 per ungrouped treatment.

Training and validation κ were averaged across the three sample variations (Figure 2c, step 17). For interpretation, κ was classified as poor (≤0.4), moderate (0.41–0.60), good (0.61–0.75), excellent (0.76–0.80), or almost perfect (>0.80) [46]. Training κ above 0.60 indicated low underfitting risk; a training–validation gap above 0.40 indicated high overfitting risk, whereas a gap below 0.15 indicated acceptable fit.

RF-hyperparameter dispersion was summarized using a custom metric, the weighted coefficient of variation by κ averaged (WCVκA):

$$WCV\kappa A=1/3 ·{\sum }_h^3\left({WSD\kappa }_h/{WMN\kappa }_h\right)$$

(3)

For each hyperparameter ($h$), κ-weighted means and standard deviations ($WMN\kappa$ and $WSD\kappa$) were calculated from configurations at or above the treatment-specific validation-κ median (Q50); the three resulting coefficients of variation were averaged.

Validation-κ Q50 and validation-WCVκA were analyzed jointly, representing median predictive performance and RF-hyperparameter dispersion among high-performing configurations. Low-, moderate-, and high-performance zones were defined using the existing Euclidean-distance equation based on each treatment’s deviation from the across-treatment means of both metrics. This criterion complemented context-dependent fixed κ thresholds and compared treatment accuracy and sensitivity to hyperparameter choice.

2.8. Treatment Comparisons, Selection, and Complementary Analyses

Normality, homoscedasticity, κ-comparison tests, paired κ differences, and effect-size procedures are detailed in Appendix A.3. Treatment evaluation proceeded from grouped to ungrouped analyses. Individual Obs, Bal, and Var axes and the combined Obs-Bal and Obs-Var groups were compared first. For ungrouped analyses, Obs_i, Bal_j, and Var_k denoted the respective axis levels. All ungrouped treatments, denoted Obs_i-Bal_j-Var_k, were then ranked by validation-κ Q50.

Ranking guided selection of contrasting Obs-No-BI and Obs-Yes-BIT series across Obs levels. They examined balancing and predictor-set effects, summarized RF-hyperparameter optima, and assessed training- and validation-κ stability. Hyperparameter statistics were calculated from classifications above the 99th percentile of validation κ. The highest three-fold validation-κ configuration per selected treatment was reevaluated in a new three-fold round, yielding six replications for standard deviation (SD) estimation.

Predictor importance was evaluated using RF mean decrease impurity [47,48]. Observation percentage and balancing state were fixed at their strongest levels, and the 100-Yes-Var series compared B, BI, and BIT. Mean decrease impurity values from six replications were averaged, ranked, and divided by the maximum within each predictor set.

The highest-ranked ungrouped treatment by validation-κ Q50 was selected for LU reconstruction.

2.9. LU Mapping and Change-Rate Analysis

For the selected treatment, the model with the highest mean κ across the three sample variations was applied to all 12 biennial mosaics, producing three C0 classification series. For each period, the three maps were combined by pixel-wise statistical mode. Missing pixels were filled using adjacent-period classifications, index-based masks, and one-pixel local-mode extrapolation; Appendix A.4 provides the complete heuristics. Final maps were aggregated from C0 to C1 and C2 (Figure 2c, step 18).

Annual exponential LU-change rates (% year⁻¹, base e) were compiled from studies overlapping the study area [5,6,30,32,49,50,51,52,53]. Literature-derived values were solved for $rate$ using a two-point formulation, $\ln \left({Area}_{Last}/{Area}_{First}\right)=\left(rate/100\right)·\left({Period}_{Last}-{Period}_{First}\right)$; whereas for this study were obtained across all periods using a continuous formulation, $\ln \left(Area\right)=\left(rate/100\right)·Period+\mathrm{c}\mathrm{o}\mathrm{n}\mathrm{s}\mathrm{t}\mathrm{a}\mathrm{n}\mathrm{t}$. Published areas, calculations, and categories were harmonized or recalculated when required to match the C2 scheme.

2.10. Software

GEE (API REST v1) was used for image processing, random point sampling, RF classification, and post-classification analysis [3]; QGIS (version 3.38) was used to harmonize polygon datasets and prepare maps [54]; and R (version 4.3.1) was used for statistical analyses and performance graphics [55].

3. Results

3.1. Color-Unbalanced and Color-Balanced Biennial Mosaics

Color-unbalanced and color-balanced biennial mosaics were compared across periods and seasons to assess whether per-band histogram matching improved the consistency of spectral predictors before classification. The comparison used winter and summer Landsat composites generated under the balancing-axis treatments.

Figure 3 shows that summer variables had more stable distributions than winter variables in both central tendency and internal variability. Most spectral bands and indices also showed lower temporal dispersion after color balancing. Minor exceptions occurred for $Bswir1\_Sum$, $Bswir2\_Sum$, $Bsi\_Sum$, and $Nbri\_Sum$, for which color-balanced values showed slightly higher variability. These exceptions were limited and did not alter the overall pattern of greater stability under summer conditions and after color balancing.

Period-level differences were also evident. P1 showed the greatest spectral misalignment, coinciding with low image availability and the inclusion of out-of-season imagery. P8 remained comparatively stable despite the dominance of L7 imagery affected by scan-line errors. Variability increased after P9, coinciding with the incorporation of L8 imagery and its different wavelength ranges and atmospheric-correction characteristics. Overall, residual spectral differences were period-specific and coincided with image availability, seasonal coverage, and sensor composition. These findings provide the empirical basis for the subsequent discussion of the effectiveness and limitations of color balancing.

3.2. Sample Distribution and Imbalance Across LC Categories and Periods

The post-classified area and observation-point distributions revealed substantial inherent imbalance across the landscape (Figure 4). Observation-point distributions remained imbalanced across LC categories and periods even after sampling and discarding. Each discarding rule also affected the opposite dimension: abundant categories were penalized most, wetlands remained unaffected, and native forests showed notable reductions. Period-based discarding reduced temporal overrepresentation, although data scarcity in P1 persisted.

Mosaic extraction affected categories unevenly and partially reintroduced category imbalance. C1.10-riverbank-and-shore and C1.12-without-vegetation showed the largest proportional reductions, mainly due to masking from snow, water, and adjacent pixels. C1.11-urban-and-industrial declined where high-reflectance urban surfaces overlapped spectrally with cloud–snow index responses. Natural meadows, scrubland, and native forests were also disproportionately reduced because of their concentration in mountainous terrain. The residual imbalance was concentrated mainly in rare or highly masked categories, including water-related (C0.16–C0.19), industrial-mining, and without-vegetation (C0.22–C0.27) classes. Overall, the sampling design controlled temporal imbalance more effectively than category imbalance.

3.3. Performance Zones Based on Validation-κ Q50 and WCVκA

Validation-κ Q50 and validation-WCVκA were analyzed jointly to evaluate median classification performance and RF-hyperparameter dispersion among high-performing configurations. Figure 5 shows an inverse, roughly linear relationship among ungrouped treatments: as validation-κ Q50 increased, validation-WCVκA decreased. Thus, higher-performing treatments were associated with lower RF-hyperparameter dispersion. Grouped and combined grouped treatments followed the same trend, indicating consistency across analysis levels. Because training-κ values were consistently high, later comparisons focused on validation-related metrics and their variation across treatments.

Across grouped and combined grouped treatments, increasing the observation percentage or applying color balancing alone improved treatment position but was insufficient to reach the high-performance zone. The grouped 100%, color-balanced, and combined 100-Yes treatments reached only moderate performance. In contrast, the grouped BIT treatment produced the largest shift and entered the high-performance zone; among Obs-BIT combinations, only 12.5-BIT remained in the moderate zone.

Among ungrouped treatments, 100-Bal_j-B and 100-Bal_j-BI reached the moderate zone, but only their Yes versions did so consistently. Even 12.5-Bal_j-BIT treatments reached the moderate zone under both No and Yes conditions, while 25-, 50-, and 100-Bal_j-BIT treatments advanced progressively toward the high-performance zone. Validation-κ Q50 exceeded 0.75 for the 100-Bal_j-BIT treatments, indicating well-fitting configurations (Figure 5). This pattern extended the BIT advantage observed in the grouped analysis across balancing states and observation levels.

3.4. Validation-κ Across Grouped Treatments and Selected Ungrouped Series

Classification performance based on validation-κ differed clearly across treatment configurations, with predictor-set design producing the strongest overall contrast. Grouped treatments were first evaluated to identify the effects of each treatment axis. The complete ungrouped-treatment ranking was then used to select Obs-No-BI and Obs-Yes-BIT as contrasting series for assessing how color balancing and predictor-set design interacted across observation levels.

Increasing the observation percentage generally improved validation-κ (Figure 6a). However, paired κ differences alternated in sign between adjacent observation levels, indicating that the increase was weak and not consistently monotonic (Figure 7a).

Color balancing produced more symmetric distributions and reduced variability (Figure 6b). Paired κ differences confirmed a positive but modest effect relative to the predictor-set axis (Figure 7b,c). Combined Obs-Bal treatments showed a similar pattern, with generally lower dispersion after color balancing, particularly at lower observation levels (Figure 6d and Figure 7d).

Within the predictor-set axis, the transition from B to BI generated negligible κ changes, indicating that spectral indices added limited information beyond the original bands. In contrast, the transition from BI to BIT produced large improvements, representing the strongest gain among the single-axis analyses (Figure 6c and Figure 7c). Obs-Var combinations reinforced this pattern: differences remained limited from Obs-B to Obs-BI but were consistently positive from Obs-BI to Obs-BIT (Figure 6e and Figure 7e). These differences remained positive across all observation levels, although their magnitude decreased at higher levels.

The complete ranking of ungrouped treatments followed the same overall trends identified in the grouped analyses and revealed no additional major patterns. One minor pattern was that Obs_i-Bal_j-BI treatments generally showed lower validation-κ dispersion than their Obs_i-Bal_j-B counterparts, although this did not consistently translate into higher performance. Detailed results are provided in Appendix B.2.

Within the contrasting Obs-No-BI and Obs-Yes-BIT series, performance followed a consistent ranking from 12.5-No-BI to 100-Yes-BIT (Figure 8). At each observation level, Obs_i-Yes-BIT outperformed the corresponding Obs_i-No-BI treatment. The Obs_i-No-BI series increased gradually, whereas Obs_i-Yes-BIT showed a stronger shift after the 25% level, widening the separation between the two series.

Overall, the grouped and selected ungrouped analyses consistently identified predictor-set design as the principal source of performance differences, while observation percentage and color balancing produced smaller and less uniform effects.

3.5. Predictor Importance

Based on the treatment comparisons, the best-performing levels within the Obs and Bal axes, 100% observations and color balancing, were held constant to compare predictor importance across the B, BI, and BIT predictor sets. Figure 9 shows a clear hierarchy in the full predictor set: territorial variables generally ranked highest, followed by spectral bands, whereas spectral indices ranked lowest. Spectral-predictor rankings were otherwise broadly consistent across the 100-Yes-B, 100-Yes-BI, and 100-Yes-BIT treatments.

Among spectral predictors, summer variables generally ranked above their winter counterparts. $Bswir2\_Sum$ and $Bswir2\_Win$ ranked highest, followed by $Bnir\_Sum$ and $Bnir\_Win$, although these predictors shifted toward intermediate positions in 100-Yes-BI. $Bswir1\_Sum$ ranked comparably to other summer bands, whereas $Bswir1\_Win$ shifted toward the lower range. Visible bands occupied intermediate positions, except $Br\_Win$ and $Bg\_Win$, which ranked among the lowest together with the thermal bands. Among spectral indices, $Evi$, $Bsi$, and $Nbri$ showed moderate importance, whereas $Gci$, $Sipi$, and $Ndwi$ consistently ranked lowest. Among territorial variables, accessibility to cities and water bodies and climatic factors occupied several of the highest-ranking positions, followed by terrain metrics. Overall, the predictor-importance results reproduced the performance hierarchy observed among the B, BI, and BIT treatments.

3.6. RF Hyperparameters

RF-hyperparameter distributions differed between the Obs-No-BI and Obs-Yes-BIT series, although the contrast varied among parameters (

Table 4. RF-hyperparameter statistics above the 99th percentile of validation-κ within the Obs-No-BI and Obs-Yes-BIT series.

Obs	Bal	Var	Bag Fraction					Number of Trees					Variables per Split					Nº	Avg CV
			Min	Q50	Max	Opt	CV	Min	Q50	Max	Opt	CV	Min	Q50	Max	Opt	CV
12.5	No	BI	0.55	1.00	1.00	1.00	20	100	163	300	100	30	6	18	24	24	44	16	31
25	No	BI	0.75	0.90	1.00	0.90	10	125	238	300	150	26	6	9	18	8	34	16	23
50	No	BI	0.80	0.90	0.95	0.90	4	125	225	300	225	23	2	3	4	2	34	16	20
100	No	BI	0.80	0.83	0.95	0.80	7	125	175	225	225	18	4	6	8	6	30	14	18
12.5	Yes	BIT	0.80	0.90	0.95	0.80	8	100	225	300	225	28	8	12	12	12	14	16	17
25	Yes	BIT	0.85	0.95	0.95	0.95	4	125	250	300	300	23	8	10	14	8	18	16	15
50	Yes	BIT	0.85	0.90	0.95	0.95	3	150	250	300	300	19	14	16	20	14	17	16	13
100	Yes	BIT	0.85	0.95	0.95	0.90	4	100	250	300	250	33	14	18	20	18	11	15	16

Obs = observation %; Bal = color balancing (No, Yes = unbalanced, balanced); Var = predictor set (BI = bands + indices, BIT = BI + territorial variables); Min = minimum; Q50 = median; Max = maximum; Opt = optimal value; CV = coefficient of variation (%); Nº = selected configurations; Avg CV = average across the three hyperparameters (%). Lower CV indicates greater convergence.

). Bag fraction and number of trees showed limited or inconsistent series-level patterns, whereas variables per split produced the clearest between-series differences.

For bag fraction, 12.5-No-BI showed the broadest range, followed by 25-No-BI, whereas the remaining treatments had similar Q50 and Opt values. High bag-fraction values were common among high-performing configurations but did not vary systematically with performance.

For number of trees, Q50 values ranged from 163 to 250, and maximum values reached 300 in most treatments. Opt values ranged from 100 to 300 but did not increase systematically with performance. Moreover, 100-Yes-BIT showed the highest number-of-trees dispersion despite being the best-performing treatment. Thus, number of trees showed no consistent series-level pattern.

Variables per split showed the clearest contrast between the two series in both dispersion and central values. In Obs-No-BI, CV was consistently higher and generally declined as observation percentage increased, while Q50 and Opt values decreased markedly from 12.5–25% to 50–100%. In contrast, Obs-Yes-BIT showed lower overall dispersion, with Q50 and Opt values increasing from 12.5–25% to 50–100%.

Overall, variables per split distinguished the two treatment series more clearly than bag fraction or number of trees.

3.7. Land-Use Areas and Transitions

Based on the ungrouped-treatment ranking, 100-Yes-BIT was identified as the highest-performing treatment and selected for LU reconstruction. Its optimal configuration– bag fraction = 0.90, number of trees = 250, and variables per split = 18 (Table 4)– achieved κ = 0.820 for Code 0 and κ = 0.910 for Code 2. Additional performance values for the contrasting series are reported in

Table A15. Q50 and Max of training-κ and validation-κ by category-pooling code.

Obs	Bal	Var	Training-κ Q50 Code			Validation-κ Q50 Code			Training-κ Max Code			Validation-κ Max Code
			0	1	2	0	1	2	0	1	2	0	1	2
12.5	No	BI	0.994	0.997	0.998	0.376	0.500	0.670	0.997	1.000	1.000	0.394	0.508	0.680
25	No	BI	0.998	1.000	1.000	0.422	0.520	0.656	0.998	1.000	1.000	0.437	0.536	0.672
50	No	BI	0.990	0.997	0.997	0.453	0.550	0.711	0.990	0.997	0.997	0.474	0.563	0.721
100	No	BI	0.985	0.997	0.997	0.490	0.586	0.735	0.985	0.997	0.997	0.507	0.598	0.743
12.5	Yes	BIT	0.997	0.999	0.999	0.550	0.636	0.752	0.997	0.999	0.999	0.582	0.663	0.762
25	Yes	BIT	0.995	0.999	0.999	0.610	0.682	0.790	0.995	0.999	0.999	0.635	0.703	0.807
50	Yes	BIT	0.991	0.997	0.997	0.739	0.778	0.858	0.991	0.997	0.997	0.769	0.803	0.871
100	Yes	BIT	0.984	0.997	0.997	0.799	0.848	0.899	0.985	0.997	0.997	0.820	0.868	0.910

Obs = observation %; Bal = color balancing (No, Yes = color-unbalanced, -balanced); Var = predictor set (BI = bands + indices, BIT = BI + territorial variables); Q50 = median; Max = maximum; Code 0, 1 = unpooled, pooled LC categories; Code 2 = pooled LU categories.

(Appendix B.2); for example, 100-No-BI achieved κ = 0.507–0.743 as the pooling code increased.

Two complementary criteria were used to examine LU areas: first-to-last changes between P1 and P12 and all-period transitions across P1–P12.

Table 5. LU area, persistence, transitions, and annual change under first-to-last (P1, P12) and all-period (P1, ···, P12) criteria.

C2	P1	P12	(P1 + P12) ⁄2	(P1 + ··· + P12) ⁄12	(d) − (e)	Un-changed (P1, P12)	Un-changed (P1, ···, P12)	(g) − (h)	P1 − (g)	P12 − (g)	P12 − P1	Rate (P1, P12)	Rate (P1,···, P12)	(m) − (n)	Log-trend R2 (P1,···, P12)
(a)	(b)	(c)	(d)	(e)	(f)	(g)	(h)	(i)	(j)	(k)	(l)	(m)	(n)	(o)	(p)
C2.1	27.7	25.2	26.5	26.3	0.2	20.7	16.2	4.6	7.0	4.5	−2.5	−0.43	−0.57	0.14	0.90
C2.2	1.2	1.4	1.3	1.4	−0.1	1.0	0.8	0.2	0.2	0.4	0.2	0.71	0.62	0.09	0.44
C2.3	9.2	9.1	9.1	9.2	−0.1	6.9	5.7	1.2	2.3	2.2	−0.1	−0.07	−0.13	0.06	0.38
C2.4	32.2	35.1	33.6	33.9	−0.3	25.7	13.4	12.3	6.6	9.4	2.9	0.39	0.52	−0.13	0.77
C2.5	19.3	18.9	19.1	18.8	0.3	15.6	12.3	3.3	3.7	3.3	−0.4	−0.10	−0.07	−0.03	0.04
C2.6	2.1	1.8	2.0	2.0	0.0	1.5	1.3	0.2	0.6	0.4	−0.3	−0.62	−0.74	0.11	0.91
C2.7	8.2	8.4	8.3	8.3	0.0	7.7	7.5	0.2	0.5	0.7	0.3	0.14	0.10	0.04	0.70
Total	100.0	100.0	100.0	100.0	0.0	79.0	57.1	22.0	21.0	21.0	0.0	0.00	0.00	0.00	0.00

(a) = code 2 (C2.1 = agricultural, C2.2 = natural meadow, C2.3 = scrubland, C2.4 = forest plantation, C2.5 = native forest, C2.6 = wetland, C2.7 = without vegetation); (b–c) = P1, P12 areas; (d–f) = temporal averages, difference; (g–i) = unchanged areas, difference; (j–l) = loss, gain, net change; (m–o) = annual exponential change rates; (p) coefficient of determination. Values are % of the study area, except rates (% year⁻¹, base $\mathcal{e}$) and R². Higher/lower R² shows more/less consistent trend.

shows that forest plantations and agricultural land jointly dominate the landscape, accounting for approximately 60% of the study area. Native forest and scrubland together accounted for approximately 28%, confirming the mixed agroforestry character of the region. Differences between temporal averages were generally small.

Regarding change rates and temporal consistency, agriculture and wetlands showed the strongest declines and a high R² value of approximately 0.90, indicating consistent downward trajectories. Forest plantations showed a relatively consistent expansion (R² = 0.77), while without-vegetation areas showed a smaller but similar regular increase. Natural meadows and scrubland had weaker consistency (R² ≈ 0.4) although opposite directions of change; whereas native forest combined little rate with the lowest R², reflecting irregular inter-period dynamics.

Categories remaining unchanged throughout P1–P12 accounted for 57.1% of the study area (Figure 10). The 22% difference between unchanged areas under the two criteria indicates substantial intermediate dynamics not captured by first-to-last comparisons. Intermediate dynamics were concentrated mainly in forest plantations and agricultural land, producing differences in losses, gains, net change, and annual rates. All-period estimates showed faster declines in agriculture and wetlands and faster forest-plantation expansion than first-to-last estimates (Table 5). Spatially, the dominant transition was from agricultural land to forest plantations, mainly within established forestry zones of the Coastal Range and Central Depression and, to a lesser extent, into native-forest patches along the Coastal Range. Among smaller categories, natural meadows expanded mainly over degraded scrubland and native forest, whereas wetlands declined through conversion to agriculture and forest plantations.

To compare the estimated LU-change rates with literature-derived values, Figure 11 summarizes annual exponential rates for the LU categories with the greatest data availability. The compiled rates were grouped into an earlier period, 1975–2001, and a more recent period, 1997–2020, and compared with the first-to-last and all-period estimates from this study.

Agricultural rates were predominantly negative across the literature, and the estimates from this study (−0.43 and −0.57) were similar to those reported for comparable spatial extents (−0.83 to −0.54 [51,52,53]). Forest-plantation rates were predominantly positive but generally lower in the recent literature group. The estimates from this study (0.39 and 0.52) were comparable to similar-area values of 0.30–0.38 [5,51], although a higher value of 0.95 was also reported [52]. Native-forest rates remained negative but approached zero in the recent group, consistent with the estimates from this study (−0.10 and −0.07) and regional values of −0.06 to −0.05 [5]. When native forests and forest plantations were pooled, the estimates from this study (0.21 and 0.31) were also consistent with the recent-group average and regional values of 0.21–0.38 [5,52].

Other categories showed less consistent evidence. Scrubland rates remained generally close to zero without a clear direction, while values for natural meadows and without-vegetation areas were scarce and variable. Wetland rates also varied substantially, although the estimates from this study (−0.62 and −0.74) closely matched the value reported for the same spatial extent (−0.60 [51]).

Overall, the reconstructed series and literature-derived comparisons identified agricultural decline and forest-plantation expansion as the clearest regional patterns, while smaller categories showed more variable trajectories.

4. Discussion

4.1. Data Quality, Harmonization, and Sample Imbalance

Integrating imagery from multiple satellites and territorial zones extends coverage but requires homogenization to limit sensor-, season-, and period-related artifacts [56,57]. Histogram matching did not fully correct spectral differences, but it reduced mismatches that could otherwise be interpreted as LC/LU change. Stable distributions despite L7 scan-line errors also supported the compositing and smoothing strategy. The limited response of $Bswir1$ and $Bswir2$ suggests that these bands were stable before harmonization, while the remaining exceptions show that improvements were not uniform. Shortwave-infrared and thermal predictors are relatively resistant to atmospheric scattering and capture vegetation structure, moisture, and surface-temperature gradients [58], which may improve discrimination of spectrally ambiguous categories. The procedure nevertheless assumes that bands can be harmonized independently. Future refinements could account for inter-band correlations using multidimensional histograms or color-space transformations [14]. Reliance on a master reference image is another limitation, while direct image-fusion methods remain difficult to adapt to multispectral data [59]. Color balancing should therefore be treated as a harmonization aid, not a universal correction.

Sample imbalance poses a related problem because equal representation cannot usually be achieved without information loss. Downsampling to the rarest class discards useful observations, whereas oversampling adds no information and may create artificial spatial dependence. Slightly imbalanced samples can outperform perfectly balanced ones [16,17], and studies have mainly addressed class imbalance rather than temporal or territorial imbalance [7,16,17]. Although category-based CVs were high, the pooled maximum-to-minimum ratio (18) was close to Zhu et al. [17] (13), and the unpooled ratio (2026) fell within the 1101–15,922 range reported by Hermosilla et al. [7]. Thus, residual imbalance was comparable to previous studies and was a practical rather than exceptional constraint. It nevertheless increased uncertainty for rare, masked, or spectrally ambiguous categories. Leaving min-leaf-population unconstrained allowed individual observations to define leaf nodes and preserved rare-class signatures, although at greater overfitting risk.

Its independent effect on accuracy could not be isolated, but RF remains suitable because of its tolerance to imbalanced and noisy training data [16,20,60,61,62]. Overall, radiometric harmonization reduced spectral inconsistencies, while controlled imbalance preserved rare-category information. These measures mitigate different input-data biases without eliminating uncertainty.

4.2. Treatment Effects and Predictor Design

Increasing the number of observations generally improved performance by providing broader class representation, but gains diminished at higher levels, suggesting an emerging saturation pattern rather than a precise sample threshold. Some computationally demanding classifications were unavailable because of GEE limits, although the small gains at the upper observation levels suggest that this missingness did not materially alter the general pattern. Future studies should test finer sample increments using less restrictive computing environments.

Color balancing addressed a different constraint by improving temporal spectral consistency rather than adding class information. Its declining marginal contribution as sample size increased suggests a partial trade-off between radiometric harmonization and observation availability. It may therefore be most valuable when reference data are limited or further sampling is impractical, while remaining complementary to predictor design rather than replacing informative variables.

The limited improvement from adding spectral indices is mainly due to redundancy. Because the seasonal indices are algebraic transformations of the same winter and summer bands already included in the baseline predictor set, they emphasize spectral relationships but add limited independent information. Moreover, the two seasonal band sets already expose RF to winter–summer differences, while RF can capture nonlinear relationships and interactions directly from the original bands and tolerate multicollinearity [20,21,63,64]. Thus, the ecological interpretability of indices does not necessarily translate into additional predictive information. Bera et al. [18] likewise found no consistent hierarchy between bands and indices, indicating that their value is context-dependent. The low importance of $Gci$, $Sipi$, and $Ndwi$ likely reflects information already represented by $Bnir$ and the visible band. However, as predictor sets were cumulative, comparisons with indices-only configurations are needed to isolate their independent contribution.

In contrast, territorial variables produced stronger gains because they introduced nonredundant spatial context not derivable directly from the seasonal spectral bands. Previous work has similarly shown that elevation, slope, and proximity to water can equal or exceed spectral predictors [7]. Although Level-2 surface-reflectance products correct major physical effects [33,34,35], they do not encode landscape position, accessibility, or environmental gradients associated with LC/LU patterns. Variables such as $WbInd$ and $AccesCity$ therefore helped distinguish spectrally similar categories, including wetlands with variable vegetation and natural versus artificial bare soil.

These contextual gradients were also consistent with established LU distributions, including agriculture on gentle, irrigated terrain, forestry on steeper slopes near roads, and native forest in higher, rugged areas [6,28,65]. These associations are explanatory rather than causal, and predictor rankings based on mean decrease impurity remain sensitive to RF stochasticity, tree number, and calibration [22,66]. Nevertheless, the results imply that informative predictor design can partly compensate for limited samples, whereas additional observations and color balancing provide complementary but smaller benefits.

4.3. Performance Zones and RF-Hyperparameter Tuning

Performance zones complemented fixed κ thresholds by combining median validation performance with RF-hyperparameter dispersion. Although κ remains useful for multi-class and imbalanced classification because it adjusts agreement for chance [45], fixed thresholds are context-dependent [45,67]. RF fit also reflects model complexity, noise, irrelevant predictors, and the bias–variance trade-off, while its ability to prevent overfitting remains debated [16,20,45,64,68,69]. Low variability across cross-validation replications further suggests that treatment configuration, rather than random sample splits, drove the main performance differences.

To place these results in broader quantitative context, Sun and Ongsomwang [23], using Sentinel-2 spectral and auxiliary data with RF, obtained κ = 0.91 (6 classes), consistent with the optimal 100-Yes-BIT configuration here (κ = 0.82–0.91; 27–7 classes). Nguyen et al. [70] compared parametric multinomial logistic regression with conventional machine-learning classifiers (k-nearest neighbors, support vector machine, and RF) using multi-season Sentinel-2 bands and obtained κ = 0.61–0.81 (11 classes). Sentinel-2 deep-learning approaches achieved κ = 0.71–0.76 (23–7 classes) with Swin-UNet [71] and κ = 0.87 (8 classes) with a multi-year IRUNet ensemble [72]. These values slightly exceeded the optimal spectral-only 100-No-BI configuration here (κ = 0.51–0.74; 27–7 classes).

Beyond numerical performance, these models illustrate an interpretability–flexibility continuum rather than a hierarchy. Parametric models offer direct parameter-level interpretation but impose stronger functional assumptions, whereas deep-learning architectures provide greater flexibility but rely on latent features less directly linked to physical predictors. Conventional machine-learning approaches such as RF occupy an intermediate position, capturing nonlinear interactions while retaining partial interpretability through predictor-set comparisons and variable importance [64]. Because study designs differ, these κ values are contextual benchmarks rather than direct model rankings.

Within this study, however, interpretation of WCVκA remains exploratory because it depends on its weighting scheme and on κ values above Q50. Alternative weighting and scaling approaches should therefore be tested. Nevertheless, the inverse relationship supports jointly using validation-κ Q50 and validation-WCVκA to select treatments based on median performance and convergence, rather than on a single optimal configuration.

The tuning results translate this convergence pattern into parameter-specific recommendations. Bag fraction was highly data-dependent, consistent with reported optimal ranges of 0.1–0.9 [18,23,44,73], and therefore requires at least coarse tuning. A value of 1.0 may be suboptimal because it removes bagging stochasticity. Number of trees showed similarly broad behavior. Studies have tested 10–5000 trees and commonly reported optima of 150–1000 [18,20,23,73], while Svoboda et al. [73] identified 150 trees as optimal after testing up to 400. Increasing ensemble size may promote stabilization, but the maximum tested value should not be assumed optimal.

The variables-per-split parameter was more sensitive to predictor composition. Common starting rules include the square root, one-third, or all predictors [18,20,44,74,75], with the one-third rule supported here as a practical baseline rather than a universal optimum. However, the two series showed opposite trends as observation percentage increased. In Obs-No-BI, variables per split decreased, possibly because smaller predictor subsets reduced redundancy among indices that largely recombined spectral-band information. In contrast, variables per split increased in Obs-Yes-BIT, potentially allowing the model to exploit the nonredundant context provided by territorial variables. This parameter should therefore be calibrated according to both predictor composition and observation availability, whereas bag fraction and number of trees may be tuned more coarsely.

4.4. Reliability and Comparability of LU Change

The LU patterns should be interpreted through classification uncertainty and temporal consistency. Agriculture and forest-plantation changes are better supported by broad spatial representation and clear regional patterns. Wetlands, natural meadows, scrubland, and small native-forest patches are sensitive to spectral confusion, limited representation, and temporal inconsistency. Sample imbalance and spectral confusion may interact because fewer observations reduce representation of within-class variability, while spectral overlap with more abundant classes can increase omission and commission errors. When independently classified periods are compared, these errors may propagate into apparent gains or losses, particularly for small categories. Their estimates are therefore more reliable for indicating direction and relative magnitude than exact pixel-level conversions.

These category-level limitations also constrain cross-study comparison despite harmonization and recalculation. Previous studies merged wetlands with agriculture or scrubland [6,30], natural meadows with agriculture and scrubland [5,50,53], and scrubland with native forest [6,32]. These differences may reflect grazing on degraded land, weak separation of managed and natural grasslands, and inconsistent scrubland definitions. Small categories were also frequently omitted or merged. Literature-derived rates should therefore be treated as contextual references rather than direct validation.

Within these constraints, reliability depended on agreement between internal temporal consistency and external evidence. Agriculture and forest plantations provided the strongest support because their trajectories were consistent and broadly agreed with studies covering comparable areas. Wetland decline was internally consistent, but small extent, spectral ambiguity, limited literature, and frequent category merging reduced broader confidence. Native forest showed a generally declining direction consistent with regional studies, although irregular inter-period dynamics weakened support for a continuous trend. Scrubland, natural meadows, and without-vegetation areas showed more variable trajectories and weaker comparative evidence. Some were also concentrated in mountainous areas with limited imagery, where residual gaps required heuristic post-classification filling. Thus, dominant and consistently defined categories support more reliable change directions than small, ambiguous, or inconsistently classified ones.

Beyond category-level limitations, the generally lower estimated rates may partly reflect the classification design. Persistent areas dominated the landscape so that sampling may have represented persistence more strongly than change. In addition, invariant territorial predictors emphasized stable spatial contrasts rather than temporal dynamics. Together, these mechanisms may have favored persistent classifications and reduced estimated change magnitudes. Future work should balance observations by change versus persistence and, where feasible, transition type, in addition to category and period. Transition-based approaches such as Land-Use 2.0 could classify transitions and persistence directly rather than infer them from independently classified periods [15], strengthening change estimates.

A further limitation is that spatial transition uncertainty was not quantified. The final LU reconstruction used the modal class of three classification series generated from the three sample variations of the single optimal RF-hyperparameter configuration in 100-Yes-BIT. Variation among these series therefore reflects sensitivity to the three-fold sample variation for that configuration, rather than variability across the high-performing RF-hyperparameter configurations evaluated for the treatment. A broader uncertainty assessment should consider both sources of variation before deriving spatial, category-level, or transition-level uncertainty for LU transitions.

5. Conclusions

This study provides four conclusions and practical recommendations for multi-decadal LC/LU mapping under limited spatial and temporal data.

First, mapping reliability depended more on observation quality, seasonal coverage, and temporal representativeness than on archive size, while radiometric harmonization improved cross-sensor and cross-period consistency. Multi-temporal workflows should therefore prioritize high-quality, seasonally consistent imagery and explicit harmonization rather than maximizing image quantity.

Second, perfect sample balance was neither feasible nor desirable because it would discard information from dominant classes or artificially replicate rare ones. However, retaining very small samples increased uncertainty and overfitting risk. Sampling designs should seek controlled balance and explicitly report uncertainty for rare, masked, and spectrally ambiguous categories.

Third, predictor design was the main determinant of performance: spectral indices contributed limited new information, color balancing improved temporal consistency, and territorial variables added nonredundant environmental and spatial context that improved class discrimination. Data-limited applications should therefore prioritize relevant territorial predictors, using color balancing and additional observations as complementary measures rather than substitutes for informative predictor design.

Fourth, model selection should jointly consider validation accuracy and RF-hyperparameter dispersion because high-performing treatments showed greater convergence across plausible configurations, whereas weaker treatments depended more strongly on specific settings. Treatments should therefore be selected using high median accuracy and low dispersion, with coarse tuning for bag fraction and tree number and predictor-dependent calibration of variables per split.

The reconstructed LU series supported the regional importance of agricultural decline and forest-plantation expansion, but transition reliability was higher for dominant, consistently classified categories than for small or spectrally ambiguous ones. Transition-specific spatial uncertainty was not quantified; future work should consider both three-fold sample variation and variability among high-performing RF-hyperparameter configurations. Literature-derived rates should be treated as contextual comparisons rather than direct validation, and future studies should balance observations by change versus persistence and, where feasible, by transition type.