Cropland mapping in rapidly urbanizing tropical regions increasingly relies on multi-sensor Earth observation,

yet the individual contribution of different sensor and feature types to classification accuracy is rarely quantified

explicitly. Building on a prior comparative study of Random Forest (RF) and Gradient Tree Boosting classifiers

were compared across combinations, and McNemar's test was used to assess whether accuracy differences

between stacks were statistically significant. Results show that SAR backscatter, not spectral indices or

phenological metrics, was the dominant driver of accuracy gains: the optical-plus-SAR combination achieved

the highest overall accuracy (74.8%) and Kappa (0.685) of all six combinations tested, exceeding even the full

fused stack (73.5%, κ=0.668), while adding spectral indices and phenological metrics without SAR reduced

accuracy below the optical-only baseline (67.0% vs. 68.9%). McNemar's tests confirmed the full stack was not

significantly different from the optical-only baseline or the optical-plus-SAR combination (p > 0.05), but was

significantly more accurate than any combination lacking SAR (p < 0.05). Gini-based variable importance

ranked VH backscatter and VH temporal variance as the two most important predictors among all 20 bands, and

this ranking was corroborated by a permutation-importance cross-check on an independently trained model,

Importance, Cropland Mapping

Agricultural land constitutes approximately 38% of the Earth's terrestrial surface and underpins food security

and ecosystem services, particularly in rapidly urbanizing regions of the developing world (Azadi et al., 2021).

Timely and accurate information on cropland extent is essential for land use planning, food security assessment,

and climate-resilient urban development (Dubey et al., 2025). Satellite remote sensing, and the Sentinel

constellation in particular, has become central to operational land cover monitoring because it provides

continuous, repeatable observation independent of ground access constraints (Du et al., 2019).

A growing body of work has demonstrated that fusing optical and radar imagery, together with derived spectral

indices and phenological metrics, improves cropland and land cover classification accuracy relative to single-

Boosting (GTB) classifiers for cropland mapping in the Abuja Municipal Area Council (AMAC), Nigeria

(Ibrahim et al., 2026). What remains less well quantified is the marginal contribution of each feature group within

such a stack: whether the added processing cost of incorporating SAR time series or phenological metrics is

justified by a commensurate gain in accuracy, relative to a simpler optical-only or optical-plus-indices approach

and, notably, whether combining every available feature type is even the best strategy, as opposed to a more

selective combination.

This distinction matters for practical deployment. In data- and resource-constrained settings, the cost of

acquiring, preprocessing, and maintaining multiple sensor time series is non-trivial, and operational agencies

benefit from evidence on which feature investments yield the greatest returns. The objective of this study is

therefore to quantify, through a systematic feature ablation design, the individual and combined contribution of

spectral indices, NDVI-derived phenological metrics, and SAR backscatter to cropland classification accuracy

characterized by distinct wet and dry seasons. Full details of the study area are provided in Ibrahim et al. (2026).

All data processing was implemented in Google Earth Engine (GEE) (Gorelick et al., 2017). The optical dataset

comprised Sentinel-2 Level-2A Surface Reflectance imagery (COPERNICUS/S2_SR_HARMONIZED) for the

year 2025, cloud-masked using the Scene Classification Layer (SCL) to exclude cloud, cloud shadow, and cirrus

pixels.

The radar dataset comprised Sentinel-1 Ground Range Detected Interferometric Wide-mode imagery

(COPERNICUS/S1_GRD) with VV and VH polarizations over the same period. Preprocessing steps follow

Ibrahim et al. (2026) and were not modified for this study, so that any accuracy differences observed here are

Table 1: Summarized Optical Spectral Bands, Spectral Indices, NDVI-derived phenological metrics, and SAR

backscatter texture measures.

specifies the sample size per class rather than in total, the resulting reference dataset comprised approximately

1,250 points across the five classes. Following a 70/30 train-validation split with a fixed random seed (42), this

yielded 373 validation points (approximately 69–77 per class), which were held fixed across all six feature-stack

combinations so that any accuracy differences reflect feature composition rather than sampling variation. We

note this reference dataset is derived from an existing global land cover product rather than independently

collected ground truth, and validation accuracy should be interpreted as agreement with WorldCover rather than

an assessment against field-verified labels.

To isolate the contribution of each feature group, six feature-stack combinations were classified independently

using an identical classifier configuration: (1) OPT only, as a baseline; (2) OPT + IDX; (3) OPT + PHEN; (4)

random seed of 42, consistent with the RF configuration in Ibrahim et al. (2026). Gradient Tree Boosting was

not re-run in this study, as the objective here is feature contribution rather than algorithm comparison; RF was

retained as the sole classifier because it was identified as the more effective algorithm for cropland-class

discrimination in the original comparison. To reduce computational overhead, training and validation pixel

values for the full 20-band stack were extracted once, and each combination's classifier was trained using the

relevant subset of columns rather than re-extracting pixel values from imagery for every combination. For each

combination, overall accuracy, the Kappa coefficient, and cropland-class precision, recall, and F1-score were

computed from the confusion matrix on the held-out validation set.

For the full-stack RF model, Gini-based variable importance (mean decrease in impurity) was extracted per band

cardinality variables a concern directly relevant here, given that several SAR and phenology-derived bands are

continuous while the underlying reflectance bands are as well this ranking was cross-checked against

permutation feature importance, computed as the mean decrease in validation accuracy over 30 random shuffles

To test whether differences in classification accuracy between feature-stack combinations were statistically

significant rather than attributable to sampling variability, McNemar's test (Foody, 2004) was applied pairwise

between the full-stack model and each reduced combination, using the paired correct/incorrect classification

outcomes for each of the 373 validation points, tracked by a persistent point identifier across all six combinations.

A significance threshold of α = 0.05 was adopted, with exact binomial McNemar p-values used given the

discordant pair counts involved.

Feature-stack construction, sampling, and classification were performed in Google Earth Engine (Gorelick et al.,

2017). McNemar's tests were computed in Python 3 (SciPy) from per-point prediction outcomes exported from

Table 4: Gini Variable Importance

in Section 3.3 offer direct support for this explanation: eleven of the twenty bands, including NDBI and NDVI

amplitude both inflated in the Gini ranking relative to permutation importance recorded near-zero or negative

permutation importance, indicating they contributed little unique predictive signal once the dominant SAR bands

were already in the model.

The finding that OPT+SAR statistically matched the full stack (Table 3; McNemar's p=0.499) while using 13 of

the 20 available bands, rather than requiring the full feature set, has direct practical value: it suggests that, in this

setting, the phenological metrics and spectral indices in the original 20-band stack could be omitted from an

operational pipeline with no significant loss of classification accuracy, simplifying preprocessing to two sensor

sources (Sentinel-2 spectral bands and Sentinel-1 SAR) rather than four derived feature groups. This runs

somewhat counter to the common assumption in the multi-sensor fusion literature that combining more

field-level crop-type or phenology data, so the explanation below should be read as an interpretation consistent

with the pattern of results and the known regional context, not as something the dataset directly confirms. The

Federal Capital Territory falls within the Guinea savanna zone of Nigeria's middle belt, where cultivation is

dominated by rain-fed smallholder farming of maize, sorghum, millet, and cassava, sown between April and

June and harvested between August and October, closely tracking the territory's single rainy season (April–

October) (FCTA/ACReSAL, 2025). Crucially, the natural vegetation matrix surrounding cropland in AMAC

classified here as 'other vegetation' and comprising the park/grassy savanna, savanna woodland, and shrub

savanna that together cover the majority of the territory greens up on largely the same rainfall-driven schedule

as the surrounding crops. Because both cover types flush green at the onset of the rains and senesce at their end,

their NDVI-derived phenological signatures (green-up timing, peak greenness, senescence) are expected to

SAR backscatter, by contrast, is sensitive to canopy structure and volume scattering rather than pigment- or

greenness-driven reflectance, and is therefore better positioned to distinguish the relatively low, seasonally

cultivated, row-structured canopy of smallholder cereal and tuber fields from the taller, woodier, structurally

ground by the October–November dry-season onset, producing a large within-year range in backscatter that a

single natural-vegetation phenology metric would not necessarily capture. Woody savanna vegetation, retaining

structural elements (trunks, branches) even through senescence, would be expected to show comparatively lower

structural variance across the same period. This structural-and-temporal contrast, rather than a spectral or

greenness-based one, offers a physically grounded explanation for why radar not optical indices or phenology

emerged as the dominant contributor to cropland discrimination in this landscape, and suggests that the value of

From a practical standpoint, these findings suggest that operational cropland mapping programs in similar

tropical savanna, data-constrained contexts should prioritize investment in radar time series and SAR

preprocessing capacity over the added complexity of deriving and maintaining phenological metrics and spectral

index stacks, at least as a first-order feature engineering decision. This recommendation is offered with the caveat

that it reflects a single growing season and a single geographic setting, and should be tested across additional

vs. Full stack had only 35 discordant pairs), so the absence of a significant difference between OPT+SAR and

the full stack should be interpreted as inconclusive rather than as strong evidence of true equivalence. Third,

reference labels were derived from the ESA WorldCover product rather than independently collected ground

truth, meaning reported accuracy reflects agreement with an existing global product rather than field-verified

classification performance. Fourth, the ablation reflects a single growing season; the relative value of SAR versus

optical/phenological features, particularly given SAR's sensitivity to soil moisture, may differ across seasons

with different rainfall patterns or during periods of more persistent cloud cover, when SAR's cloud-penetration

advantage would be expected to matter even more.

This study conducted a systematic feature ablation to quantify the individual and combined contribution of

spectral indices, NDVI-derived phenological metrics, and SAR backscatter to cropland classification accuracy

in the Abuja Municipal Area Council, Nigeria, building on the multi-sensor feature stack established in Ibrahim

et al. (2026). Contrary to the assumption that a fully fused feature stack maximizes accuracy, SAR backscatter

emerged as the dominant contributor across accuracy comparison, McNemar's significance testing, and Gini

investment in operational cropland monitoring in data- and resource-constrained tropical settings favoring SAR-

a reproducible, GEE-based ablation framework that can be applied to other study areas, land cover classes, and

classifiers in future work.

directions. ISPRS journal of photogrammetry and remote sensing, 114, 24-31.Breiman, L. (2001).

Random forests. Machine Learning, 45(1), 5-32.

3. Du, J., Watts, J. D., Jiang, L., Lu, H., Cheng, X., Duguay, C., & Tarolli, P. (2019). Remote sensing of

environmental changes in cold regions: Methods, achievements and challenges. Remote Sensing, 11(16),

1952.

4. Dubey, A., Singh, P. K., Singh, S., Manzoor, U., Sahoo, C., & Saini, A. (2025). Impact of Climate Change

and Land Degradation on Agriculture. In Eco-Resilience: Climate Change, Land Degradation and

Sustainable Solutions (pp. 49-95). Cham: Springer Nature Switzerland.

5. FCTA/ACReSAL. (2025). FCT: Agro-Climatic Resilience in Semi-Arid Landscapes. Federal Capital

Territory Administration. Retrieved from https://acresal.gov.ng/states/fct/

of Abuja Municipal Area Council, Nigeria. International Journal of Latest Technology in Engineering,

Management & Applied Science (IJLTEMAS), 15(4), 156–

164. https://doi.org/10.51583/IJLTEMAS.2026.150400014

9. Li, Y., & Xiao, X. (2025). Deep learning-based fusion of optical, radar, and LiDAR data for advancing

land monitoring. Sensors, 25(16), 4991.

10. Van Tricht, K., Gobin, A., Gilliams, S., & Piccard, I. (2018). Synergistic use of radar Sentinel-1 and