Mapping Banana and Peach Palm in Diversified Landscapes in the Brazilian Atlantic Forest with Sentinel-2

Abstract

Mapping banana and peach palm in heterogeneous landscapes remains challenging due to spatial heterogeneity, spectral similarities between crops and native vegetation, and persistent cloud cover. This study focused on the municipality of Jacupiranga, located within the Ribeira Valley region and surrounded by the Atlantic Forest, which is home to one of Brazil’s largest remaining continuous forest areas. More than 99% of Jacupiranga’s agricultural output in the 21st century came from bananas (Musa spp.) and peach palms (Bactris gasipaes), underscoring the importance of perennial crops to the local economy and traditional communities. Using a time series of vegetation indices from Sentinel-2 imagery combined with field and remote data, we used a hierarchical classification method to map where these two crops are cultivated. The Random Forest classifier fed with 10 m resolution images enabled the detection of intricate agricultural mosaics that are typical of family farming systems and improved class separability between perennial and non-perennial crops and banana and peach palm. These results show how combining geographic information systems, data analysis, and remote sensing can improve digital agriculture, rural management, and sustainable agricultural development in socio-environmentally important areas.

1. Introduction

The Atlantic Forest (AF) is one of the most biodiverse biomes on the planet [1]. About 21% of its remaining area is located in the Ribeira Valley, in Brazil’s southwest, between São Paulo and Paraná states [1]. This region is home to traditional communities and family farmers whose livelihoods depend on the forest’s natural resources [2,3].

Banana (Musa spp.) and peach palm (Bactris gasipaes) cultivation play a central role in the Ribeira Valley, with considerable economic, social, cultural, and environmental significance. Banana production ensures income for family farmers and contributes to local food security [4,5], while integrating peach palm heart into agroforestry systems offers a sustainable alternative that combines low-impact management, diversified production, and rural employment generation [5,6].

Diversified systems, designed to simulate the structure of tropical forests, provide greater stability and adaptability under climate vulnerability [7,8]. They also contribute to ecosystem recovery by supporting soil conservation, water cycling, and pollination [9,10,11]. Empirical studies have shown that such systems ensure more stable income and productivity over time, with greater resilience to seasonal losses. In contrast, specialized systems tend to be more fragile and economically volatile [12,13].

At the landscape scale, diversification enhances environmental and economic resilience by mitigating volatility in agricultural returns and stabilizing productivity in the face of market and environmental uncertainties [14]. Agroforestry approaches, by incorporating structural and functional diversity, reinforce ecosystem services and multifunctionality in rural landscapes [15].

Digital agriculture (DA) has emerged as a strategic tool for monitoring, planning, and territorial management [16]. The combined use of remote sensing, data analysis, and geographic information systems (GISs) enables detailed spatial and temporal characterization of land use and cover, supporting the identification of productive patterns and decision-making at multiple scales [16,17]. In Brazil, a national survey revealed that 84% of rural producers already use at least one digital technology, and 95% expressed interest in expanding their use [17], underscoring the growing role of DA in fostering innovation and socioeconomic development.

DA offers a wide range of potential applications, including monitoring crop health and productivity in various contexts. However, these applications depend on accurate crop mapping. Nevertheless, mapping tropical agroforestry systems with orbital remote sensing remains technically challenging due to their spatial heterogeneity [18]. The coexistence of multiple species within small properties, overlapping canopies surrounded by native forests, and similar phenological cycles often result in spectral confusion, particularly in medium-resolution sensors such as Sentinel-2 [19,20,21].

In recent years, Sentinel-2 imagery has proven effective for monitoring heterogeneous and fragmented tropical landscapes [22]. Its high spatial resolution (10–20 m), short revisit interval (up to five days), and rich spectral range make it especially suitable for detecting land-cover dynamics even under persistent cloud cover [23,24]. However, the limitations of conventional mapping methods remain evident. For instance, in the Ribeira Valley, one of Brazil’s leading banana-producing regions, national initiatives such as the MapBiomas Project failed to correctly classify perennial crop areas, labeling them mostly as “Mosaic of Uses” [25,26].

Given this context, the present study aims to detect banana and peach palm cultivation within the broader land-use and land-cover mosaic of Jacupiranga (São Paulo State), located in the Ribeira Valley. To this end, we employ multi-temporal optical Sentinel-2 imagery combined with a hierarchical classification system supported by field and remote data. The study also evaluates the effectiveness of this approach in enhancing the detection of perennial crops in complex agricultural landscapes, thereby improving the use of geospatial data for agricultural monitoring and supporting sustainable rural development, with a particular emphasis on small- and medium-sized producers.

2. Materials and Methods

2.1. Study Area

Jacupiranga (Figure 1), located in the Ribeira Valley (São Paulo State), features predominantly rugged terrain, with 41.9% of its area classified as strongly undulating (20–45% slope), 13.7% as mountainous (45–75% slope), 23.1% as undulating (8–20% slope), and only 10.9% as flat areas, reaching elevations up to 1280 m [27]. The municipality stands out for its diverse soils, partly attributed to the availability of detailed pedological surveys, with Red-Yellow Argisols (45%) and Haplic Cambisols (41%) as the dominant classes, in addition to Gleisols (≈9%) associated with dissected relief. Land tenure is relatively balanced, comprising 1218 rural properties, of which 854 are smallholdings, 163 small farms, 154 medium farms, and 47 large farms [27].

The Ribeira Valley, where Jacupiranga is located, is characterized by rugged relief and soils predominantly classified as Red Yellow Argisols and Fluvic Cambisols, representative of the lower Ribeira de Iguape basin. These soils are acidic, kaolinitic, and rich in fine particles, characteristics that directly influence their chemical and physical properties [28]. Studies highlight their environmental role in retaining contaminants such as lead (Pb) and cadmium (Cd), owing to their high specific surface area, cation-exchange capacity, and the presence of kaolinite and iron oxides [28]. Thus, Jacupiranga can be considered representative of the geomorphological and pedological context of the Ribeira Valley and is relevant for environmental studies and land management in a region that harbors extensive remnants of the AF and agricultural practices closely linked to biodiversity conservation.

The humid tropical to subtropical climate ensures abundant water availability: potential evapotranspiration is estimated at 921 mm and actual evapotranspiration at 878 mm, indicating a favorable water balance for agricultural production [27]. Figure 2 shows the monthly rainfall (mm) and air temperature (°C) in Jacupiranga, comparing the 1995–2024 climatology with observations from 2024. The region exhibits a strongly seasonal climate, with a 30-year mean annual rainfall of 1738 mm, based on Climate Hazards Center InfraRed Precipitation with Station data (CHIRPS) processed in Google Earth Engine (GEE). The mean air temperature for the same period was 20.9 °C, according to data from ERA5-Land, also obtained in GEE. In contrast, 2024 was atypical, with an annual rainfall of 1476 mm, below the climatological mean in 9 out of 12 months, and a mean air temperature of 21.6 °C, approximately 0.7 °C above the historical average. The intense and protracted rainy season causes persistent cloud cover for several consecutive months, particularly during the summer, which significantly restricts the ability to obtain satellite images without clouds [29].

In 2023, the municipality of Jacupiranga, which covers an area of 704.1 km², stood out in agricultural production with banana cultivation, totaling 80,000 tons, distributed over more than 3000 hectares of harvested area. This volume represented approximately 97.08% of the municipality’s total agricultural production and about 8.19% of the overall banana production in the state of São Paulo. The second most relevant crop was peach palm, which reached a production of 880 tons across 220 hectares, accounting for nearly 2% of agricultural output in the same year [3].

In municipalities such as Jacupiranga, where family farming and smallholdings predominate, productive diversification plays a central role in rural sustainability. By expanding income sources, reducing dependence on a single crop, and improving food security, diversification enables households and communities to consume a wider variety of foods year-round. This practice embodies traditional ecological knowledge, developed empirically and transmitted across generations, that warrants recognition and integration into spatial planning and public policy [30].

An aerial photograph captured during fieldwork (Figure 3) illustrates a representative smallholder property in Jacupiranga (24°37′15.6″ S, 48°02′52.8″ W). The image shows a mosaic production system combining banana and peach palm (Bactris gasipaes) with citrus trees (lemon and orange), complemented by livestock and essential rural infrastructure situated near Atlantic Forest remnants. This example highlights how diversified family farming systems simultaneously sustain livelihoods, strengthen local food security, and contribute to biodiversity conservation across the Ribeira Valley.

2.2. Methodological Approach

The methodological approach to conducting the study involved four main steps: (i) image acquisition and processing using gap-filling and smoothing techniques; (ii) design of a hierarchical classification scheme with three levels of detail to provide greater granularity for classes that tend to be confused with one another; (iii) fieldwork in the municipality of Jacupiranga and imagery inspection to collect reference samples, which were used to prepare datasets composed of both field and remote samples; and (iv) classification using the Random Forest algorithm. These steps were mainly implemented in Google Earth Engine (GEE) and Google Colab environments, with support from QGIS for spatial visualization and post-processing. Finally, classification accuracy was assessed through statistical analyses, including the confusion matrices and derived performance metrics. Figure 4 illustrates the methodological flow used in the study, and these will be discussed in detail in the following topics.

2.3. Image Acquisition

Satellite data used in this study were obtained from the Sentinel-2 SR Harmonized Collection (COPERNICUS/S2_SR_HARMONIZED) using the GEE platform. The analysis focused on the year 2024, and the image collection was filtered by date (1 January 2024 to 31 December 2024) and spatially constrained to the study area.

A sequence of preprocessing steps was applied to prepare the images for time-series analysis and land-use classification: (a) Cloud and shadow masking: A custom function was implemented to mask clouds, shadows, and other low-quality pixels based on the Scene Classification Layer (SCL). The mask removed pixels classified as shadow (3), cloud (4), thin cloud (8), cloud shadow (9), snow (10), and no data (11), ensuring higher-quality input for subsequent processing. (b) Optical band rescaling: Reflectance values for optical bands B2, B3, B4, B5, B8, and B11 were scaled using a factor of 0.0001 following ESA recommendations, while the original SCL band was preserved. (c) Spatial clipping: All images were clipped to the boundaries of the study area using a shapefile, improving processing efficiency. (d) Band Resampling: The shortwave infrared band (B11), initially with a spatial resolution of 20 m, was resampled to 10 m using bilinear interpolation to harmonize spatial resolution across all bands. (e) Index Calculation—NDVI (Normalized Difference Vegetation Index) from Rouse (1974) [31], NDWI (Normalized Difference Water Index) from Gao et al. (1996) [32], and the BSI (Bare Soil Index) from Rikimaru et al. (2002) [33] were computed for each image according to Equations (1)–(3). These indices were selected because they are some of the most efficient indices for land-use mapping [34] and for representing the three fundamental aspects of agricultural and natural environments: vegetation cover, water content, and bare soil, respectively.

NDVI = (B8 − B4)/(B8 + B4)

(1)

NDWI = (B3 − B8)/(B3 + B8)

(2)

BSI = ((B11 + B4) − (B8 + B2))/((B11 + B4) + (B8 + B2))

(3)

where B2 is blue, B3 is green, B4 is red, B8 is near infrared (NIR), and B11 is shortwave infrared 1 (SWIR1).

The processed image collection, now including NDVI, NDWI, and BSI, was exported with a spatial resolution of 10 m and a coordinate reference system of EPSG:4326. Each image export included metadata (e.g., cloud cover percentage, vegetation cover, etc.) saved in CSV format archives.

2.4. Time Series Gap-Filling and Resampling

To reconstruct regular time series and mitigate missing values, each spectral index was compiled into a temporal stack and processed in Google Colab. Using Python 3.12.11 libraries such as rasterio, pandas, scipy, and pywt, pixel-wise interpolation and smoothing were applied: (i) linear interpolation filled temporal gaps based on the original image dates, and (ii) wavelet smoothing reduced high-frequency noise.

Due to cloud contamination and acquisition gaps in the Sentinel-2 time series, several missing observations were identified within the annual index stacks. On average, 67.8% of observations were missing across the study area in 2024 (metadata are available for users). These gaps were addressed exclusively in the temporal domain. For each pixel, a one-dimensional linear interpolation was applied to its time series using the scipy interp1d function, based on the Julian day integers of the original acquisition dates. This procedure connects the nearest valid observations before and after each missing point, ensuring temporal consistency without resorting to spatial neighbors or external images, and has been shown to yield highly accurate estimates in previous studies [35,36].

To guarantee full annual coverage, each time series was resampled to a regular 15-day interval from 1 January to 31 December, thereby producing 25 observations per index and reducing data dimensionality. Pixels with fewer than two valid input observations were flagged and excluded to prevent unreliable extrapolation. To further improve the quality of the reconstructed series, a discrete wavelet transform (db4) with soft thresholding was applied to the interpolated profiles, attenuating high-frequency noise while preserving seasonal trends. The final output consisted of cloud-free and temporally smoothed stacks of NDVI, NDWI, and BSI for 2024, stored as raster series at 15-day intervals. This preprocessing step reduces the influence of missing data and noise on the subsequent classification, although it may slightly attenuate abrupt short-term changes. All Python notebooks with the code for these procedures are available to the reader.

The choice of a 15-day resampling interval was motivated by both environmental and methodological considerations. First, due to the climatic characteristics of the study area, marked by persistent cloud cover, it is not feasible to reliably generate cloud-free composites at 5-day intervals. Second, increasing temporal resolution also expands the dimensionality of the feature space; in order to avoid overfitting and maintain a balanced signal-to-noise ratio given the available sample size, a 15-day interval was adopted. Third, the region does not exhibit extensive multi-cropping systems; instead, land use and cover are largely stable throughout the year, dominated by forests, perennial crops, and pastures, all favored by high water availability. Finally, it was previously demonstrated that even with bimonthly intervals, Sentinel-2 imagery can yield agricultural classifications with error rates below 25% in tropical environments [37].

After interpolation, the wavelet transforms reduced high-frequency noise while preserving seasonal trends and abrupt phenological changes. The Discrete Wavelet Transform (DWT) was implemented using the Daubechies 4 (db4) wavelet, which is well suited for time series with irregular patterns and has been shown to perform effectively in remote sensing applications [38]. The decomposition was performed at Level 1, separating the signal into approximation (low-frequency) and detail (high-frequency) coefficients. To suppress noise, soft thresholding was applied to the detail coefficients. The threshold value was computed adaptively as half of the standard deviation of the interpolated signal (σ/2). The signal was then reconstructed using the inverse DWT, which reintegrated the smoothed trend and discarded residual high-frequency fluctuations. The DWT was chosen over filters such as Savitzky–Golay because it more effectively suppresses high-frequency noise while better preserving the original phenological curve, particularly when employing db4 (i.e., four coefficients), which provides an intermediate level of smoothing [39].

The combined use of linear interpolation and wavelet smoothing ensured a continuous and denoised temporal profile for each index, suitable for input into time-series classification models. The output for each index was a multi-band raster with 25 layers, each corresponding to a 15-day period in 2024. These processed stacks of BSI, NDVI, and NDWI were used as covariates in the subsequent stages of sample extraction and classification.

2.5. Hierarchical Classification System and Sample Collection

Level 1 separated anthropogenic agricultural areas (AAs) from native vegetation (NV). Planted forests were grouped with NV because of their low representation in the study area. For this level, 1000 random sample points (500 per class) evenly distributed were collected using cloud-free Sentinel-2 10 m true and false color compositions and Google Earth high-resolution images, both from 2024.

At Level 2, agricultural areas were subdivided into permanent (PC) and non-permanent crops (NPCs), with 100 samples per class. For this level, observations were hybrid: (i) 83 points for PC (banana and peach palm) were collected on-site during a two-day field campaign between 13 and 14 May 2025 using AgroTag [40], and (ii) 17 points for PC and 100 points for NPC (annual crops and pasturelands) were gathered using cloud-free Sentinel-2 imagery from 2024, along with vegetation indices temporal profiles provided by Embrapa with the SatVeg platform [41]. Figure 5 exhibits some photographs taken during field campaigns.

Level 3 provides a more detailed classification of PC, identifying banana and peach palm plantations. To complement the sample dataset and build a larger sample size, 58 remote samples were manually extracted with the same procedure used in Level 2, supported by geolocation from visited farms. The total number of samples for this level was 83 for bananas and 74 for peach palm. Sampling followed a convenience scheme, as specific production farms were purposefully selected in collaboration with local producers to ensure access and accurate field information. Peach palm (Bactris gasipaes) samples were restricted to the visited farms and were not extrapolated beyond these areas, given the challenges of accurately interpreting this type of cultivation through photointerpretation. This approach ensured data reliability while acknowledging the limited spatial representativeness of the sample set. The spatial distribution of the samples in the Jacupiranga area is shown in Figure 6.

2.6. Land-Use Land-Cover Classification

Raster values were extracted for each sample point using geographic coordinates, projected to the raster CRS. For every vegetation index, each point was associated with 25 consecutive time-series values obtained at 15-day intervals, representing its complete temporal profile. These temporal vectors were then organized into .csv files for each classification level.

The Random Forest (RF) classifier [42] was selected for its robustness in handling high-dimensional and non-parametric data [43]. Seven vegetation index (VI) combinations were tested at each classification level using their complete annual time-series stacks (NDVI, NDWI, and BSI individually; NDVI + NDWI, NDVI + BSI, BSI + NDWI, and NDVI + NDWI + BSI). Each stack contained all 2024 acquisition dates after interpolation and smoothing, providing a full temporal profile for model training. Hyperparameters were optimized through grid search, considering the following ranges: (i) n_estimators = [100, 200, 300]; (ii) max_depth = [3, 5, 7, 9, 11]; and (iii) min_samples_split = [2, 3, 5]. Model validation relied on a repeated stratified K-fold approach (5 folds, 10 repeats) with fixed random seeds to ensure reproducibility. The dataset was split into training (70%) and testing (30%) subsets using a stratified random sampling strategy to preserve class proportions. This ensured robust model evaluation while maintaining representativeness across classes. Performance was assessed using overall accuracy, F1-score, recall, precision, and training time. The best model for each level was selected according to metric performance, with training time serving as a tiebreaker when necessary.

Using the best-performing model and feature set, spatial predictions were generated across the entire study area. Smoothed raster stacks were restructured into pixel-wise temporal vectors, and predictions were processed in batches of 100,000 pixels to optimize memory use. The classified raster was exported as a single-band uint8 GeoTIFF.

Post-processing steps included (i) applying a 3 × 3 modal filter to reduce pixel-level noise and enhance spatial coherence; (ii) masking urban and water classes using MapBiomas land-cover data (Collection 9–2023) [25]; and (iii) enforcing hierarchical consistency across classification levels. Specifically, Level 2 predictions were restricted to pixels labeled as AA in Level 1, while Level 3 predictions (banana or pupunha) were confined to pixels identified as PC in Level 2.

3. Results

3.1. Accuracy Assessment

Table 1. Random Forest performance results by classification level/metric and combination of vegetation indices. Cells highlighted in grey indicate the best combination for each classification level. Arrows indicate the direction of variables: “Combination ↓” represents the list of vegetation index combinations by rows, and “Level →” represents the classification levels by columns.

Level → Combination ↓	Overall Accuracy			Kappa
	Level 1	Level 2	Level 3	Level 1	Level 2	Level 3
NDVI	0.892	0.864	0.847	0.783	0.723	0.692
NDWI	0.868	0.898	0.782	0.735	0.795	0.552
BSI	0.857	0.885	0.782	0.714	0.767	0.547
NDVI + NDWI	0.923	0.855	0.913	0.846	0.766	0.821
NDVI + BSI	0.927	0.878	0.913	0.853	0.752	0.821
NDWI + BSI	0.881	0.858	0.804	0.763	0.713	0.594
NDVI + NDWI + BSI	0.937	0.891	0.934	0.874	0.780	0.864

presents the overall accuracy (OA) and Kappa index for all spectral index combinations tested at classification levels 1, 2, and 3. At Level 1, the combination of all three indices (NDVI + NDWI + BSI) yielded the best performance, with an OA of 93.7% and a Kappa index of 0.874. These results enhanced the role of capturing complementary aspects of canopy vigor, moisture, and soil background in the separability of very distinct classes (native versus anthropogenic). At Level 2, unlike the other levels, the NDWI alone produced the highest results (OA = 89.8%, Kappa = 0.795), outperforming even the index combinations, suggesting that for this intermediate level, water-related information was more decisive in distinguishing between classes. At Level 3, the full combination of indices again achieved the best performance (OA = 93.4%, Kappa = 0.864), highlighting the importance of integrating multiple spectral dimensions in more complex classification tasks.

The confusion matrices for the assessed models using vegetation indices individually or in combination are exhibited in

Table 2. Confusion matrices and commission and omission errors (CE and OE, respectively) from classification levels 1, 2, and 3. Matrix cell values indicate the number of validation samples.

Level 1				Level 2				Level 3
Class	NV	AA	OE%	Class	NPC	PC	OE%	Class	BAN	PAL	OE%
NV	127	11	7.97	NPC	75	11	12.79	BAN	26	0	0
AA	7	143	4.67	PC	4	58	6.45	PAL	3	17	15
CE%	5.22	7.14	-	CE%	5.06	15.94	-	CE%	10.34	0	-

NV is native vegetation, AA represents the anthropogenic agricultural areas, NPCs are the non-perennial crops, PC are the perennial crops, and BAN and PAL represent banana and peach palm cultivation, respectively.

. They indicate that mapped areas will generally align well with reality at Level 1, given the low error rates. The separation between natural and anthropic areas is robust, and errors were consistently below 10%. At Level 2, perennial crops are expected to be overestimated due to high commission error, while non-perennial areas may be partially misclassified. This suggests greater spectral overlap between crop types, especially in perennial classes, where structural, phenological similarities, or even spectral mixing from many crops within one pixel complicate discrimination. Based on the matrix, at Level 3, banana is likely to be overrepresented in the final maps, whereas peach palm will be underestimated, reflecting the asymmetrical error pattern observed.

3.2. Importance of Variables

Figure 7 shows the 20 most relevant variables for distinguishing between land-use and land-cover classes during the training stage, based on the Mean Decrease Accuracy (MDA) metric.

Across all analyzed levels, there was a notable concentration of variables associated with the dry season (between April and September). At the first level, NDWI and BSI variables from late September predominated, underscoring the influence of vegetation moisture and soil background during the transition between dry and wet seasons in distinguishing NV from AA. Notably, NDVI did not rank among the top variables at this stage, suggesting that greenness alone was insufficient for broad land-cover separation.

At the second level, variables specific to this stage dominated the ranking, spanning almost all months from April to December. This pattern evidences the pronounced role of canopy and soil moisture in differentiating perennial systems from other land uses.

At the third level, NDVI variables from the early dry season, particularly April to June, along with December observations, emerged as the most influential. This distribution highlights the importance of canopy vigor and phenological differences in distinguishing banana from peach palm, where greenness indices become the primary discriminators, albeit still enhanced by water- and soil-related indices.

3.3. Digital Classification Results

The classification results from the best-performing classes and hierarchical levels were used to spatialize land-use and land-cover patterns in Jacupiranga (Figure 8). At Level 1, natural vegetation dominated the landscape, occupying approximately 53,000 ha, while anthropogenic areas accounted for around 17,500 ha, underscoring the prevalence of native cover over human-modified land. At Level 2, anthropogenic areas were further stratified, with non-perennial crops prevailing (≈12,400 ha) compared to perennial crops (≈4300 ha). At Level 3, within the perennial class, banana cultivation emerged as the most extensive (≈3458 ha), followed by peach palm with 860 ha.

4. Discussion

4.1. Mapping Banana and Peach Palm Cultivation in Heterogeneous Smallholder Landscapes

Using NDVI, NDWI, and BSI time series across 25 dates in 2024, we achieved overall accuracies above 93% for both Level 1 (native vegetation vs. anthropogenic) and Level 3 (banana vs. peach palm). Predicted areas closely matched state and municipal statistics (IBGE), indicating that the model realistically captures the distribution of perennial crops. These results are comparable to UAV-based or ensemble approaches but with lower cost and greater scalability [44,45].

International experiences corroborate these findings. In Nigeria, [44] combined UAV imagery with Sentinel-2 and SAR data, reporting 90–93% accuracy and Kappa coefficients of 0.86–0.89 for banana plantations. In Australia, [45] obtained 88% user and 79% producer accuracies from 10-cm orthophotos. In Uganda, [46] modeled banana suitability without pixel-level mapping, achieving AUC values of 0.895–0.907.

At Level 2 (perennials vs. non-perennials), NDWI alone delivered 89% accuracy, with commission and omission errors between 5% and 16%. ROC analysis confirmed its superior class separability (mean AUC = 0.837), clearly outperforming NDVI (AUC = 0.751) and BSI (AUC = 0.169). This reflects the ecological setting: in humid, rainforest-dominated landscapes, perennial crops maintain stable canopy water content detectable by NDWI, whereas NDVI tends to saturate under high biomass, and BSI contributes little under dense cover. Future research may explore NDWI in combination with indices less prone to saturation, such as EVI [47].

Sampling design accounted for the fine-grained mosaic of the Brazilian Atlantic Forest, where forest remnants, agroforestry systems, and smallholder plots coexist. To enhance representativeness, we targeted crop–forest interfaces, mixed-use parcels, and topographic gradients. This context-specific strategy mirrors international practices that adapt to landscape structure, farm size, and data availability [44,45,46,48,49].

A distinctive contribution of this study is the explicit mapping of peach palm (Bactris gasipaes) as a separate perennial class. While most tropical mapping efforts focus on bananas [44,45,46] or broader perennial categories [25,26], few differentiate species in diverse agricultural settings. Including peach palm acknowledges its rising ecological and economic role in southern Brazil, improves estimates of cultivated area, and enhances monitoring of agroforestry systems where it coexists with bananas. This novelty not only fills a gap in tropical land-use mapping but also establishes a baseline for future studies and policies related to sustainable palm heart production and non-timber forest products.

4.2. Hierarchical Classification as a Tool for Mapping Diversified Farming Systems

Hierarchical classification has proven to be an effective strategy for addressing the complexity of agricultural systems [49,50]. Unlike flat schemes, hierarchical approaches organize the decision process across successive levels, beginning with broad categories (e.g., natural vegetation versus anthropogenic use) and progressing to finer distinctions, such as specific crop types. This structure enhances class separability and offers flexibility for integrating various data sources [51,52,53,54]. As highlighted in FAO [55] guidelines, hierarchical systems also enhance consistency and comparability across scales, supporting the production of adaptable cartographic outputs. Broader classes are particularly valuable for regional or national monitoring, while subclasses provide insights for local planning, technical assistance, and public policies targeting family farming [52,53,54,55].

Evidence from diverse contexts reinforces this broader applicability. Jiao et al. [56], for instance, mapped coastal wetlands in the Yellow River estuary and showed that combining local classifiers by levels reduced confusion between rice crops and other covers. In Poland, Sentinel-2 imagery processed through a hierarchical scheme improved accuracy from 90% to 99%, particularly in fragmented agricultural mosaics [48]. Similarly, in Turkey, Demirkan et al. [57] observed accuracy gains of 4–10 percentage points compared with traditional approaches, with overall accuracies surpassing 80% across agricultural and forest classes.

In Brazil, several studies have employed hierarchical methods in large-scale agricultural contexts, utilizing Harmonized Landsat and Sentinel-2 (HLS) time series. Parreiras et al. [58], working in western Bahia (Cerrado biome), demonstrated that hierarchical classification efficiently handled spatial and temporal heterogeneity: overall accuracy remained high from Level 1 (95.9%, Kappa 0.917, separating agricultural areas from native vegetation) to Level 3 (91.3%, Kappa 0.808, distinguishing soybeans from other annual crops). Likewise, Bendini et al. [59] classified multiple crop types and patterns using resampled vegetation indices from Landsat 8, reporting unprecedented performance in fragmented agricultural areas.

The methodology adopted here enabled the mapping of diversified systems characteristic of family farming, especially in the Ribeira Valley, where heterogeneous landscapes and productive mosaics prevail. Such landscapes are marked by small plots (often <5 ha) and intercropping, which cannot be captured by medium-resolution sensors (250–30 m) [60]. In this context, high-resolution data combined with machine learning can identify dominant crops at early stages [61]. In future work, the same approach may be extended to other crops common in the region (e.g., passion fruit and guava), though this will require more detailed imagery and field data.

The classification results confirm this potential. At Level 1, the best model estimated approximately 53,000 ha of native forest and 17,500 ha of agricultural areas—figures consistent with MapBiomas, which reported ~52,000 ha of native cover in Jacupiranga [25,26], reinforcing the framework’s reliability for broad-scale mapping. At Level 2, perennial crops were predicted at 4300 ha, compared with 3300 ha reported by the Municipal Agricultural Survey [3] and 4690 ha from the state census (LUPA, 2017–2018) [62]. These values suggest the model more accurately reflects actual extents, especially when contrasted with MapBiomas outputs, which identified only 14 ha of perennial crops in the municipality. At Level 3, integration of NDVI, NDWI, and BSI predicted ~3600 ha of banana production, compared with 3000 ha reported by producers in 2023 [3]. The slight overestimation is consistent with error patterns in the confusion matrices (Table 2). Importantly, although banana and peach palm samples were balanced during training, predictions did not artificially inflate peach palm areas, indicating stable model behavior.

4.3. Advances and Limitations

This study demonstrates significant methodological progress in mapping diverse agricultural systems in tropical environments. The use of dense Sentinel-2 spectral time series with hierarchical classification effectively addressed persistent cloud cover. This approach improved class separation accuracy in small-scale plots and heterogeneous mosaics. These findings support recent evidence that temporal gap-filling techniques can reliably distinguish land uses with similar spectral signatures in fragmented landscapes [63,64,65].

Incorporating both field data and remote samples expanded the temporal and spatial representativeness of the training set. This enhanced model generalization and predictive accuracy. Temporal interpolation methods, compatible with approaches previously validated in tropical regions, such as adaptive NDVI reconstruction [66], maintained continuity in spectral profiles under high cloud incidence. Adopting a 15-day temporal resolution reduced feature dimensionality and stabilized the model, but it also limited detection of rapid phenological changes [67,68].

While optical data remain advantageous for scalability, storage, and interpretability, adding microwave observations could further improve crop discrimination under persistent cloud cover. Combining Sentinel-1 SAR with Sentinel-2 imagery offers particular promise for capturing subtle phenological variations [69,70,71]. Sentinel-2’s spatial resolution is appropriate for regional assessments. However, it can hinder the identification of crops in small plots or intercropped systems, which are typical of family farms and agroforestry in the Ribeira Valley. Classification uncertainties may also arise from spectral overlap between certain crops and native vegetation at specific growth stages [71].

Temporal scalability presents another challenge. This study analyzed a single agricultural year (2024), which was atypical due to below-average rainfall and elevated temperatures (Figure 3). These conditions may reduce temporal generalization. Spectral and canopy patterns differ from climatically typical years [72,73,74]. Broader, multi-year analyses are necessary to understand long-term land-use dynamics, climate responses, and market-driven shifts.

Hierarchical classification systems, despite their benefits, are vulnerable to error propagation [75,76]. Misclassifications at higher levels prevent accurate assignment at lower levels. Grouping spectrally similar classes and limiting hierarchy depth mitigated this effect. In this study, we deliberately avoided using deep or overly complex hierarchies, opting instead to group spectrally similar classes. Nevertheless, some degree of error propagation is an expected outcome of the approach.

Transferring hierarchical schemes to other regions depends on prior knowledge of local conditions. Broad categories (e.g., vegetated vs. non-vegetated) are generally transferable across different contexts. However, finer distinctions require contextual understanding of plot size, crop rotations, intercropping, and management practices [49,50]. Finally, the absence of spatial blocking and the relatively small dataset are other limitations of the study, which may constrain broader generalizations. These issues should be addressed in future studies.

5. Conclusions

Integrating dense Sentinel-2 time series with a hierarchical classification approach proved to be a robust, scalable, and low-cost strategy for mapping diversified agricultural systems in complex tropical landscapes. In Jacupiranga, São Paulo, where family farming coexists with persistent forest cover and high cloud cover, the method achieved accuracies above 93% at the broadest levels (native vegetation vs. anthropogenic use) and at the most detailed (banana vs. peach palm) (>89%), with predicted areas consistent with municipal and state statistics. These values are comparable to those obtained by UAV-based approaches or international ensembles but with greater operational scalability.

Building on these mapping results, the analysis revealed that NDWI was decisive in separating perennial crops from annual crops and pastures in humid environments, outperforming NDVI and BSI. This evidence confirms that indices sensitive to canopy water content are especially effective in tropical forests and fine agricultural mosaics. The explicit inclusion of the peach palm as an independent class represents an important innovation: in addition to filling a gap in tropical mapping, which is generally restricted to bananas, it provides a baseline for policies on agroforestry and non-timber forest products.

Taken together, these results demonstrate that hierarchical approaches facilitate the progressive distinction between classes, maintaining high consistency across scales and supporting regional planning decisions, technical assistance, and public policy formulation for family farming. The study also highlights the potential of digital agriculture technologies to increase the visibility of smallholder farmers, thereby contributing to ecological restoration and sustainable rural management.