Accurate land-cover classification maps are essential geospatial resources that support a wide range of societal applications, including urban development, environmental monitoring, and disaster risk management. With the growing availability of satellite remote sensing data, pixel-wise classification methods, which assign a single land-cover class to each image pixel, have become widely adopted. However, when medium-resolution satellite imagery, such as Sentinel-2, is used, a single pixel often contains more than one land-cover type. This phenomenon, commonly referred to as the mixed-pixel problem, leads to the loss of information about minority classes within a pixel, making it difficult to accurately characterize fine-scale surface conditions and limiting the practical utility of classification results. One promising solution to this issue is compositional classification, which estimates the proportion of each land-cover class present within a pixel rather than assigning a single label. This approach retains information about all classes, including those that occupy only a small fraction of a pixel. Compositional data are subject to two mathematical constraints: all values must be non-negative, and the proportions across all classes must sum to one. Yet, the implementation of such compositional classification is challenging, as machine learning or deep learning model architectures do not consider these characteristics properly. This study proposes a deep learning framework for compositional land-cover estimation at 10 m spatial resolution. We explore combinations of input features, model architectures, and loss functions suited to compositional outputs. Rather than seeking a single definitive solution, we aim to provide practical insights into how these design choices interact and influence estimation quality.

As an experiment, we examined two inputs, three models, and two loss functions. Two types of 10-m level input features were examined. The first was spectral reflectance data from 10 multispectral bands (B2-8, B8A, B11, and B12) in Sentinel-2, which directly capture surface characteristics across visible and infrared wavelengths. The second was Embedding V1, a 64-dimensional feature vector generated by AlphaEarth Foundations, a geospatial embedding model that integrates multiple data streams, including optical, radar, and LiDAR observations from multi-temporal satellite imagery. We used OpenEarthMap (OEM), a publicly available global dataset with high-resolution eight land-cover labels at 0.25–0.5 m spatial resolution, as reference data. To generate training targets at medium resolution, we aggregated all OEM pixels within each 10 m pixel and computed the fractional coverage of each land-cover class. This aggregation produced pixel-level composition vectors that served as ground truth for supervised learning, establishing a direct correspondence between medium-resolution inputs and high-resolution reference proportions. Three deep-learning model architectures were tested: a multilayer perceptron (MLP), a two-dimensional convolutional neural network (2D-CNN), and a three-dimensional convolutional neural network (3D-CNN). All models consisted of two fully connected layers followed by a Softmax activation function to ensure outputs formed valid compositional proportions. The CNN-based models were included to capture spatial context from neighboring pixels, while the MLP was used as a simpler baseline that processes each pixel independently. The dataset was split into 70% for training, 10% for validation, and 20% for testing. We compared two loss functions: Mean Absolute Error (MAE), a widely used metric in regression tasks, and the Aitchison distance, a measure derived from compositional data analysis. The Aitchison distance evaluates the relative differences between components rather than their absolute errors by applying a logarithmic transformation and centering to each proportion. To eliminate redundancy associated with the compositional constraint, we applied a linear transformation using an orthonormal basis, projecting the data into a D−1 dimensional space. This metric enables stable distance computation while reflecting the geometric structure and constraints inherent to compositional data. This combination of architectures and loss functions allowed for a systematic comparison across multiple design dimensions. We evaluated the performance of each combination using MAE and the Aitchison distance to identify the most effective framework for compositional land-cover classification mapping. All processing and training are implemented with open-source Python geospatial/ML tools, and we will release code and experiment configurations to support reproducibility.

The experimental results revealed clear differences based on the combination of inputs and model architecture. When using only Sentinel-2 spectral bands as input, 3D-CNN models outperformed the MLP (MAE of 0.1126 versus 0.1185), indicating that spatial context from convolutional operations benefits models when input features are limited to raw reflectance. In contrast, when using Embedding V1 as input, the MLP achieved the best performance (MAE of 0.0989) among all tested models. This suggests that Embedding V1 already encodes rich spatial and semantic information through its pretraining process, making simpler models better suited to leverage these representations without overfitting to their internal structure. Regarding the choice of loss function, the resulting maps showed a meaningful difference between MAE and Aitchison distance. Models trained with MAE tended to produce smoothed predictions, with estimated proportions concentrated near intermediate values and rarely approaching 0 or 1. This smoothing effect is a known limitation of MAE in compositional contexts, where extreme values are physically meaningful. When the Aitchison distance was used instead, the estimated proportion maps showed sharper contrasts and more realistic spatial patterns that better matched the actual distribution of land-cover classes. These results suggest that using a loss function aligned with the mathematical properties of compositional data leads to more credible and interpretable outputs.

The findings demonstrate that the combination of input features, model architecture, and loss function substantially impacts the quality of fractional land-cover estimation. Under our experimental conditions, the MLP model using Embedding V1 and the Aitchison distance achieved the highest accuracy, providing practical insights into the design of compositional land cover classification frameworks in remote sensing.