Accurate crop-type classification from satellite time series is essential for agricultural monitoring. Consequently, various machine learning (ML) algorithms, aiming on enhancing classification performance on data-scarce tasks, have been developed. While previous evaluations demonstrated the effectiveness of these algorithms in certain situations, these studies frequently lacked real-world scenarios. Hence, the performance of the algorithms in challenging practical applications has not yet been profoundly evaluated. To facilitate future research in this domain, we present the first comprehensive benchmark for evaluating supervised and self-supervised learning (SSL) methods for crop-type classification under real-world conditions. This benchmark study relies on the EuroCropsML time-series dataset, which combines farmer-reported crop data with Sentinel-2 satellite observations from Estonia, Latvia, and Portugal. Our findings indicate that variants of Model-Agnostic Meta-Learning (MAML; Finn et al., 2017) achieve slightly higher accuracy compared to supervised transfer learning and SSL. For instance, algorithms belonging to the MAML-family show on average a 31% higher overall accuracy for a 20-shot benchmark task when compared to regular transfer learning. However, compared to simpler transfer learning, the improvement of meta-learning comes at the cost of increased computational demands and training time. Moreover, supervised methods benefit most when pre-trained and fine-tuned on geographically close regions. In addition, while SSL generally lags behind meta-learning, it demonstrates advantages over training from scratch—particularly in capturing fine-grained features essential for real-world crop-type classification—and also surpasses standard transfer learning. This highlights its practical value when labeled pre-training crop data is scarce. Our insights highlight the trade-offs between accuracy and computational demand in selecting supervised machine learning methods for real-world crop-type classification tasks and underscore the difficulties of knowledge transfer across diverse geographic regions. Furthermore, they demonstrate the practical value of SSL approaches when labeled pre-training crop data is scarce. The corresponding code is publicly available at https://github.com/jsreu/EurocropsML-Meta-Learning and https://github.com/jsreu/EuroCropsML-SSL . • Comprehensive benchmark study on real-world crop-type classification from multi-spectral satellite imagery using the highly granular, multi-class EuroCropsML dataset. • Profound evaluation of regular transfer learning, various meta-learning, as well as self-supervised learning algorithms. • Investigation of transfer and meta-learning algorithms’ capabilities for knowledge transfer across different geographical regions. • In-depth analysis of knowledge transfer (specific to transfer and meta-learning approaches) for crop types that are shared or distinct between different geographical regions.

The Ice, Cloud, and Land Elevation Satellite-2 provides a great opportunity to measure water surface and hydrophytic vegetation in complex wetlands. Obtaining reliable signal photons from ICESat-2 data in wetlands is challenging because there are many types of noise photons, such as specular return photons, after-pulse photons, and noise photons caused by sunlight. In addition, the high photon density difference between the water and hydrophytic vegetation makes it difficult to find accurate hydrophytic vegetation photons. Therefore, this research aims to propose a method to obtain high-accuracy signal photons and classify water body photons and hydrophytic vegetation photons in complex wetlands. First, we introduced the modified elevation histogram statistics vector-based (MEHSV) method to filter out noise photons caused by sunlight. The MEHSV method was developed to retain sparse canopy photons. Therefore, the MEHSV method can retain sparse hydrophytic vegetation photons. Second, peak analysis of the elevation histogram statistics removed the specular return photons and after-pulse photons caused by the water surface. Finally, the manually labeled photons and reference water surface level data were used to assess the proposed method. The filtering results showed that the F value of the proposed method achieved 0.99. Compared with other reference methods, the proposed method both preserved hydrophytic vegetation photons being misrecognized and removed all types of noise photons effectively. The water photons and hydrophytic vegetation photons were distinguished accurately. Additionally, the accuracy of water surface level (R2 = 0.97, and RMSE = 0.84 m) witnessed the good performance of the proposed method.

Predicting crop yield using deep learning (DL) and remote sensing is a promising technique in agriculture. In smallholder agriculture (<2 ha), where 84% of the farms operate globally, it is crucial to build a model that can be useful across several fields (high spatial transferability). However, enhancing spatial model transferability in a small-scale setting faces significant challenges, including spatial autocorrelation, heterogeneity and scale dependence of spatial dynamics, as well as the need to address limited data points. This study aimed to test the hypothesis that spatial cross validation (SCV) is a more suitable model validation practice than random cross validation (RCV) to enhance model transferability for spatial prediction in a small-scale farming setting. We compared the performances of DL models that predict crop yield for several settings including three crop types and two DL architectures based on RCV with and without overlapping samples and SCV. Notably, we conducted model performance tests on external, equally sized fields instead of the field used for training. We used high resolution RGB imagery taken with a drone as input. Our results show that the models using SCV outperformed those using RCV when the models were tested on external fields (on average r = 0.37 for SCV, r = 0.18 for RCV with overlap and r = 0.07 without), even though the models using SCV showed a substantially lower performance for cross validation (CV) than those using RCV (r with SCV and RCV w/o overlap = 0.73 and 0.98/0.73, respectively). The results suggest that RCV leads to over-optimism by overfitting the spatial structure and remembering image-specific information (so called memorization). Our study offers the first empirical evidence in agriculture that SCV is preferable to RCV in small field settings for making DL models more transferable.

CNES is currently carrying out a Phase A study to assess the feasibility of a future hyperspectral imaging sensor (10 m spatial resolution) combined with a panchromatic camera (2.5 m spatial resolution). This mission focuses on both high spatial and spectral resolution requirements, as inherited from previous French studies such as HYPEX, HYPXIM, and BIODIVERSITY. To meet user requirements, cost, and instrument compactness constraints, CNES asked the French hyperspectral Mission Advisory Group (MAG), representing a broad French scientific community, to provide recommendations on spectral sampling, particularly in the Short Wave InfraRed (SWIR) for various applications.This paper presents the tests carried out with the aim of defining the optimal spectral sampling and spectral resolution in the SWIR domain for quantitative estimation of physical variables and classification purposes. The targeted applications are geosciences (mineralogy, soil moisture content), forestry (tree species classification, leaf functional traits), coastal and inland waters (bathymetry, water column, bottom classification in shallow water, coastal habitat classification), urban areas (land cover), industrial plumes (aerosols, methane and carbon dioxide), cryosphere (specific surface area, equivalent black carbon concentration), and atmosphere (water vapor, carbon dioxide and aerosols). All the products simulated in this exercise used the same CNES end-to-end processing chain, with realistic instrument parameters, enabling easy comparison between applications. 648 simulations were carried out with different spectral strategies, radiometric calibration performances and signal-to-noise Ratios (SNR): 24 instrument configurations × 25 datasets (22 images + 3 spectral libraries).The results show that spectral sampling up to 20 nm in the SWIR range is sufficient for most applications. However, 10 nm spectral sampling is recommended for applications based on specific absorption bands such as mineralogy, industrial plumes or atmospheric gases. In addition, a slight performance loss is generally observed when radiometric calibration accuracy decreases, with a few exceptions in bathymetry and in the cryosphere for which the observed performance is severely degraded. Finally, most applications can be achieved with a realistic SNR, with the exception of bathymetry, shallow water classification, as well as carbon dioxide and methane estimation, which require the optimistic SNR level tested. On the basis of these results, CNES is currently evaluating the best compromise for designing the future hyperspectral sensor to meet the objectives of priority applications.