Convolutional neural networks (CNNs) are routinely reported at near-perfect accuracy for image-based fire detection, but almost always on test data drawn from the training distribution, so their reliability under domain shift remains largely unmeasured. We benchmark ten CNN architectures for binary fire/no-fire classification across two distinct domains. Each model is trained at three input resolutions (32×32, 64×64, and 128×128 px) over five random seeds (150 configurations in total), and timed for inference on an RTX 5090 GPU and a Raspberry Pi 5 edge CPU. Internal and external performance diverge sharply due to architectural inductive biases: models within 0.5 points on internal validation separate drastically externally. MobileNet-V3-Large loses 15.7 points, while DenseNet-121 transfers best (84.99$\pm$2.48% external accuracy, 92.51% ROC-AUC) due to structural feature reuse. For latency-constrained deployment, ResNet-18 at 32×32 is fastest on the GPU (0.0354 ms per image), while MobileNet-V2 leads on the edge CPU (13.22 ms). The accuracy gained from higher resolutions is strictly architecture-dependent, exposing instability in models prone to high-frequency overfitting. We argue that input resolution must be treated as a hardware-constrained hyperparameter, and we provide concrete, computation-aware guidelines for real-world deployment.
The integration of SAR and optical satellite time series has been widely applied in crop mapping, while previous studies have mainly employed traditional machine learning approaches with information-lossy temporal compositing. Leveraging the capability of Transformer-based models for irregular time series, this study investigated Sentinel-1's contribution to Landsat 8 and Sentinel-2 for within-season major crop type classification using six HLS tiles across major U.S. agricultural regions and CDL-derived samples. A hierarchical Transformer-based model architecture was designed to flexibly fuse the raw irregular time series from multi-sensor time series. Seven sensor combination cases were performed to investigate the individual and combined crop type classification contributions of Landsat 8, Sentinel-2, and Sentinel-1 time series. Using the raw irregular time series improved overall accuracy (OA) by 0.7% during entire crop season, and by up to approximately 2% during the early season compared to temporally composited data. When only current-year observations were used, Sentinel-1 provided improvements at some early-season mapping dates, whereas its contribution became limited (<0.1% OA difference) from mid-season onward. With one year of historical data included, the OA difference remained <0.2% throughout the season. Feature analyses showed limited incremental Sentinel-1 contribution: t-SNE Silhouette Scores were similar for only optical and optical and SAR inputs and gradient-based sensitivity scores for SAR VV/VH were <0.01, compared with 1.0 for the optical green band. These findings indicate that, under the evaluated Transformer framework, raw Sentinel-1 VV/VH temporal observations provided limited incremental classification benefit beyond Landsat 8 and Sentinel-2 time series.
Flood change detection based on Synthetic Aperture Radar (SAR) is a critical component of disaster management. Existing methods primarily concentrate on developing building change detection systems using optical images, which poses significant challenges in adapting to the intricate structural characteristics of floods. SAR images are susceptible to various forms of degradation, including speckle noise, the presence of dark targets that resemble water bodies, and pseudo-changes. In particular, floods have a wide distribution and complex structures. To address these issues, a global-local feature collaborative and temporal-channel enhancement network (GLTENet) is proposed to enhance the capability of extracting flood change features. First, a convolution visual state-space (CVSS) block is designed to extract global features while simultaneously capturing local features, in order to accommodate both large scale and fragmented flooded areas. Second, to highlight change regions and suppress the adverse effects of false changes, we construct a temporal channel difference enhancement module (TCDEM) to mine differential features from bi-temporal images. Subsequently, multiscale feature aggregation module (MFAM) was developed to aggregate difference feature maps from various scales and enhance the structural integrity of the change maps. The experimental results demonstrate the efficacy of the proposed method, as evidenced by its exceptional performance on the S1GFloods and OMBRIA flood change datasets. The method attained F1 scores of 97.62% and 85.15%, respectively, along with IoU scores of 95.35% and 74.15%. Furthermore, the model shows promising transferability to other change detection applications beyond floods, achieving an F1 score of 95% and an IoU of 90.48% on the WHU-CD building change dataset.
Mapping ecological peatland types is vital for greenhouse gas (GHG) inventories and biodiversity conservation. As significant GHG sources, boreal peatlands undergo changes due to peat extraction cessation and restoration. Therefore, detailed land use information is essential for guiding environmental policies regarding EU biodiversity goals. Remote sensing (RS) and machine learning (ML) offer efficient large-scale mapping methods for peatland ecotypes. However, complex gradational vegetation structures, such as dense canopies and diverse microtopography, challenge classification success. Multimodal RS data, including optical, radar, and LiDAR, improve mapping but increase computational complexity. Additionally, sparse ground-truth data in remote areas complicates ML training. This study compares four computationally inexpensive ML models that integrate a large set of 92 multimodal predictors to classify an extensive number of 39 peatland site types and their five fertility levels. We incorporate a novel feature selection and evaluation procedure with genetic algorithm and permutation importance for optimizing the classification process. Results produced on three study areas in different mire vegetation zones in Finland show that prediction accuracy for peatland site type and fertility level range from 34 48% and 41–59% (majority baseline 9–25% and 16–43%), respectively, with random forest, logistic regression and k-nearest neighbors performing similarly. The classification performance is restricted by the limited training data and depends on the selected set of predictor variables highlighting the influence of local geographical factors. These results can facilitate national-scale mapping of boreal peatlands to improve the GHG inventory and sustainable management of biodiversity-rich eutrophic fens and naturally forested peatland site types.
GNSS Reflectometry (GNSS-R) is a technique first proposed in 1993 to address the challenge of sampling ocean topography at the mesoscale. This paper aims to recount the genesis of GNSS-R first-hand, highlight the major events that shaped its development, and review the achievements in GNSS-R altimetry from its inception to the present day. In other words, a first-hand historical perspective rather than a comprehensive review of all GNSS-R developments.
