Ten deep learning techniques to address small data problems with remote sensing

International Journal of Applied Earth Observation and Geoinformation, Journal Year: 2023, Volume and Issue: 125, P. 103569 - 103569

Published: Nov. 18, 2023

Researchers and engineers have increasingly used Deep Learning (DL) for a variety of Remote Sensing (RS) tasks. However, data from local observations or via ground truth is often quite limited training DL models, especially when these models represent key socio-environmental problems, such as the monitoring extreme, destructive climate events, biodiversity, sudden changes in ecosystem states. Such cases, also known small pose significant methodological challenges. This review summarises challenges RS domain possibility using emerging techniques to overcome them. We show that problem common challenge across disciplines scales results poor model generalisability transferability. then introduce an overview ten promising techniques: transfer learning, self-supervised semi-supervised few-shot zero-shot active weakly supervised multitask process-aware ensemble learning; we include validation technique spatial k-fold cross validation. Our particular contribution was develop flowchart helps users select which use given by answering few questions. hope our article facilitate applications tackle societally important environmental problems with reference data.

Language: Английский

---

Automatic detection of snow breakage at single tree level using YOLOv5 applied to UAV imagery

International Journal of Applied Earth Observation and Geoinformation, Journal Year: 2022, Volume and Issue: 112, P. 102946 - 102946

Published: Aug. 1, 2022

The assessment of forest abiotic damages such as snow breakage is important to ensure compensation owners. Currently, information on the extent gathered through time-consuming and potentially biased field surveys. In situations where surveys are still common practice, unmanned aerial vehicles (UAVs) increasingly being used provide a more cost-efficient objective methods answer needs. Further, advent sophisticated computer vision techniques convolutional neural networks (CNNs) offers new ways analyze image data efficiently accurately. We proposed an object detection method automatically identify trees classify them according damage by based YOLO CNN architecture. UAV imagery collected across 89 study areas over course entire year were manually annotated into total >55 K single classified healthy, damaged, or dead. trees, along with corresponding train YOLOv5 model. Furthermore, we tested effect seasonality, varying atmospheric lighting conditions model's performance. Based independent test set found that general model including all (i.e. any seasons, conditions, time day) outperformed other scenarios precision = 62 %; recall 61 %). despite fact damaged represented minority class 16 % trees), they detected largest (76 %) (78 Finally, transferred well variation in illumination making it suitable for usage acquisition.

Language: Английский

---

Out-of-year corn yield prediction at field-scale using Sentinel-2 satellite imagery and machine learning methods

Computers and Electronics in Agriculture, Journal Year: 2023, Volume and Issue: 209, P. 107807 - 107807

Published: April 3, 2023

Crop yield prediction for an ongoing season is crucial food security interventions and commodity markets decisions such as inventory management, understanding trends variability. This work considers corn at field-scale with input variables derived from satellite environmental data. data were obtained consecutively 2017 to 2021 a total of 1164 fields in the US states Iowa Nebraska. We forecast "out-of-year", i.e. we test year using machine learning methods trained on other years. study includes evaluating what spectral information raw Sentinel-2 bands best explains observed variability yields, but also how time considered temporal resampling. found that resampling annual series thermal biophysical parameters estimates increased R2 average by 0.25 0.42 when extrapolate performed different ones covered training samples, compared calendar spectrum.

Language: Английский

---

From spectra to plant functional traits: Transferable multi-trait models from heterogeneous and sparse data

Remote Sensing of Environment, Journal Year: 2023, Volume and Issue: 292, P. 113580 - 113580

Published: April 21, 2023

Language: Английский

---

UAV-based reference data for the prediction of fractional cover of standing deadwood from Sentinel time series

ISPRS Open Journal of Photogrammetry and Remote Sensing, Journal Year: 2023, Volume and Issue: 8, P. 100034 - 100034

Published: March 8, 2023

Increasing tree mortality due to climate change has been observed globally. Remote sensing is a suitable means for detecting and proven effective the assessment of abrupt large-scale stand-replacing disturbances, such as those caused by windthrow, clear-cut harvesting, or wildfire. Non-stand replacing events (e.g., drought) are more difficult detect with satellite data – especially across regions forest types. A common limitation this availability spatially explicit reference data. To address issue, we propose an automated generation using uncrewed aerial vehicles (UAV) deep learning-based pattern recognition. In study, used convolutional neural networks (CNN) semantically segment crowns standing dead trees from 176 UAV-based very high-resolution (<4 cm) RGB-orthomosaics that acquired over six in Germany Finland between 2017 2021. The local-level CNN-predictions were then extrapolated landscape-level Sentinel-1 (i.e., backscatter interferometric coherence), Sentinel-2 time series, long short term memory (LSTM) predict cover fraction deadwood per Sentinel-pixel. CNN-based segmentation UAV imagery was accurate (F1-score = 0.85) consistent different study sites years. Best results LSTM-based extrapolation fractional -2 series achieved all available --2 bands, kernel normalized difference vegetation index (kNDVI), water (NDWI) (Pearson's r 0.66, total least squares regression slope 1.58). predictions showed high spatial detail transferable Our highlight effectiveness algorithms rapid large areas imagery. Potential improving presented upscaling approach found particularly ensuring temporal consistency two sources co-registration medium resolution data). increasing publicly on sharing platforms combined mapping will further increase potential multi-scale approaches.

Language: Английский

---

Evaluating capabilities of machine learning algorithms for aquatic vegetation classification in temperate wetlands using multi-temporal Sentinel-2 data

International Journal of Applied Earth Observation and Geoinformation, Journal Year: 2023, Volume and Issue: 117, P. 103202 - 103202

Published: Jan. 25, 2023

Different perspectives use of machine learning (ML) algorithms have proven their performance depends on the quality reference data. This is particularly true when targets are complex environments, such as wetlands, which a vast majority studies site-specific and based single date. With this work, an extensive dataset about 400,000 samples was collected, covering nine different sites multiple seasons, to be considered representative temperate wetland vegetation communities at continental scale. Starting from dataset, selected ML classifiers compared for detailed type mapping, using spectral indices (SI) derived multi-temporal composites Sentinel-2 input. Global per-class accuracy metrics were computed four independent training testing subsets, extracted overall impacts input features variation in number covered assessed. Our results show generally higher predictive power ensemble methods, Random Forest (RF) eXtreme Gradient Boosting (XGBoost), standalone ones, with notable exception Support Vector Machine (SVM); latter fact, algorithm that scored highest (0.977 ± 0.001) F-score all target classes. Decreasing resulted classification losses, less marked RF than SVM, while showed more stability thus indicating SVM stronger transferability XGBoost.

Language: Английский

---

A deep learning approach for deriving winter wheat phenology from optical and SAR time series at field level

Remote Sensing of Environment, Journal Year: 2023, Volume and Issue: 298, P. 113800 - 113800

Published: Sept. 21, 2023

Information on crop phenology is essential when aiming to better understand the impacts of climate and change, management practices, environmental conditions agricultural production. Today's novel optical radar satellite data with increasing spatial temporal resolution provide great opportunities derive such information. However, so far, we largely lack methods that leverage this detailed information at field level. We here propose a method based dense time series from Sentinel-1, Sentinel-2, Landsat 8 detect start seven phenological stages winter wheat seeding harvest. built different feature sets these input compared their performance for training one-dimensional U-Net. The model was evaluated using comprehensive reference set national network covering 16,000 observations 2017 2020 in Germany against baseline by Random Forest model. Our results show are differently well suited detection due unique characteristics signal processing. combination both types showed best 50.1% 65.5% being predicted an absolute error less than six days. Especially late can be with, e.g., coefficient determination (R2) between 0.51 0.62 harvest, while earlier like stem elongation remain challenge (R2 0.06 0.28). Moreover, our indicate meteorological have comparatively low explanatory potential fine-scale developments wheat. Overall, demonstrate image Sentinel sensor constellations versatility deep learning models determining timing.

Language: Английский

---

Accurate delineation of individual tree crowns in tropical forests from aerial RGB imagery using Mask R‐CNN

Remote Sensing in Ecology and Conservation, Journal Year: 2023, Volume and Issue: 9(5), P. 641 - 655

Published: May 13, 2023

Abstract Tropical forests are a major component of the global carbon cycle and home to two‐thirds terrestrial species. Upper‐canopy trees store majority forest can be vulnerable drought events storms. Monitoring their growth mortality is essential understanding resilience climate change, but in context storage, large underrepresented traditional field surveys, so estimates poorly constrained. Aerial photographs provide spectral textural information discriminate between tree crowns diverse, complex tropical canopies, potentially opening door landscape monitoring trees. Here we describe new deep convolutional neural network method, Detectree2 , which builds on Mask R‐CNN computer vision framework recognize irregular edges individual from airborne RGB imagery. We trained evaluated this model with 3797 manually delineated at three sites Malaysian Borneo one site French Guiana. As an example application, combined delineations repeat lidar surveys (taken 3 6 years apart) four estimate upper‐canopy 65 000 across 14 km 2 aerial images. The skill automatic method delineating unseen test was good ( F 1 score = 0.64) for tallest category excellent 0.74). predicted previous studies, found that rate declined height tall had higher rates than intermediate‐size Our approach demonstrates learning methods automatically segment widely accessible This tool (provided as open‐source Python package) has many potential applications ecology conservation, estimating stocks phenology restoration. package available install https://github.com/PatBall1/Detectree2 .

Language: Английский

---

Deforestation detection using a spatio-temporal deep learning approach with synthetic aperture radar and multispectral images

ISPRS Journal of Photogrammetry and Remote Sensing, Journal Year: 2023, Volume and Issue: 199, P. 87 - 101

Published: April 8, 2023

Language: Английский

---

Towards operational UAV-based forest health monitoring: Species identification and crown condition assessment by means of deep learning

Computers and Electronics in Agriculture, Journal Year: 2024, Volume and Issue: 219, P. 108785 - 108785

Published: March 6, 2024

Uncrewed Aerial Vehicles (UAVs) have emerged as a promising tool for complementing terrestrial surveys, offering unique advantages forest health monitoring (FHM). UAVs the potential to improve or even replace core tasks such crown condition assessment, bridging gap between ground-based surveys and traditional remote sensing platforms. However, present approaches not yet fully exploited very high temporal resolution flexible convenient utilization that offer under cloudy skies. In this paper, we provide standardized data pipeline semi-automatically generate reference by merging UAV-based related species-specific health. Furthermore, investigated of Convolutional Neural Networks (CNNs) classify main tree species their conditions based on data. Therefore, acquired multispectral drone imagery 235 different ICP large scale plots (Level-I plots) distributed across Bavaria three consecutive years (2020–2022). Using highly heterogeneous time-series dataset, encompassing diverse weather lighting conditions, stand characteristics, spatial distribution study areas, successfully classified five species, genus level classes dead trees, including status occurring in Germany. This way managed 14 distinct with an average macro F1-score 0.61 using EfficientNet CNN architecture. The highest class-specific apart from class trees (0.97) was achieved Picea abies healthy (0.80). If participating countries Forests program adopt our approach harmonize monitoring, many could be reduced replaced, leading significant time cost savings. We open-source analysis strategies can potentially extended throughout Europe. Our findings demonstrate UAV deep learning modernize management efficiency sustainability. recommend integrating drones ground systems take advantage benefits.

Language: Английский

---