OpenForest: a data catalog for machine learning in forest monitoring

Environmental Data Science, Journal Year: 2025, Volume and Issue: 4

Published: Jan. 1, 2025

Abstract Forests play a crucial role in the Earth’s system processes and provide suite of social economic ecosystem services, but are significantly impacted by human activities, leading to pronounced disruption equilibrium within ecosystems. Advancing forest monitoring worldwide offers advantages mitigating impacts enhancing our comprehension composition, alongside effects climate change. While statistical modeling has traditionally found applications biology, recent strides machine learning computer vision have reached important milestones using remote sensing data, such as tree species identification, crown segmentation, biomass assessments. For this, significance open-access data remains essential data-driven algorithms methodologies. Here, we comprehensive extensive overview 86 datasets across spatial scales, encompassing inventories, ground-based, aerial-based, satellite-based recordings, country or world maps. These grouped OpenForest, dynamic catalog open contributions that strives reference all available datasets. Moreover, context these datasets, aim inspire research applied biology establishing connections between contemporary topics, perspectives, challenges inherent both domains. We hope encourage collaborations among scientists, fostering sharing exploration diverse through application methods for large-scale monitoring. OpenForest is at following url: https://github.com/RolnickLab/OpenForest .

Language: Английский

---

Instance segmentation of individual tree crowns with YOLOv5: A comparison of approaches using the ForInstance benchmark LiDAR dataset

ISPRS Open Journal of Photogrammetry and Remote Sensing, Journal Year: 2023, Volume and Issue: 9, P. 100045 - 100045

Published: Aug. 1, 2023

Fine-grained information on the level of individual trees constitute key components for forest observation enabling management practices tackling effects climate change and loss biodiversity in ecosystems. Such tree crowns (ITC's) can be derived from application ITC segmentation approaches, which utilize remotely sensed data. However, many approaches require prior knowledge about characteristics, is difficult to obtain parameterization. This avoided by adoption data-driven, automated workflows based convolutional neural networks (CNN). To contribute advancements efficient we present a novel approach YOLOv5 CNN. We analyzed performance this comprehensive international unmanned aerial laser scanning (UAV-LS) dataset (ForInstance), covers wide range types. The ForInstance consists 4192 individually annotated high-density point clouds with densities ranging 498 9529 points m-2 collected across 80 sites. original was split into 70% training validation 30% model assessment (test data). For best performing model, observed F1-score 0.74 detection rate (DET %) 64% test outperformed an approach, requires 41% 33% DET %, respectively. Furthermore, tested reduced (498, 50 10 per m-2) performance. YOLO exhibited promising F1-scores 0.69 0.62 even at m-2, respectively, were between 27% 34% better than that knowledge. areas segments resulting close reference (RMSE = 3.19 m-2), suggesting YOLO-derived used derive level.

Language: Английский

---

Incorporation of neighborhood information improves performance of SDB models

Remote Sensing Applications Society and Environment, Journal Year: 2023, Volume and Issue: 32, P. 101033 - 101033

Published: July 20, 2023

Language: Английский

---

Linearly interpolating missing values in time series helps little for land cover classification using recurrent or attention networks

ISPRS Journal of Photogrammetry and Remote Sensing, Journal Year: 2024, Volume and Issue: 212, P. 73 - 95

Published: May 4, 2024

Satellite time series data, widely used for land cover classification, often contain missing values due to cloud contamination, which can negatively affect classification. Numerous strategies have been developed reconstruct the produce regular machine learning classifiers, among compositing followed by linear interpolation is most used. However, classification improvement of has not examined. Recently deep models such as long short term memory (LSTM) and Transformer allow examination they classify with values. In this study, we compared composites (without interpolation) linearly interpolated values) About 18 thousand Harmonized Landsat Sentinel-2 (HLS) images acquired over Amur River Basin China (890,308 km2) in 2021 were composited 14 16-day periods. Two classified, i.e., (i) without that on average 15.35% periods (ii) no The classifications showed (1) between there was < 0.2% overall accuracy differences bidirectional LSTM (Bi-LSTM) 0.5% both smaller than model training randomness; (2) computation be saved using interpolation. findings suggested it unnecessary use time-consuming Bi-LSTM Transformer-based classifications. confirmed experiments sensitivity number cloud-free different legends crop type It implied algorithm cannot reliable historical method more about mitigating inability traditional classifiers handle rather improving Linear necessary capability datasets codes study are made publicly available.

Language: Английский

---

Challenges in data-driven geospatial modeling for environmental research and practice

Nature Communications, Journal Year: 2024, Volume and Issue: 15(1)

Published: Dec. 19, 2024

Machine learning-based geospatial applications offer unique opportunities for environmental monitoring due to domains and scales adaptability computational efficiency. However, the specificity of data introduces biases in straightforward implementations. We identify a streamlined pipeline enhance model accuracy, addressing issues like imbalanced data, spatial autocorrelation, prediction errors, nuances generalization uncertainty estimation. examine tools techniques overcoming these obstacles provide insights into future AI developments. A big picture field is completed from advances processing general, including demands industry-related solutions relevant outcomes applied sciences. In this scoping review, authors explore challenges implementing data-driven models—namely machine learning deep algorithms—in research.

Language: Английский

---

Integrating geographic knowledge into deep learning for spatiotemporal local climate zone mapping derived thermal environment exploration across Chinese climate zones

ISPRS Journal of Photogrammetry and Remote Sensing, Journal Year: 2024, Volume and Issue: 217, P. 53 - 75

Published: Aug. 22, 2024

Language: Английский

---

Supervised machine learning for predicting and interpreting dynamic drivers of plantation forest productivity in northern Tasmania, Australia

Computers and Electronics in Agriculture, Journal Year: 2023, Volume and Issue: 209, P. 107804 - 107804

Published: April 12, 2023

The capacity of a plantation forest to grow and produce timber is locally constrained by topography, climate, soil conditions, external factors such as fire harvesting. Accurate estimation productivity supports effective management. However, efficiently generating accurate models hampered the need gather, process integrate large volumes disparate, high dimensional data that require computationally intensive analysis processing methods. Recent developments in cloud-based machine learning systems offer means address this problem. This research investigates use supervised model predict across pine (Pinus radiata) plantations northern Tasmania, Australia. Forest are generated integrating 23 predictive features, including multi-temporal LiDAR (Light Detection Ranging) derived topographic attributes, climate (rainfall temperature) information, edaphic conditions (geology soil). Five (ML) regression algorithms compared for task: Linear Regression (LR), Polynomial (PR), Decision Trees (DT), Random Forests (RF) Gradient Boosted (GBDT). best performing algorithm, terms optimal bias-variance trade-off, was RF (RMSE 2.08 Bias −0.72) followed closely GBDT 2.13 −0.68) DT 2.94 −0.68). Tuning Model Complexity used provide clear understanding relationship interactions between input features productivity, resulting more interpretable models. In contrast, we conclude results reliable performance than RF, transferability unseen assessing spatial autocorrelation. Across top models, rainfall most important factor driving geological class, position index (TPI), landscape aspect Digital Elevation (DEM). work demonstrates usefulness techniques generate efficient predictions from diverse datasets. Moreover, users afforded ability gain insight into changes affect through time, increasing risks wildfire change identifying contribute tree growth. By delivering framework understand complex dynamic drivers pipeline enhanced systems, managers provided with easily accessible tools maximisation productivity.

Language: Английский

---

Improving spatial transferability of deep learning models for small-field crop yield prediction

ISPRS Open Journal of Photogrammetry and Remote Sensing, Journal Year: 2024, Volume and Issue: 12, P. 100064 - 100064

Published: April 1, 2024

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

Language: Английский

---

The analysis of ecological security and tourist satisfaction of ice-and-snow tourism under deep learning and the Internet of Things

Scientific Reports, Journal Year: 2024, Volume and Issue: 14(1)

Published: May 10, 2024

Abstract This paper aims to propose a prediction method based on Deep Learning (DL) and Internet of Things (IoT) technology, focusing the ecological security tourist satisfaction Ice-and-Snow Tourism (IST) solve practical problems in this field. Accurate predictions IST have been achieved by collecting analyzing environment behavior data combining with DL models, such as convolutional recurrent neural networks. The experimental results show that proposed has significant advantages performance indicators, accuracy, F1 score, Mean Squared Error (MSE), correlation coefficient. Compared other similar methods, improves accuracy 3.2%, score 0.03, MSE 0.006, coefficient 0.06. These emphasize important role IoT technology predicting IST.

Language: Английский

---

Influence of temperate forest autumn leaf phenology on segmentation of tree species from UAV imagery using deep learning

Remote Sensing of Environment, Journal Year: 2024, Volume and Issue: 311, P. 114283 - 114283

Published: July 2, 2024

Remote sensing of forests has become increasingly accessible with the use unoccupied aerial vehicles (UAV), along deep learning, allowing for repeated high-resolution imagery and capturing phenological changes at larger spatial temporal scales. In temperate during autumn, leaf senescence occurs when leaves change colour drop. However, influence in on tree species segmentation using a Convolutional Neural Network (CNN) not yet been evaluated. Here, we acquired UAV over forest Quebec, Canada seven occasions between May October 2021. We segmented labelled 23,000 crowns from 14 different classes to train validate CNN each acquisition. The CNN-based showed highest F1-score (0.72) start colouring early September lowest (0.61) peak fall October. timing events occurring senescence, such as fall, varied substantially within according environmental conditions, leading higher variability remotely sensed signal. Deciduous evergreen that presented distinctive less temporally-variable traits individuals were better classified. While heterogenous remains challenging, learning show high potential mapping species. Our results strong autumn best performance onset this change. • Effect phenology drone is well known. U-Net semantic yieled good tree-cover most was found colours. Species segmented. A dataset crown annotations growing season generated.

Language: Английский

---