Project Acronym: FIRE-RES
Project Name: Innovative technologies and socio-ecological-economic solutions for fire resilient territories in Europe
Call ID: H2020-LC-GD-1-1-2020 (Preventing and fighting extreme wildfires with the integration and demonstration of innovative means)
Work Package: 2
Task Number: 2.1
Lead Beneficiary: ICGC
Contributing Beneficiaries: CTFC, ICGC, INRAE, IRTA, NIBIO
This document was produced under the terms and conditions of Grant Agreement No. 101037419 of the European Commission. It does not necessarily reflect the view of the European Union and in no way anticipates the Commission’s future policy in this area.
Copyright © All rights reserved. This document or any part thereof may not be made public or disclosed, copied or otherwise reproduced or used in any form or by any means, without prior permission in writing from the FIRE-RES Consortium. Neither the FIRE-RES Consortium nor any of its members, their officers, employees or agents shall be liable or responsible, in negligence or otherwise, for any loss, damage or expense whatever sustained by any person as a result of the use, in any manner or form, of any knowledge, information or data contained in this document, or due to any inaccuracy, omission or error therein contained.
All Intellectual Property Rights, know-how and information provided by and/or arising from this document, such as designs, documentation, as well as preparatory material in that regard, is and shall remain the exclusive property of the FIRE-RES Consortium and any of its members or its licensors. Nothing contained in this document shall give, or shall be construed as giving, any right, title, ownership, interest, license or any other right in or to any IP, know-how and information.
The information and views set out in this publication does not necessarily reflect the official opinion of the European Commission. Neither the European Union institutions and bodies nor any person acting on their behalf, may be held responsible for the use which may be made of the information contained therein.
AI: Artificial Intelligence
ALS: Airborne Laser Systems
CBD: Canopy Bulk Density
CBH: Canopy Base Height
CFL: Canopy Fuel Load
CHM: Canopy High Model
CTFC: Centre de Ciència i Tecnologia Forestal de Catalunya
DBH: Diameter at breast height
DSM: Digital Surface Model
DTM: Digital Terrain Model
EO: Earth Observation
IA: Innovative Action
ICGC: Institut Cartogràfic i Geològic de Catalunya
INRAE: Institut National de Recherche pour l'Agriculture, l'Alimentation et l'Environnement
IRTA: Institut de Recerca i Tecnologia Agroalimentàries
LiDAR: Light Detection And Ranging
NDVI: Normalized Differential Vegetation Index
NIBIO: Norwegian Institute of Bioeconomy Research
ONF: French National Forest Service
RADAR: Radio Detection And Ranging
RS: Remote Sensing
SCL: Scene Classification Layer
SWIR: Short Wave Infrared
TLS: Terrestrial LiDAR System
WP: Work Package
WUI: Wildland User Interface
The main purpose of this document, as MS (Milestone) 2.1 associated to Task 2.1: Assessment and upgrade of data acquisition and management systems, fire simulators, vegetation dynamics models and management planning techniques for resistant and resilient landscapes, is to describe the data sources and its usage in the subtasks involved in the Innovative Action 2.1. The main data sources used in this IA are LiDAR, photogrammetric CHM, Sentinel-2 and field work.
This document is focused on the data specifications, the compliance criteria and what regulations are applicable to the work done in the following subtasks:
Subtask 2.1.1. A dynamic high-resolution map of the state of the forest and fuel: It will build from Copernicus satellite data and available upcoming missions e.g. the ESAs “Living Planet Programme”. This will be combined with data from existing remote sensing-based forest resource maps. Moving from mapping to monitoring the action will aim at a 3–5-day update schedule throughout the fire season. The map will provide fire fighters, decision makers and other stakeholders updated information on the state of the landscape namely with respect to fuel amount (IA 2.1).
Subtask 2.1.2. Innovative methodologies for fuel structure assessment: LiDAR digital models (of elevation and surface), as well as point clouds and eventually waveforms (if available) will be processed to compute relevant metrics such as vertical and horizontal distributions of fuel bulk density, fuel availability/continuity, understorey amount, crown base height and canopy height models, or fuel mapping in WUI. The combination of different temporal layers will allow to detect changes and forest recovery after fire events. The trafficability of forest tracks will also be assessed by using innovative Artificial Intelligence methodologies to enhance dispatch decisions and faster and safer access to the forest in case of fire events (IA 2.1)
Subtask 2.1.3. Fire management models, adaptive landscape management strategies and operational management options including machinery specifications (IRTA, CTFC, CNR). These will be derived from WP1 recommendations and corresponding upgrade of existing fire and landscape management tools. For that purpose, parameters and indicators (e.g., forest cover, understory cover, fuel availability) over which managers have control will be selected and incorporated into fire simulators and DSS for cost-efficient evaluation of novel management alternatives under climate and fire scenarios provided by WP1. This subtask also aims to combine preventive activities (fuel reduction) with maximum value recovery of biomass to address circular economy concerns while allowing fuel treatments of larger forested areas thanks to its lower net cost per hectare. It will include management planning techniques to create resistant and resilient landscape (IA 2.2).
To conclude, this document highlights the continuous nature of progress, and the insights gained from past experiences. Exploring how work evolves over time and emphasizing the importance of applying lessons learned to drive further improvement.
Modelling forest status has been extensively accomplished using airborne discrete-return LiDAR (Light Detection and Ranging). This remote sensing technology provides high-resolution, three-dimensional data, enabling detailed analysis and accurate representation of forest structures. Within the scope of Task 2.1, a wide set of applications leverage LiDAR analysis to assess forest conditions. These applications use LiDAR datasets that must meet specific minimum requirements in terms of point density, penetrability, and classification. Ensuring these standards are met is crucial for obtaining reliable data that can effectively support the applications developed.
LiDAR flight campaign in Catalonia was planned to achieve a minimum density of 8 pts/m2 and an average density of 10 pts/m2 on last or single returns on the surface.
The minimum point density required to individually detect trees, as established by several authors (Sparks et al., 2022; Gupta et al., 2020) is 8 pts/m2. Given that WUI analysis primarily relies on single tree detection, this density serves as the minimum requirement for this purpose.
The analysis of tracks trafficability requires a DTM with a spatial resolution of 0.5m. To meet this criterion, considering the minimum distance between ground points obtained with this density, a minimum density of 4pts/m2 is required. Nonetheless, a higher density would enhance the delineation of track and slope boundaries.
It is referenced in the scientific literature that a density of 7-8 points per square meter (pts/m²) is necessary to provide robust measurements of structural diversity in forests for temporal or spatial comparisons (LaRue et al., 2022, Campbell et al., 2018). This should be the minimum reference density for estimating new vegetation morphological metrics.
In the context of fuel metrics estimation with the approach developed in the WP 2.1 (Martin-Ducup et al 2024 in revision) the minimum point density used was 4.5 pts/m² (in Portugal) and the maximum was 203 pts/m² (in France). The average values were 65, 16.2 and 22 pts/m² in France, Portugal and Catalonia respectively (Table 1). The approach is generalist and should not be influenced by point density but a minimum of 7-8 pts/m² is also recommendable for this application.
The LiDAR penetration rate is quantified as the ratio of ground points to the total emitted laser beams within a given area. Factors that may influence penetration rate include the flight epoch during the year, the altitude above ground level (AGL), pulse mode and forest cover (Hyyppä et al. 2004). Morsdorf et al. (2008), Takahashi et al, (2008); and Naesset (2009) proposed AGL as a crucial determinant impacting penetration, while Huang and Shih (2008), Morsdord et al 2008 and Zhao and Popescu 2009 underscore the significance of the relationship between the field of view (FOV) and penetration. Hsu, Wei-Chen et al., (2015) observed that the AGL exhibits a weak correlation with penetration rate under similar vegetation cover. Moreover, they reported that employing the full waveform recording technique could increase penetration rates by approximately 15%. Their study advocates against conducting laser surveys during wet conditions due to the near-zero water responses of airborne topographic LiDAR wavelengths in the near-infrared spectrum, leading to decreased echo returns. Additionally, land-cover type and canopy density are evidently influential factors in determining penetration rates.
Considering the lack of consensus on the factors influencing penetrability and aiming to provide valid recommendations for different types of LiDAR sensors, extents to cover, and forest typologies, it is concluded that the following scenarios should be avoided:
Flying in wet vegetation conditions, especially after extensive rains.
High altitudes Above Ground Level (AGL).
Wide Field Of View (FOV).
Forests with high canopy coverage and density under leaf-on conditions.
Whenever possible and feasible, the full-waveform mode is strongly recommended for dense forests.
In the Catalonia dataset, the LiDAR data acquisition did not have optimal parameters for achieving a high penetration rate. The target was to cover the entire territory of Catalonia (32.107km2) with LiDAR data. Due to economic constraints and operational flight limitations, it was not feasible to implement strictly the recommended practices exposed before. However, the parameters are adequate to achieve a favorable penetration rate. With a maximum AGL altitude of 2200 meters and a FOV of 40º, the average penetration rate achieved in a sample forested area is 75%.
Classification in a LiDAR point cloud is critical for extracting meaningful information from the point cloud. The accurate identification and categorization of different objects and surfaces within the point cloud enable the detailed analysis of terrain, vegetation, and built environments. Effective classification enhances the precision of derived metrics, such as canopy height and structural diversity in forests.
Furthermore, it facilitates the differentiation between ground and non-ground points, which is essential for creating high-resolution Digital Terrain Models (DTMs) and for normalizing heights above ground.
Advanced classification algorithms can automatically identify complex patterns and features, reducing the need for manual intervention and minimizing errors. However, manual review and editing of terrain are still recommended to enhance the quality of the products, especially in the analysis of forest tracks trafficability.
All the LiDAR applications developed in the task 2.1 require a minimum classification of ground, vegetation, and noise. Below, a series of recommendations are proposed to enhance classification processes:
Ground classification. In the developed methods, ground classification has been performed using automatic algorithms from Terrasolid. Specially, the software employs the progressive densification algorithm for filtering, effectively separating ground points from non-ground points. Following this automated process, it is recommended to perform a manual review of the terrain for further refinement and accuracy validation.
Vegetation. It is recommended to use NDVI to classify the vegetation of LiDAR point clouds. In this scenario, an integrated photogrammetric camera captured data in red, green, blue, and infrared channels. These four colour channels were assigned to each point, enabling the calculation of NDVI. All points not previously classified as ground or buildings, with an NDVI above 0.2 and with a height above ground between 0.30 and 70 meters were automatically classified as vegetation.
Noise. Noisy points or outliers are erroneous data points that do not represent actual objects or surfaces. These can be caused by various factors, including sensor errors, atmospheric conditions, or reflections from unintended surfaces. In this case, noise was filtered using HxMap software, by Leica Geosystems, which is the processing software provided by the LiDAR system manufacturer, and subsequently refined with Terrasolid.
Furthermore, the data utilized in this task included classification of buildings, other ground, towers, and power lines, which allowed for optimization of result quality and minimization of errors stemming from vegetation classification.
Buildings. Non-ground points which are in the location of ground class holes are checked for planarity conditions. If these points additionally satisfy an NDVI below 0.2, they are classified as building points. The process is entirely automated using Terrasolid.
Other ground. Points with a height above ground under 0.30 meters and not classified as ground are included in this class.
Towers and power lines. These objects were initially classified automatically using artificial intelligence methods (Carós et al 2024) and subsequently underwent manual revision to enhance the accuracy of the results.
Digital Surface Models (DSMs) are generated from orthophotos using various image matching techniques. Subsequently, these DSMs are normalized with a Digital Terrain Model (DTM) to derive a new product called Canopy Height Models (CHMs).
Various image matching techniques are used, such as feature-based matching (FBM), cost-based matching (CBM) and least squares matching (LSM) to produce highly dense point clouds. It is generated using Trimble/Inpho’s software package MATCH-T DSM. The process follows a hierarchical approach starting from an upper level of the image pyramid and generating an approximate DSM for the next lower pyramid level. Different levels of smoothing can be applied as a function of terrain roughness to filter or reject outliers from the generated point cloud. Large point clouds (>5 million points) are automatically split into a squared tile structure. From the final point cloud (tiles) a raster file with the selected 1-m grid size is interpolated. The same aerial photogrammetric images at 0.25 m–0.35 m used to produce the orthophoto are employed, thus guaranteeing a good consistency between these products.
Several factors can contribute to errors in dense image matching: textureless surfaces, occlusions, illumination variations, sensor characteristics, image artifacts, misalignments, overlapping coverage and computational resources. By addressing these potential sources of error and implementing preprocessing steps and quality control measures it is possible to mitigate errors in dense image matching and improve the accuracy of 3D reconstructions. Below, a series of recommendations are proposed to enhance the accuracy and reliability of dense image matching processes:
Ensure adequate overlap between images during the data capture planning to provide sufficient redundancy for matching. The recommended minimum overlap between images can vary depending on factors such as the terrain complexity, the image resolution, and the characteristics of the dense matching algorithm being used. However, a commonly recommended overlap range is between 60% to 80%.
Acquire imagery under consistent lighting conditions to minimize variations in appearance. Flying at specific times of the day and year when the sun is at an optimal angle can minimize shadows.
Conduct thorough pre-processing, including noise reduction and image enhancement, to improve matching performance.
Use GNSS Ground Control Points (GCPs) to accurately georeference images and improve registration accuracy.
Employ feature-based matching algorithms that can effectively handle textureless surfaces and occlusions.
Utilize multi-view stereo techniques to leverage information from multiple images for robust matching.
Implement quality control measures to identify and mitigate errors, such as outlier rejection and consistency checks.
Regularly calibrate and maintain imaging equipment to ensure accurate and consistent results over time.
Optical satellite data has been used as a source for forest and landscape information for decades. In WP2 in the FIRE-RES project we have used data from the Sentinel-2 satellites, provided as open data by the European Space Agency (ESA) through the Copernicus Programme. The Sentinel-2 satellites acquire data in 13 spectral bands with wavelengths ranging from ~ 440 nm to ~ 2200 nm. The spatial resolution is between 10 m and 60 m, depending on the band. The temporal resolution is ~ 5 days, meaning that every location on land will have a new image every fifth day. Note that in the northern parts of Europe the satellite orbits overlap, and the revisit time is in many locations 2-3 days, when data from all orbits are considered.
Three bands from Sentinel-2 can be combined to images for visual interpretation. A combination of the bands B4, B3 and B2 (red, green, blue) gives an image with natural colors, whereas combination of other bands can produce images of for example Normalized Difference Vegetation Index (NDVI) values. The latter will hold information on vegetation greenness and health. Utilization of Sentinel-2 data is however often not done through direct visualization, but through further processing, typically as input to statistical or machine learning models. The following recommendations are intended for data from Sentinel-2 in models. The presence of clouds is one of the most key factors to consider when using optical satellite imagery. Sentinel-2 images are acquired at fixed points in time, and any clouds within the present scene will be part of the image, obscuring the ground below. The portion of the current scene covered by clouds is given in the metadata for each Sentinel-2 image. A depiction of which pixels are covered by clouds - or are in cloud shadows - is given as part of the scene classification layer which is provided as part of the SAFE format download. Not that the presence of clouds in an image can influence the cloud-free part of the image in at least two ways: degraded georeferencing due to occlusion of ground features and affecting the atmospheric correction.
Baseline versions: The processing of the Level-2A product sometimes changes. The processing version is indicated by a processing baseline version. One should make sure that all Sentinel-2 data involved in building models and applying them have been processed using compatible processing baselines. Otherwise, one must take measures to convert all images to a common baseline.
Mosaics: In order to obtain cloud free images, it is common to combine cloud free pixels from multiple acquisitions into cloud free mosaics, covering larger areas. For some modelling applications mosaics inhibit some undesirable properties. For example, neighboring pixels acquired at different points in time, or unknown adjustments such as normalization of spectral values across acquisitions. Mosaics should be used with care, and for some applications avoided.
Clouds: Acquisitions of satellite imagery over Europe completely without clouds is often not the case. It is therefore necessary to weigh the consequences of a partly cloudy image against having updated information.
Field and ALS data from southeastern France, Spain, and Portugal were used for evaluating the ability of ALS data to describe meaningful fuel metrics. While French field data allowed us to evaluate the ability of ALS data to describe vertical vegetation profiles from ground to canopy top, Portuguese and Spanish data were used to evaluate several metrics related to canopy fuels (i.e. CBH, CFL and CBD). Field and ALS data are described in Table 1 and in the following sections.
Field data used to assess the ability of ALS point clouds to describe the vertical vegetation profile were collected in 2022 by the French National Forest Service (ONF) on 296 10m radius plots in southeastern France. Plots were sampled within a variety of forest types representative of the diversity of vegetation types in French Mediterranean ecosystems with 24 different dominant species and plot heights ranging from 0.5 m to 44 m with an average height of 14.7 m.
Vegetation cover was estimated in seven layers (i.e., 0-0.5 m; 0.5-1 m; 1-2 m; 2-3 m; 3-4 m; 4-5 m; and > 5 m) on each plot. The field approach consisted of a team of two operators, visually estimating the percentage of vegetation cover in each layer. Finally, to facilitate visual estimation, the 10-m plots were divided into four quarters using the mean value in each layer. The percentage of vegetation cover was estimated in tenths. Even if the measurement remains coarse, the objective was to obtain such quantitative data over a large set of forest types, to compare LiDAR estimates to the forester’s perception of fuel density and distribution. This protocol is proved relevant to compare ALS and field data for vertical fuel stratification in forest stands (Marino et al. 2018).
Portuguese data were acquired for six areas, covering approximately 34,109 ha of land mostly occupied by vegetation but also including infrastructure and buildings. The study areas include different forest species distributions. In these areas, 409 circular plots with a radius of 12.62 m (500 m2) were sampled (Mihajlovski et al., 2023). Tree height, crown base height (CBH), and the diameter at breast height (DBH) of each tree within a plot were measured. These metrics were used for estimating canopy fuel load (CFL) and canopy bulk density (CBD) from species-specific allometries for leaf fraction of biomass using the Portuguese National Forest Inventory (PNFI) and general allometries for fine wood element fraction. These data were acquired between April 2020 and June 2021.
Spanish data consist in 141 10 m radius plots with a radius of 10 m distributed across four very different ecological areas in Catalonia. There are 38 plots along the inland northern Mediterranean coast, characterized by a high recurrence of fires and rapid vegetation growth; 38 plots along the inland central Mediterranean coast, located near urban residential areas and affected by drought; 45 plots in the Pyrenees, where there is a large accumulation of forest biomass; and 20 plots in a riparian vegetation area. These plots represent nine of the most common tree species in Catalonia, as well as riparian vegetation and scrublands. For each plot, the CBH from the first living leaf was measured, along with other metrics such as understory height and coverage, as well as the number of trees and tree coverage for each species in the plot.
Data needed to parametrize forest management scenarios came from the ORGEST management guidelines (Piqué et al., 2017), Spanish National Forest Inventories, inputs from WP1 of FIRE-RES project, including field data and prescribed fires data, and expert criteria. The parameterization involved converting typical dasometric parameters that characterize a stand into metrics related to fire behavior (surface fuel and canopy models), to serve as inputs for fire simulators. Three management scenarios were parameterized: 1) multifunctional, focused on production and fire prevention, 2) conversion to silvopastoral systems, and 3) the use of prescribed burns, for the main forest types in Catalonia (9 typologies) and three stages of forest development. Forest stand characteristics before and after the implementation of 127 treatments (following forest management guidelines, FMG) were parameterized, and potential fire behavior was simulated for each case.
Forest management guidelines (FMG) were developed for each main forest type in Catalonia. Once the FMG were defined, the structure of both surface and canopy fuels was parameterized using surface fuel classification algorithms (Krsnik et al., 2020) and canopy biomass allometries (Ruiz-Peinado et al., 2011, 2012). The surface fuel classification algorithms are based on the models of Scott & Burgan (2005). For this parameterization, some of the original parameters (ex: 1h load, Life woody load, moisture of extinction, 1h SAV, among others) were modified to generate specific fuel models for Catalonia. To do so, data from the 4th National Forest Inventory (DGDRPF, 2017) and R packages as Medfuels (De Cáceres et al., 2019) and Medfate (De Cáceres et al., 2023) were used.
To parameterize potential fire behavior, the data required to build meteorological scenarios was taken from Busquets et al. (2019), life fuel moisture data was estimated using Medfate (De Cáceres et al., 2023), and dead fuel moisture data was taken from Rothermel (1983) or Scott & Burgan (2005).
Quality control of a LiDAR flight is crucial to ensure that the collected data meets the required standards of precision and accuracy for specific applications. Below, the key steps followed and recommended to guarantee a right quality assessment of this process are described:
Pre-Flight Planning: It is conducted before data collection to ºensure that the flight planning parameters are optimized to achieve the desired point density on the surface, adequate penetration in forested areas, appropriate coverage of ground control points and complete coverage of the entire area of interest.
System calibration: A calibration flight is performed to determine system alignment, combined with rigorous data analysis to maintain the reliability and accuracy of the system.
In-Flight Data Collection: During the flight, LiDAR data is collected through aerial surveys of the area of interest. It's crucial to maintain a precise and stable flight trajectory to ensure the consistency and quality of the collected data. Furthermore, real-time checks are conducted to ensure that the data collection parameters are in line with predefined requirements.
Flight trajectory, Differential GPS (DGPS). The data from the permanent GNSS stations are used to correct the positional information collected by the airborne GNSS receiver. Visual and statistical analysis of the post-processed trajectory are performed to quantify the accuracy and precision of the data, which must be within 10 centimetres in all the 3 axes (X, Y, Z).
Data validation: After data collection, the post-processing is initiated to process and analyze the collected data. This analysis includes verifying that the captured point cloud and imagery are free from cloud covers, check that there is adequate overlap between adjacent flight lines, confirm that the entire project area has been covered without gaps, verify that the LiDAR point density meets project specifications, ensure that the image resolution is as specified, inspect the data for artefacts such as shadows, reflections or sensor noise and apply noise filtering algorithms to remove any unwanted noise from the LiDAR data.
Images radiometric normalization: A LUT (Look-Up Table) is created for each flight session to ensure consistency in color representation across different flight lines, eliminating or reducing the effects of light or calibration variations. The LUT is then applied to the original images to normalize their radiometric values. Afterwards, normalized color is assigned to each corresponding LiDAR point.
LiDAR accuracy: Strip adjustment techniques are applied to minimize discrepancies between overlapping flight lines. Ground Control Points (GCPs) are used to validate height accuracy of the adjusted flight lines, ensuring that the mean deviations are always within 10 cm.
LiDAR classification: Visual inspection of the classified point cloud is performed to detect inconsistencies in the classification results. Errors are addressed by modifying the parameters of the classification algorithms or through manual editing.
Digital Surface Model and Canopy Height Model: The quality of both LiDAR DTM and DSM was checked against a high number of independent check points (García et al 2022), measured with GNSS at a vertical accuracy of approximately 4 cm. In case of the LiDAR DTM, the check points were located on soccer fields. For the DSM the available country-wide network of photogrammetry control points was used. Only points located on the ground were selected (around one thousand points). From the check points the following empirical vertical accuracy values (Root Mean Squared Errors - RMSE) were derived and in consequence, according to results, recommended has an approach of quality thresholds:
LiDAR DTM: better than 15 cm.
DSM in the Pyrenees: better than 40 cm.
DSM in the rest of Catalonia: better than 30 cm.
CHM in the Pyrenees: better than 45 cm.
CHM in the rest of Catalonia: better than 35 cm.
Since the DSM and the CHM are automatically generated products, their quality can be considerably decreased in areas where the matching algorithm did not achieve optimal results (e.g. in shadow areas).
It should be also noted that in areas covered with some kind of forests and mildly sparse trees the DSM/CHM does not always represent the height of the canopy, depending on the tree density and the presence of foliage.
The criteria below were used for selecting Sentinel-2 optical satellite images for updating fuel and vegetation maps in subtask 2.1.1 in FIRE-RES.
Processing: Level-2A single image Sentinel-2 data. This is data which has been atmospherically corrected to represent the surface reflectance.
Resolution: Spatial resolution of 10m, meaning that the spectral bands with native resolution 20 m were resampled to 10 m.
Clouds: Sentinel-2 images with less than 50% cloud cover.
Scene classification – discarded pixels: Pixels classified as clouds (SCL value 8,9,10), cloud shadows (SCL value 3) and snow (SCL value 11) were discarded from analysis.
Three main forest management scenarios were defined: multifunctional management aimed at wildfire prevention, conversion to sylvopastoral system, and direct prescribed burning. The first two management scenarios can be followed by different slash management treatments: manually with chainsaw, mechanized by truck with a chain-type haulm slasher, or prescribed burning. These three forest management scenarios are all aimed at either reducing the surface and canopy fuel load or at creating a forest structure that reduces the potential fire behavior (higher horizontal and vertical discontinuity). Topographical and accessibility constraints influence which of the three forest management scenarios can be implemented.
The specific parameters that describe forest structure before and after the treatment, required for parameterization, should be integrable into fire simulators and decision support systems (DSS). This integration allows the quantification of the reduction in potential fire behavior following treatment, facilitating an estimation of the cost-effectiveness of the treatment. These parameters include stand height, canopy base height, canopy cover, tree density, DBH, basal area, shrub cover, and shrub height. Forest structure measurements must be taken at stand level through forest inventory plots. If it is not possible to conduct this forest inventory, resources such as LIDAR data and National Forest Inventories can serve as alternatives.
Additional data required to be able to decide on the most cost-effective management alternatives are: forest types, fire behavior related to the different forest structures before and after treatments (FMG), and economic information related to different forest management scenarios and their associated forest management guidelines (FMG), covering both implementation costs and an estimation of potential revenues. Forest cover types do not need to be georeferenced. If national or regional fuel models are not available, the use of standard fuel models such as Rothermel (1983) or Scott & Burgan (2005) is recommended.
Suitability maps for woody crops resulted from a multicriteria analysis combining climatic, soil and topographic indicators. Climatic indicators are usually the ones with more weight when compared to soil and topographic (Alsafadi et al 2019). For this reason, robust climatic data is needed to increase the accuracy of the results. Daily maximum and minimum temperature, as well as accumulated precipitation should be used to correctly identify potential negative impacts like late spring frost. The period including this data should be from at least 30 years to provide representative reference values for the climatic conditions, as recommended by the World Meteorological Organization (WMO, 2017). National meteorology agencies should be able to provide this database at least at 1x1 km resolution which should be the lowest resolution to be used. Downscaling techniques should be considered, especially if soil data is obtained at higher resolution than 1x1 km.
Soil data will probably be the most difficult to obtain. If not available in the study zone the European Soil Data Centre (ESDAC) is a good starting point to obtain all kinds of indicators that may be used in the suitability assessment. Topographic data is easily obtainable from multiple sources and may be adapted to the study zone.
Based on the guidelines provided by the parametrization process (section 3.4.1), the areas requiring mechanical treatment are identified. These are prioritized through the assessment of the territorial stakeholders (in FIRE-RES this role was played by the LL leader). This leads to a well-defined type of intervention and a specific setting (e.g., maintenance of fuel-break infrastructure). A general description of the types of vegetation and the most common silvicultural goals are defined. These entail an unavoidable degree of variability: continuing the case of fuel-break management, these are distributed all over the landscape (at regional level) and may involve different vegetation or be interested by different land uses and protection degrees (e.g., areas of active grazing vs Natura2000 protected sites). The selection of the innovative work system considers the flexibility to operate in the given setting improving the current work system and considering financial capacity of the local stakeholders, the requirement to maintain high levels of manual work or maximize mechanization, other constrains such as soil or aesthetic protection.
The analysis of the vegetation in the Wildland-Urban Interface (WUI) should adhere to the local regulations applicable to the study area. In this case, adherence is directed by the DECREE 123/2005, issued by the Generalitat de Catalunya on June 14, pertaining to measures for forest fire prevention in urbanized areas not directly contiguous with the urban zone.
According to this decree, the minimum width of the protection belt must be 25 meters, and the clearing operations must be conducted to meet canopy fuel and terrain slopes specifications outlined within the decree. Parameters stipulated for the regulation include tree and shrub coverage, tree density and minimum distance between tree trunks.
In France, legal obligations for vegetation clearing (OLD) aim to regulate the spaces between forests, urban areas, and infrastructures to reduce the occurrence of fires and the vulnerability of people and property by reducing the combustible mass and the continuity of vegetation structures. These obligations fall on the owners of buildings, construction sites, and installations of any kind up to a maximum distance of 50 meters, which can be extended to 100 meters, and on the managers of transport infrastructures over a maximum width of 20 meters, in territories or zones identified as fire risk areas (which is the case of most of the Aquitaine living lab) throughout the national territory.
This methodology has been developed following the characteristics defined in the GUIDANCE GUI.INVE.003 V1.2020: “Morphological and usage characteristics of forest roads for the prevention and extinguishing of forest fires”, and also the Operative Guide GUI.INVE.002.V1 "Maneuvers with technical fires 1". Both by the Departament d’Interior de la Generalitat de Catalunya.
This guides contain recommendations for the minimum geometric parameters that forest tracks of the basic network for forest fires must comply with, meeting the specific needs of firefighting vehicles such as fire engines (BRP, heavy rural pump truck).
The parameters established to classify each track based on its trafficability will be tailored to each specific country/area, considering the type of vehicles that will drive on them.
Forest fire prevention legal measures in Catalonia are outlined across several laws and decrees. Decree 64/1995, sets out general forest fire prevention measures; Decree 268/1996, specifies fire prevention measures for areas near power lines; Decree 130/1998, regulates fire prevention along road margins; Law 5/2003 and Decree 123/2005, regulate fire prevention in wildland-urban interface, and Law 20/2009, focuses on preventing and controlling contaminant activities impacting the environment.
Prescribed burning activities are regulated by the following decrees: Decree 312/2006, regulates the use of technical fire by firefighting specialists, and Order MAH/120/2006, regulates the use of prescribed burning to improve pastureland in high-elevation mountainous areas.
Finally, in the Vall d’Aran region, a pioneering integrated fire management plan was approved in March 2022: “Strategic Plan for the Sustainable Fire Regime in Vall d’Aran”.
The workflow for analysing vegetation in the Wildland-Urban Interface (WUI) will be enhanced to enable prompt assessment of vegetation status, facilitating rapid response once LiDAR data is available, thereby accelerating silvicultural works.
The detection of forest tracks trafficability requires an initial position to identify the trafficable area. To eliminate this initial dependency on track position, an Artificial Intelligence (AI) model is being developed to detect forest tracks without the need for prior location data. This model is being trained using the terrain's LiDAR intensity map, shadow map, and slope map. Incorporating orthophotos into the model is under consideration, but it has been temporarily dismissed due to the potential introduction of errors in track segments covered by vegetation or shadows presence.
New training datasets will be created to enhance the AI model for calculating new vegetation morphological metrics. Training work is costly, involving fieldwork for validation and a rigorous visual interpretation of point clouds to accurately classify the various strata and species.
Characterization and detection of changes in forest and vegetation using optical satellite data is a topic with ongoing research and development. The current progress and development in machine learning and AI methods means that there is a potential for improvements of methods and models. For example, a better utilization of spatial or temporal relationships between pixels. Deep learning methods such as semantic segmentation can incorporate the spatial context of each pixel in the modelling process, in contrast to traditional methods which have typically applied modelling at the single pixel level.
The use of pre-trained foundation models can reduce the need for large sets of training data which has been a limitation with some AI model types. Foundation models are generic models trained one a large collection of unlabelled data, which can be fine-tuned to specific modelling tasks with a smaller set of labelled training data. The use of advanced AI models for monitoring and change detection using optical satellite imagery should be subject to further research.
Results of the parameterization and fire behavior simulation will be integrated into decision-support tools for the cost-effective evaluation of management alternatives and the prioritization of treatments for fire prevention planning. Also, these management alternatives and priority areas defined for fire prevention should be transferred to forest owners and forest administration. Additionality, current parameterization of forest management alternatives has been applied to pure forest types but in the future it is expected to include mixed forest types.
The developed analysis of the vegetation WUI offers valuable insights into wildfire prevention strategies. One key lesson learned is the importance of maintaining clearances and managing vegetation density within these zones to mitigate fire risk effectively. Utilizing LiDAR technology for vegetation analysis has proven highly accurate, providing detailed information on vegetation structure and density. However, a significant drawback is the temporal gap between LiDAR data collections, limiting the ability to monitor vegetation changes continuously.
To address this limitation, it's crucial to explore alternative data sources that offer more frequent temporal coverage. For example, photogrammetric CHM and satellite imagery can provide regular updates on vegetation dynamics, allowing for more frequent monitoring of protection strip conditions. Although these current temporal coverages are far from achieving the same precision, scale of work and level of detail as LiDAR, they can still detect significant changes in vegetation conditions, such as forest clearings or substantial vegetation growth that might alter the fire spread risk previously analysed with LiDAR data. Integrating these diverse data sources into a comprehensive monitoring framework can enhance the effectiveness of wildfire prevention efforts by enabling more timely identification of vegetation growth within protection zones.
Knowing in advance whether emergency services can transit through a forest track is crucial for effective emergency response in rural or forested areas. Access to these areas is often limited and can be impeded by various factors such as terrain conditions, vegetation density, and infrastructures. Therefore, conducting pre-assessments of forest tracks to determine their trafficability for emergency vehicles is essential for preparedness and timely response to incidents such as wildfires, medical emergencies, or search and rescue operations.
An important key from analysing the trafficability of forest tracks is the significance of data temporality, similarly to what has already been observed in the WUI. The temporal aspect of input data is critical for ensuring the currency of resulting data, but LiDAR coverages are lacking in terms of data temporality.
Forest track trafficability can indeed be impacted by various factors, including heavy rainfall, strong winds, or other contingencies. These natural events have the potential to alter terrain conditions, affecting the accessibility with which vehicles can traverse these tracks. Therefore, ongoing monitoring and assessment are essential for adapting to changing conditions and detecting alterations in trafficability along forest tracks.
Classification methods of LiDAR point clouds with AI optimize the interpretation process, reducing the time required for human analysis. By leveraging artificial intelligence algorithms, these methods can automatically identify and classify various vegetation strata within LiDAR data. This automation significantly accelerates data processing and facilitates the extraction of valuable information from LiDAR datasets, thereby expanding the range of vegetation metrics available for forest analysis, such as fuel load.
However, while AI-based classification offers notable efficiency gains, it comes with its own set of challenges. One of the primary requirements is the need for extensive training datasets. Training AI models to accurately classify LiDAR point clouds requires large volumes of well-labelled data, which can be huge time-consuming and resource-intensive to collect and prepare. Moreover, ensuring the quality and representativeness of training data is crucial for achieving reliable classification results.
Near real-time monitoring using satellite data is a challenging task, and several aspects can be mentioned for later reference. One key factor is the trade-off between reliable predictions and the utilization of satellite images partly covered with clouds. Clouds are typically present, and with a revisit time of 3-5 days, some locations might experience several weeks between cloud-free acquisitions. Some modelling methods require multiple consecutive satellite images to reduce uncertainty and noise in the predictions. This means that a near real-time monitoring based on optical satellite imagery might in practice have a time lag of several months for certain locations, depending on the methods used.
One other factor to consider for change detection using optical satellite data is the threshold to set for the reliability of prediction. Depending on the use of the data one can choose models and methods to favour either early detection or high certainty of the detected changes. Deciding a balance between these two properties will be needed when using optical satellite data to monitor changes.
[1] Alsafadi, K., Mohammed, S., Habib, H., Kiwan, S., Hennawi, S., & Sharaf, M. (2020). An integration of bioclimatic, soil, and topographic indicators for viticulture suitability using multi-criteria evaluation: A case study in the Western slopes of Jabal Al Arab—Syria. Geocarto International, 35(12), 1466–1488. https://doi.org/10.1080/10106049.2019.1583291
[2] Gupta, A., Byrne, J., Moloney, D., Watson, S., & Yin, H. (2020). Tree annotations in LiDAR data using point densities and convolutional neural networks. IEEE Transactions on Geoscience and Remote Sensing, 58(2), 971–981. https://doi.org/10.1109/TGRS.2019.2942201
[3] Campbell, M. J., Dennison, P. E., Hudak, A. T., Parham, L. M., & Butler, B. W. (2018). Quantifying understory vegetation density using small-footprint airborne lidar. Remote Sensing of Environment, 215, 330–342. https://doi.org/10.1016/j.rse.2018.06.023
[4] Carós, M., Just, A., Seguí, S., & Vitrià, J. (2024). Effective training and inference strategies for point classification in LiDAR scenes. Remote Sensing, 16(2153). https://doi.org/10.3390/rs16122153
[5] García, C., Mora, O., & Pérez-Aragüés, F. (2022). CatLC: Catalonia multiresolution land cover dataset. Scientific Data, 9(554). https://doi.org/10.1038/s41597-022-01674-y
[6] Hsu, W.-C., Shih, P. T.-Y., Chang, H.-C., & Liu, J.-K. (2015). A study on factors affecting airborne LiDAR penetration. Terrestrial, Atmospheric and Oceanic Sciences, 26(241). https://doi.org/10.3319/TAO.2014.12.02.08(EOSI)
[7] Hyyppä, H., Yu, X., Hyyppä, J., Kaartinen, H., Kaasalainen, S., Honkavaara, E., & Rönnholm, P. (2004). Factors affecting the quality of DTM generation in forested areas. International Archives of Photogrammetry, Remote Sensing and Spatial Information Sciences, 36. https://doi.org/10.3390/rs10122014
[8] LaRue, E. A., Fahey, R., Fuson, T. L., Foster, J. R., Matthes, J. H., Krause, K., & Hardiman, B. S. (2022). Evaluating the sensitivity of forest structural diversity characterization to LiDAR point density. Ecosphere, 13(9), e4209. https://doi.org/10.1002/ecs2.4209
[9] Sparks, A. M., Corrao, M. V., & Smith, A. M. S. (2022). Cross-comparison of individual tree detection methods using low and high pulse density airborne laser scanning data. Remote Sensing, 14(3480). https://doi.org/10.3390/rs14143480
[10] Marino, E., Montes, F., Tomé, J. L., Navarro, J. A., & Hernando, C. (2018). Vertical forest structure analysis for wildfire prevention: Comparing airborne laser scanning data and stereoscopic hemispherical images. International Journal of Applied Earth Observation and Geoinformation, 73, 438–449. https://doi.org/10.1016/j.jag.2018.07.015
[11] Martin-Ducup, O., Dupuy, J. L., Soma, M., Guerra-Hernandez, J., Marino, E., Fernandes, P. M., Just, A., Corbera, J., Toutchkov, M., Sorribas, C., Bock, J., Piboule, A., Pirotti, F., & Pimont, F. (2024). Unlocking the potential of ALS data for direct assessment of fuel load and vertical structure. SSRN. https://doi.org/10.2139/ssrn.4779351
[12] Mihajlovski, B., Fernandes, P. M., Pereira, J. M. C., & Guerra-Hernández, J. (2023). Comparing forest understory fuel classification in Portugal using discrete airborne laser scanning data and satellite multi-source remote sensing data. Fire, 6(327). https://doi.org/10.3390/fire6090327
[13] World Meteorological Organization. (2017). WMO guidelines on the calculation of climate normals (WMO-No. 1203). World Meteorological Organization. https://library.wmo.int/index.php?lvl=notice_display&id=19834
[14] Busquets, E., Castellnou, M., González-Olabarria, J.R., & Piqué, M. (2019). Escenaris meteorològics per la simulació d'incendis forestals. Servidor PREVINCAT.
[15] De Cáceres, M., Casals, P., Gabriel, E., & Castro, X. (2019). Scaling-up individual-level allometric equations to predict stand-level fuel loading in Mediterranean shrublands. Annals of Forest Science, 76(3). https://doi.org/10.1007/s13595-019-0873-4
[16] De Cáceres, M., Molowny-Horas, R., Cabon, A., Martínez-Vilalta, J., Mencuccini, M., García-Valdés, R., Nadal-Sala, D., Sabaté, S., Martin-StPaul, N., Morin, X., D’Adamo, F., Batllori, E., & Améztegui, A. (2023). MEDFATE 2.9.3: A trait-enabled model to simulate Mediterranean forest function and dynamics at regional scales. Geoscientific Model Development, 16, 3165–3201. https://doi.org/10.5194/gmd-16-3165-2023
[17] DGDRPF. (2017). Cuarto Inventario Forestal Nacional (2014-2016): Catalunya. Ministerio de Agricultura, Pesca y Alimentación.
[18] Krsnik, G., Busquets, E., Piqué, M., Larrañaga, A., Cardil, A., García-Gonzalo, J., & González, J.R. (2020). Regional Level Data Server for Fire Hazard Evaluation and Fuel Treatments Planning. Remote Sensing, 12(4124), 1–17. https://doi.org/10.3390/rs12244124
[19] Piqué, M., Vericat, P., & Beltrán, M. (2017). Resource communication. ORGEST: Regional guidelines and silvicultural models for sustainable forest management. Forest Systems, 26(2). https://doi.org/10.5424/fs/2017262-10627
[20] Rothermel, R.C. (1983). How to predict the spread and intensity of forest and range fires. USDA Forest Service.
[21] Ruiz-Peinado, R., Montero, G., & del Río, M. (2012). Biomass models to estimate carbon stocks for hardwood tree species. Forest Systems, 21(1), 42–52. https://doi.org/10.5424/fs/2112211-02193
[22] Ruiz-Peinado, R., del Río, M., & Montero, G. (2011). New models for estimating the carbon sink capacity of Spanish softwood species. Forest Systems, 20(1), 176–188. https://doi.org/10.5424/fs/2011201-11643
[23] Scott, J.H., & Burgan, R.E. (2005). Standard Fire Behavior Fuel Models: A Comprehensive Set for Use with Rothermel’s Surface Fire Spread Model. USDA Forest Service.
