SCO FrichesAgricoles

Project completedFrom October 2022 to March 2024, Safer Occitania, in association with CNES and national federation of Safer companies (FN Safer), led SCO Friches Agricoles, an innovative project based on satellite imagery and Artificial Intelligence, enabling automated pre-spotting of agricultural wasteland with a 65% performance rate. The tools and processes developed will enable annual updates of the inventory to be produced, facilitating the implementation of monitoring and limiting the costs involved.  

OVERVIEW

Territorial planning actors are faced with various challenges, including adaptation to climate change, improving food sovereignty, maintaining biodiversity and a healthy and safe living environment.

In Occitania, the development of agricultural wasteland is a major concern in a context where the risk of fire and health problems continue to increase. These risks strongly involve wasteland, which is the source of fire outbreaks and also a reservoir of diseases (e.g. flavescence dorée, vine disease) and parasites (e.g. Xylella fastidosia, a bacterial killer of olive trees). This is why knowledge of wastelands - their location and state of evolution - is essential to prevent and improve the effectiveness of emergency services and the protection of inhabited areas, to monitor and fight against the development of harmful organisms.

The reclamation of wasteland land is also a lever for territorial revitalization and food sovereignty, the requalification of areas to preserve the environment, and carbon sequestration through adapted cultivation practices that preserve living soil.

However, the detection of agricultural wasteland is a complex issue, for which several inventory strategies have been successively carried out by a limited number of DDTs (departmental directorate for the territories) and by the Safer (land development and rural establishment company, Société d'aménagement foncier et d'établissement rural), with disparate resultss.

This is why the SCO FrichesAgricoles project was launched, with 4 objectives: Create an inventory method for automated identification of agricultural wasteland using WaSaBI ©CNES software, trained on the basis of field data from the Vigifriche mobile application. Produce a database of agricultural wastelands in Occitania Develop a web interface for consulting and querying data Enable tools and methods to be transposed to all regions of France.

Methodology

The methodology used to produce the WaSaBI brownfields inventory in Occitania is summarized below, in 5 steps:

• Step 1: Production of the "potential brownfields" layer

In line with the definition of agricultural wasteland¹ adopted by the Occitania regional agricultural wasteland community (Communauté Régionale des Friches Agricoles d'Occitanie CRéFAO), only certain areas are subjected to Artificial Intelligence analysis in order to produce the "WaSaBI wasteland" inventory. For this reason, cartographic pre-processing is carried out to create the "potential wasteland" layer. To do this, various vector data are cross-referenced (RPG sources, OCS GE, BD Foret from IGN, Corinne Land Cover, etc.) in order to eliminate urban and forest areas, as well as agricultural plots on which an activity is declared (details in the diagram below). Further processing is carried out to exclude from the database all plots with a surface area of less than 500 m².

¹Agricultural wasteland is defined as an area or piece of land with no active human occupants, which is consequently not or no longer exploited, productive or even maintained. It is the result of agricultural abandonment of the land (absence of development, definitive abandonment or abandonment over a long period), unlike fallow land, which is merely a temporary period of rest for the soil.

• Step 2: Training WaSaBI with a "ground truth" database

The WaSaBI software is used to train an artificial intelligence model to recognize wasteland based on "ground truths" (wasteland and non-wasteland) identified using the Vigifriche mobile application. To produce the 1st vintage of WaSaBI wastelands, some 5,000 plots of "ground truth wasteland" (2,500 ha) and 127,000 plots of "non-wasteland" (51,000 ha) were evaluated via 6 iterative learning loops. Once the model has been trained, the WaSaBI software is able to automatically recognize "WaSaBI wasteland" in new geographical areas.

• Step 3: Automatic recognition of "WaSaBI wastelands"

The trained artificial intelligence model is made up of 150 decision trees using 96 descriptors to characterize each of the plots in the "potential wasteland land" layer. Based on the descriptor values (radiometric and texture indices, overlap areas, etc.), each of the decision trees makes a choice (wasteland or non-wasteland). The decision as to whether or not the plot belongs is determined by a "vote", which defines the plot as "WaSaBI wasteland" if the number of trees choosing " wasteland" exceeds a certain decision threshold. The result of this "vote" is called the reliability index (value between 0 and 100%). If the results of the various decision trees easily lead to a choice of "WaSaBI wasteland", the reliability index will be high (over 80%). On the other hand, if the choice is complex, the reliability index will be average. Below a certain decision threshold (≃ 30%), the software will classify the plot as "non-WaSaBI wasteland".

• Step 4: WaSaBI performance evaluation

During the training phase, the WaSaBI software calculates various indicators (recall, precision and fscore) that measure the model's ability to identify WaSaBI wasteland accurately and exhaustively. These indicators are calculated on a specific dataset that was not used for training (= validation base made up of 800 ha of ground-truth wastelands spread over some 50 communes). This analysis resulted in a WaSaBI software performance rate of 65%.

• Step 5: WaSaBI wasteland verification

As wasteland is a complex element that changes over time, it's a good idea to check out "WaSaBI wasteland" in the field, before moving on to the operational phase of a project. This step, carried out via the Vigifriche mobile app, will enable the typology of the wasteland to be specified (9 categories available) or, if necessary, its qualification to be modified (land, meadow, vineyard...).

Contributing to the "ground-truth" database, these verifications will help WaSaBI software to learn continuously, thus improving results over time.

Application site(s)

 The Occitania region, in particular :

• Commune of L’Honor de Cos - Tarn et Garonne
• Commune of  Nyls Ponteilla - Pyrénées Orientales
• Commune of Figeac - Lot
• Commune of Moissac - Tarn et Garonne

DATA

Satellites

• Sentinel-2 de niveau 3A
• SPOT6/7

Other

• IGN databases BD TOPO, BD Forêt, OCS-GE and RPG available in OpenData
• Cadastral repository available in OpenData

RESULTS - FINAL PRODUCTS

The results are as follows:

Building an effective method and operational tools

👉 Allows automated pre-identification of agricultural wasteland on a regional scale, with a performance rate of around 65%.

Creation of the 1st regional inventory of agricultural wasteland: the WaSaBI wasteland inventory for the Occitania region

👉 The results demonstrate the benefits of using artificial intelligence and satellite imagery, compared with a method based solely on field surveys or vector data cross-referencing (speed of processing and "cleaning" of the vector layer).

Consolidation of tools and processes

👉 Produce annual inventory updates to facilitate monitoring (quantitative and spatial changes) and limit the costs involved.

Creating a web interface

👉 Provide access to a range of geographic data and functionalities that can be mobilized quickly and easily to help users understand the agricultural, environmental and forestry issues affecting their project area:

• Visualization and export of WaSaBI wasteland and ground-truth wasteland,
• Access to reference funds, environmental and agricultural zoning,
• Access to cadastral data (for beneficiaries),
• Automatic queries of "wasteland" summary sheets at plot, commune and department level,
• Visualization of the reliability index for WaSaBI wasteland data.

OUTLOOK

The operation is due to be extended nationwide, in successive stages, thanks to the mobilization of the National Safer Federation (Fédération Nationale des Safer), which will support other regional Safer companies in the production of their inventories (funding is currently being sought):

• First phase of experimentation and consolidation in several new territories complementary to the initial project (project in Auvergne-Rhône-Alpes, Centre, Nouvelle-Aquitaine, Provence-Alpes-Côte d'Azur),

• Second phase of national inventory production, based on the results of the extended trial,
• Third phase of industrialization of the process to guarantee annual renewal of national data.

To achieve each of the post-SCO development stages, we need to forge partnerships both within the Safer group (FN Safer and regional Safer) and with institutional structures that can provide financial and/or technical support. An agreement with one or more technical service providers is currently under consideration.

References

SCO FrichesAgricoles on GEO Knowledge Hub, a long lasting digital repository created by the Group on Earth Observations: https://doi.org/10.60566/hv421-ze780

PROJECT NEWS

• 26/04/2024: Webinar of the Occitania regional agricultural wasteland community (Communauté Régionale des Friches Agricoles d'Occitanie) with the presentation of the results of the SCO Friches Agricoles and launch of Izifriche, a service for local players, giving access to tools (mobile application, web interface, WaSaBI inventory of wasteland), by jurisdiction.
• 05/04/2024: Agricultural wasteland, discover the Izifriche solution
• 21/12/202 : Présentation du projet SCOFrichesAgricoles et de ses avancées lors de la 11^ème Trimestrielle du SCO France « Des services innovants pour des pratiques agricoles résilientes »