Evaluating different classification methods to effectively delineate tree cover on cattle farms in Colombia

 

Beef production in Colombia is the fastest-growing in Latin America. In order to further the growth of beef production, opening trades with the EU and USA would be instrumental; however, footprinting the commodity-linked deforestation is mandatory, as half of tropical deforestation occurs in Latin America. 

 

The European Union (EU) addresses deforestation through its EU deforestation regulation (EUDR), targeting cocoa, rubber, cattle, palm oil, coffee, soy, and wood. The United States of America (USA) advances similar goals with the Fostering Overseas Rule of Law and Environmentally Sound Trade (FOREST) initiative. Both require exporters to show that no deforestation occurred during production from the year 2020 onwards.

 

Thus, this study aims to determine whether accurate and robust annual forest cover detection models can be developed for the Republic of Colombia, using freely available satellite data, ground truth data, and drone images. Specifically, the study evaluates the feasibility of using these data sources to monitor deforestation relevant to regulatory requirements. The study was conducted on two ranches in Monteria, Córdoba Department, with contrasting landscapes: El Rosario ranch, dominated by estrella grass in open spaces and mombasa grass on hilly areas, and Costa Rica ranch, which is mostly hilly and dominated by Toledo grass. The study is a part of a collaborative project between Japan and Colombia under the Science and Technology Research Partnership for Sustainable Development (SATREPS) program.

 

To achieve the study’s objectives, monthly cloud cover assessment was conducted first on both regions from 2020 using Sentinel-2’s Cloud Probability collection. The assessment showed that at least a month of cloud free satellite data can be generated for each region. Then, different land classification methods were evaluated to determine which best fits the application utilizing Sentinel-1 and 2 data. The methods considered were random forest (RF), support vector machine (SVM), gradient tree boost (GTB) machine learning models, mixed tuned match filtering (MTMF), trend analysis using fourier series (FS) and combination of these methods. The training and validation for the methods were derived from drone images and the tree inventory survey conducted over the El Rosario ranch. Each method’s implementation utilized two different approaches on building training dataset, vector-based approach and grid-based approach. The latter was used to consider the coarse resolution of Sentinel-2. To ensure model’s robustness, each model was tested on both ranches. Lastly, the methods were evaluated according to the accuracy metrics and also its integrability with the on-going farm management system in Colombia.

 

The best method identified was the RF using grid-based approach, producing an accuracy of 88.58%, and with the advent of freely accessible geospatial platforms such as Google earth engine, its integrability to any current system is very straightforward. The method is then used to produce an annual forest cover map and detect forest cover loss. Through this, a clear picture of the impact of beef production was created, and the risk assessment requirements by the EU and USA through their regulations were fulfilled.

 

Key words: EU deforestation regulation, cattle farms, remote sensing, classification