Texture analysis for detecting avocado orchards in Uruapan, Mexico

Abstract
This study focuses on addressing ecological challenges located in Michoacán’s Trans-Mexican Volcanic Belt due to the expansion of avocado orchards, which is a risk to local wildlife, water resources, and forest fragmentation, but in the meantime has economic benefits to the area.

Applying machine learning models for land use classification, the research explores the potential improvement through texture analysis, specifically applying Gray-Level Co-occurrence Matrix (GLCM) within the municipality of Uruapan, Michoacán, known for significant avocado cultivation and biodiversity. The investigation encompasses varying resolutions, from high-resolution Planetscope imagery of 3 meters to lower-resolution Sentinel-2 and Landsat 8 data varying from 10 to 60 meters.

In the first research question, Random Forest (RF) and Support Vector Machine (SVM) models are compared for avocado orchard localization using Planetscope data. The RF model has the greater accuracy, achieving an F-score of 0.9545 compared to SVM's 0.9481. Following the application of GLCM texture features provides minimal improvement in RF predictions with a F-score of 0.9551 but adversely affects SVM in the spatial prediction, even with an F-score of 0.9453. Lastly, the study applies the Sentinel-2 and Landsat 8 data with coherent GLCM textures to the machine learning models which shows a decreased F-score and spatial predictions.

Following these results, several recommendations are discussed for alternative accuracy metrics, feature extraction methods, and the application of deep learning with pre-trained models. To finally conclude that with the current results the GLCM texture features did not improve the overall accuracy. However, can be of potential use when accounting for the recommended model improvements.