Identifying soil management strategies in olive groves through satellite imagery using conventional and machine learning approaches.

The agricultural landscape in Southern Spain, particularly in the Córdoba countryside, is in an on-going transformation due to the expansion of woody crops, like olives orchards [1], which has implications for erosion risk in the area. In that sense, the use of remote sensing to determine actual soil management strategies is a useful technique [2] to calibrate erosion models’ factors, e.g. the cover and management factor in RUSLE [3].

This communication explores the performance of several algorithms for the identification of soil managements in olive orchards. For this, it were considered 3 classes: i) Bare soil (BS), with any combination of herbicide application and/or tillage; ii) Partial soil cover of the lane (alternate lanes of bare soil and cover crop, or narrow cover crop strips, less than 1 m wide, in all the lanes) by temporary cover crops (TCC), defined as those grown during the rainy season (autumn and winter) which are controlled in early spring); and iii) Full ground cover along all the lanes (FCC), also controlled as temporary cover crops. A total of thirty-four olive farms with a known soil management strategy were selected within the study area, located in the countryside of Cordoba (Southern Spain); more details in [1]. Fifty-percent of the farms were used for training, 25% for calibration and 25% for validation, balancing among treatments.

A comparison of five different techniques using the same Sentinel satellite imagery was performed. The techniques were: 1- Support Vector Machines (SVM); 2- Linear Discriminant Analysis (LDA); 3- Random Forest (RF); 4- Boosted Regression Trees (BRT); 5-Dense Neural Networks (DNN). The used dataset consisted of 8 vegetation indexes (ARVI, AVI, EVI, GNDVI, MBI, MCARI, NDVI, SAVI) and ten spectral-bands.

Preliminary results demonstrated that the dataset derived from vegetation indexes exhibited greater accuracy for the five techniques (range 80-99%) than those based on combining several spectral bands (range 40-75%). This assertion was valid for distinguishing between BS and combined TCC&FCC and among BS, TCC and FCC. There was a similar performance among techniques for distinguishing between BS vs. TCC&FCC. For distinguishing among BS, TCC and FCC, LDA and DNN showed the best results. Overall the predictive capability using spectral indexes worsen for distinguishing among three treatments (around 80%) as compared for BS vs. TCC&FCC (around 99%).

This study provides a comparative framework for assessing the response of spectral indices and spectral-bands to different soil management strategies widely used in Mediterranean conditions.

Acknowledgements: This work is supported by the projects SCALE (EJP Soil Horizon 2020 GA 862695), TUdi (Horizon 2020, GA 101000224) and PID2019-105793RB-I00 (Spanish Ministry of Science and Innovation).

References:

[1] Guzmán et al. 2022. Expansion of olive orchards and their impact on the cultivation and landscape through a case study in the countryside of Cordoba (Spain). Land Use Policy, 116, 106065.

[2] Almagro et al. 2019. Improving cover and management factor (C-factor) estimation using remote sensing approaches for tropical regions. International Soil and Water Conservation Research., 7(4), 325–334.

[3] Renard et al. 1997. Agricultural Handbook 703, USDA-ARS. Washington, DC.