Mapping cover crop dynamics in Mediterranean perennial cropping systems through remote sensing and machine learning methods

Abstract

About 1.5 Mha of olive orchards are found in the southern Spanish region of Andalusia, representing over 15% of the world olive surface. Some of the most critical rates of soil erosion in Mediterranean agriculture have been found in the local steep slopes of olive orchards (> 61 t ha-1 year-1), where soil is frequently tilled to avoid crop-weeds competition. Conservation agriculture has been proposing alternative strategies such as the use of inter-row cover crops (CC), sown or indigenous, during the period of lowest evaporative demand, with effective (chemical or mechanical) control in spring to avoid significant inter-specific competition for water during the critical period. However, despite the efforts of policy making and scientific research, the use of CC has not been fully adopted yet and a high variability regarding the fraction of ground cover is still observed in the region. In this sense, a better understanding on the main factors driving such variability is required and the development of an up-scaled methodology for mapping and analyzing CC dynamics in olive orchards could considerably contribute to it. In this light, we developed and tested a ‘big data’ approach trained to quantify the fractional green canopy cover (FGCC) as a key diagnostic variable of CC dynamics. We started by collecting the time-series of summer vegetation signals in order to represent FGCC in the absence of CC, assuming that the fraction of bare soil was maximum in summer as CC was controlled before the maximum evaporative period. Therefore, the FGCC of olive trees was directly derived from summer imagery and inter-row FGCC (%CC) was calculated as the difference between 'real time' and summer FGCC (assumed as constant for mature plants in the absence of pruning or other canopy-reducing factors). A validation dataset (N=1600) was built from Deimos-2 satellite data (4x4m), assessed with an image processing package (Fiji Image-J) and based on a binary classification according to the structure of each pixel brightness histogram. Different machine learning (ML) methods taking into account all satellite bands were tested against standard vegetation indices (NDVI, EVI, BI). A higher robustness in predicting FGCC was achieved when using ML methods rather than vegetation indexes, especially for the case of PLS regression, Bayesian Ridge or Multiple Linear Regression Models (MLR). A model based on PLS was tested on Sentinel-2 data for more than 16.500 plots and evaluated with both the Deimos-2 validation dataset and field observations. The PLS model revealed a satisfactory potential to be used from crop field (10x10m) to landscape scale, with a temporal resolution of 5-10 days in cloud-free conditions. Pixel classification showed higher accuracy when distinguishing between higher CC densities (high from >60 to medium

Plant phenotyping relevance

衛星画像と機械学習によりオリーブ園の植生被覆率（FGCC）を推定する手法を開発・検証しており、植物状態の定量的取得が研究の中心である。

Code and data availability

公開本文の所在を確認できませんでした。非公開または購読が必要な可能性があります。

No evidence-backed public reproduction asset is currently recorded.