Census Parcels Cropping System Classification from Multitemporal Remote Imagery: A Proposed Universal Methodology

Abstract

A procedure named CROPCLASS was developed to semi-automate census parcel crop assessment in any agricultural area using multitemporal remote images. For each area, CROPCLASS consists of a) a definition of census parcels through vector files in all of the images; b) the extraction of spectral bands (SB) and key vegetation index (VI) average values for each parcel and image; c) the conformation of a matrix data (MD) of the extracted information; d) the classification of MD decision trees (DT) and Structured Query Language (SQL) crop predictive model definition also based on preliminary land-use ground-truth work in a reduced number of parcels; and e) the implementation of predictive models to classify unidentified parcels land uses. The software named CROPCLASS-2.0 was developed to semi-automatically perform the described procedure in an economically feasible manner. The CROPCLASS methodology was validated using seven GeoEye-1 satellite images that were taken over the LaVentilla area (Southern Spain) from April to October 2010 at 3- to 4-week intervals. The studied region was visited every 3 weeks, identifying 12 crops and others land uses in 311 parcels. The DT training models for each cropping system were assessed at a 95% to 100% overall accuracy (OA) for each crop within its corresponding cropping systems. The DT training models that were used to directly identify the individual crops were assessed with 80.7% OA, with a user accuracy of approximately 80% or higher for most crops. Generally, the DT model accuracy was similar using the seven images that were taken at approximately one-month intervals or a set of three images that were taken during early spring, summer and autumn, or set of two images that were taken at about 2 to 3 months interval. The classification of the unidentified parcels for the individual crops was achieved with an OA of 79.5%.

Object-Based Greenhouse Horticultural Crop Identification from Multi-Temporal Satellite Imagery: A Case Study in Almeria, Spain

“…Classification through the DT algorithm is increasingly applied in remote sensing data. In fact, it has been the selected classifier in recent OBIA investigations focused on outdoor crop identification [24,25,48]. Other advantages of DT classifiers are: (1) they fit well in the OBIA procedure; (2) DTs are computationally fast and make no assumptions regarding the distribution of the data; (3) they are able to take numerous input variables and perform rapid classification without being severely affected by the "curse of dimensionality"; and (4) the DT methods provide quantitative measurements of each variable's relative contribution to the classification results, so allowing users to rank the importance of input variables.…”

“…This classification accuracy rose up to 88% by using support vector machine (SVM) or multilayer perceptron (MLP) neural network classifiers [51]. In another study, twelve outdoor crops and two non-vegetative land use classes were classified with 80.7% OA at cadastral parcel levels through multi-temporal remote sensing images [48]. Furthermore, working on mapping outdoor crops (classes corn, soybean and others), an OA of 88% was reported by Zhong et al [2].…”

Greenhouse Crop Identification from Multi-Temporal Multi-Sensor Satellite Imagery Using Object-Based Approach: A Case Study from Almería (Spain)

“…Inglada et al [67] reported an OA above 80% for most sites around the world using RF. In another study, 12 outdoor crops and two non-vegetative land-use classes were classified with 80.7% OA at the cadastral parcel level through multi-temporal remote sensing images [68]. Using DT and 13 Landsat 8 multi-temporal images, Patil et al [69] achieved an OA of 81%, while an OA of 88% was reported by Zhong et al [70] working on corn, soybean, and others, and using Landsat imagery.…”

Crop Classification from Sentinel-2 Time Series with Temporal Convolutional Neural Networks

“…vegetation indices such as NDVI or EVI). For instance, [2] proposes a method based on a Decision Tree (DT) that considers as features the mean value of the pixels within each parcel for several spectral bands (Red, Green, Blue and NIR) and three vegetation indices. Besides, a Random Forest (RF) method has been proposed in [1] to classify 9 crop types in Khorezm (Uzbekistan) considering as features the NDVI and EVI mean and standard deviation val- [4,3].…”